title
stringlengths 14
154
| paper_url
stringlengths 42
42
| authors
sequencelengths 1
21
| type
stringclasses 3
values | abstract
stringlengths 413
2.52k
| keywords
stringlengths 4
397
| TL;DR
stringlengths 5
250
⌀ | submission_number
int64 2
14.3k
| arxiv_id
stringlengths 10
10
⌀ | embedding
sequencelengths 384
384
| github_url
stringlengths 0
126
| github_stars
int64 0
55k
| num_models
int64 0
82
| num_datasets
int64 0
7
| num_spaces
int64 0
100
|
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
Conservative Contextual Bandits: Beyond Linear Representations | https://openreview.net/forum?id=SThJXvucjQ | [
"Rohan Deb",
"Mohammad Ghavamzadeh",
"Arindam Banerjee"
] | Poster | Conservative Contextual Bandits (CCBs) address safety in sequential decision making by requiring that an agent's policy, along with minimizing regret, also satisfies a safety constraint: the performance is not worse than a baseline policy (e.g., the policy that the company has in production) by more than $(1+\alpha)$ factor.
Prior work developed UCB-style
algorithms for this problem in the multi-armed (Wu et al., 2016) and contextual
linear (Kazerouni et al., 2017) settings.
However, in practice the cost of the arms
is often a non-linear function, and therefore existing UCB algorithms are ineffective in such settings.
In this paper, we consider CCBs beyond the linear case and develop two algorithms $\mathtt{C\text{-}SquareCB}$ and $\mathtt{C\text{-}FastCB}$, using Inverse Gap Weighting (IGW) based exploration and an online regression oracle.
We show that the safety constraint is satisfied in high probability and that the regret for $\mathtt{C\text{-}SquareCB}$ is sub-linear in horizon $T$, while the the regret for $\mathtt{C\text{-}FastCB}$ is first-order and is sub-linear in $L^*$, the cumulative loss of the optimal policy.
Subsequently, we use a neural network for function approximation and online gradient descent as the regression oracle to provide $\tilde{\mathcal{O}}\big(\sqrt{KT} + K/\alpha\big) $ and $\tilde{\mathcal{O}}\big(\sqrt{KL^*} + K (1 + 1/\alpha)\big)$ regret bounds respectively.
Finally, we demonstrate the efficacy of our algorithms on real world data, and show that they significantly outperform the existing baseline while maintaining the performance guarantee. | Contextual Bandits, Safety, Neural Bandits, Constrained Bandits | Algorithms for safe exploration in Conservative Contextual Bandits, ensuring performance stays within a safe range of a baseline policy for general non-linear costs. | 12,575 | 2412.06165 | [
-0.08450839668512344,
-0.034220319241285324,
-0.0741073414683342,
0.011819466017186642,
0.045240368694067,
0.05241065099835396,
0.043894167989492416,
0.03125086426734924,
0.04065052047371864,
0.026423228904604912,
-0.002405219478532672,
-0.015366367995738983,
0.09932896494865417,
-0.00863101240247488,
0.001844304846599698,
0.010678264312446117,
0.1071549654006958,
0.025683628395199776,
-0.012650461867451668,
0.015668630599975586,
-0.061613768339157104,
0.0010808921651914716,
0.00331631931476295,
-0.09724508970975876,
0.0010465803788974881,
-0.07359758019447327,
-0.026193829253315926,
0.006255189888179302,
0.05064179748296738,
0.010256300680339336,
0.029261816293001175,
-0.010433130897581577,
0.03238613158464432,
-0.08217647671699524,
-0.03541186451911926,
0.044069502502679825,
-0.14365141093730927,
-0.009442795999348164,
-0.009402274154126644,
0.061132654547691345,
-0.042034536600112915,
0.057939399033784866,
0.0008454194176010787,
0.06991951912641525,
0.027879448607563972,
-0.05977476015686989,
-0.00791349820792675,
-0.007135310675948858,
0.03972876816987991,
-0.03650679439306259,
-0.03845934569835663,
0.015134992077946663,
-0.026005160063505173,
-0.04780334606766701,
-0.012573130428791046,
-0.04171144589781761,
0.03659864515066147,
0.03910849615931511,
-0.002317924750968814,
0.03298569098114967,
0.00729124853387475,
-0.14262470602989197,
-0.016591249033808708,
-0.06344205886125565,
-0.023048914968967438,
0.012761739082634449,
-0.04524098336696625,
0.022937817499041557,
0.023895462974905968,
0.09757303446531296,
-0.0020602839067578316,
0.0022110508289188147,
-0.1047801524400711,
0.020596591755747795,
0.051838796585798264,
0.015798088163137436,
0.09864480048418045,
-0.011315837502479553,
0.03663134574890137,
-0.015469898469746113,
-0.003851643530651927,
-0.014462254010140896,
0.025098633021116257,
-0.03322843834757805,
0.05702057108283043,
-0.01886443980038166,
0.020849717780947685,
-0.011462322436273098,
0.16379643976688385,
-0.03823857381939888,
0.062423765659332275,
-0.032648008316755295,
-0.021349668502807617,
0.011506705544888973,
0.010470900684595108,
-0.023393258452415466,
0.023715421557426453,
-0.02263079211115837,
-0.06539808213710785,
0.07949290424585342,
0.062421467155218124,
0.038929689675569534,
-0.03328021988272667,
-0.03520960733294487,
-0.008808651939034462,
-0.061320457607507706,
-0.026872990652918816,
0.06426763534545898,
0.045711152255535126,
-0.05213232338428497,
0.020149579271674156,
0.05339478328824043,
0.05860939621925354,
0.01100391335785389,
-0.02328387275338173,
-0.012427475303411484,
-0.00029483193065971136,
-0.045802559703588486,
0.04508155584335327,
0.08795060962438583,
0.009497294202446938,
0.02506181038916111,
0.07475967705249786,
-0.0061266543343663216,
-0.026212161406874657,
-0.026621321216225624,
0.0032720016315579414,
4.068407472140349e-35,
0.01041547954082489,
-0.017337452620267868,
0.012509129010140896,
-0.10073443502187729,
-0.015688959509134293,
-0.030402012169361115,
0.036046139895915985,
-0.04475106671452522,
-0.08073478937149048,
0.06274302303791046,
0.05987299233675003,
0.019900480285286903,
-0.06660544127225876,
0.07297962158918381,
0.10914276540279388,
-0.03559447079896927,
-0.0010743109742179513,
0.02119695395231247,
0.012642497196793556,
-0.08409982919692993,
-0.0013062573270872235,
-0.0425700768828392,
0.006651383358985186,
-0.08629529923200607,
0.04944157972931862,
-0.03406797721982002,
0.001253716298379004,
-0.032381195574998856,
0.019487885758280754,
0.02719646319746971,
-0.10691303014755249,
0.05206991359591484,
-0.11633707582950592,
0.04332680627703667,
0.015913959592580795,
0.0010078110499307513,
-0.08802429586648941,
-0.04679325222969055,
-0.03930067643523216,
0.027337318286299706,
-0.05679774656891823,
0.07621393352746964,
0.06702383607625961,
-0.01434092503041029,
-0.03569214791059494,
-0.12730024755001068,
-0.009967273101210594,
0.010798156261444092,
-0.05303701385855675,
-0.05998682230710983,
-0.03527338430285454,
0.04320176690816879,
0.04277756065130234,
0.009889290668070316,
-0.04413371905684471,
0.02823207899928093,
-0.031613536179065704,
0.04302551597356796,
0.007837860845029354,
0.035852838307619095,
0.062260501086711884,
-0.0175953209400177,
-0.0031547374092042446,
0.037446554750204086,
-0.08764037489891052,
0.09790193289518356,
-0.0894923210144043,
0.01529123168438673,
0.056451234966516495,
-0.021064363420009613,
-0.003382706083357334,
-0.028413468971848488,
0.0989317074418068,
-0.07626140862703323,
0.05629776045680046,
0.007437147665768862,
0.07178681343793869,
0.0034558323677629232,
-0.006061087362468243,
-0.08622346818447113,
0.033233851194381714,
0.07191697508096695,
-0.0024902229197323322,
0.02471892163157463,
-0.054072894155979156,
-0.03765932098031044,
0.0010872522834688425,
-0.07199136167764664,
-0.01811613328754902,
-0.011956467293202877,
-0.10560385137796402,
0.032275594770908356,
-0.05866596847772598,
0.05687849596142769,
0.0300111286342144,
-8.337114704680378e-34,
0.0279172845184803,
-0.09716513752937317,
0.058824826031923294,
-0.04160299524664879,
0.0349978543817997,
-0.08613549917936325,
0.0136960344389081,
-0.15130196511745453,
0.0034402478486299515,
-0.034242164343595505,
-0.09189848601818085,
0.016126200556755066,
-0.043081723153591156,
0.009916172362864017,
0.07176990807056427,
-0.05725967884063721,
0.07421515136957169,
0.06618695706129074,
-0.040839217603206635,
-0.028523843735456467,
0.09671086817979813,
0.011459059081971645,
-0.10099652409553528,
0.044386740773916245,
-0.0022966910619288683,
-0.009201389737427235,
-0.04767507314682007,
0.04906144738197327,
-0.03212956711649895,
0.007486398797482252,
-0.055480487644672394,
-0.005704306531697512,
-0.01562800072133541,
0.038010548800230026,
-0.07735660672187805,
0.04268129914999008,
0.02546841651201248,
0.05005759000778198,
-0.04610366374254227,
0.08749163150787354,
0.0592757947742939,
0.012832192704081535,
0.010590141639113426,
-0.01506125833839178,
0.06399624794721603,
0.03318556770682335,
-0.03307519853115082,
-0.034124456346035004,
0.037923336029052734,
0.041147250682115555,
-0.02867799438536167,
0.026680603623390198,
-0.07680226117372513,
0.07290563732385635,
-0.07955022156238556,
-0.0687231495976448,
0.005265643820166588,
-0.0537109375,
0.06467994302511215,
0.03745151311159134,
-0.04194631800055504,
0.065040722489357,
-0.052433744072914124,
0.021601390093564987,
0.0005660160095430911,
0.04374728724360466,
-0.05644752085208893,
-0.025892209261655807,
-0.008410724811255932,
-0.04197332635521889,
-0.04850216954946518,
0.06244025379419327,
0.016202393919229507,
0.006305963732302189,
-0.01145867072045803,
0.06314846128225327,
0.023565370589494705,
0.0036280013155192137,
0.004362212494015694,
-0.017545662820339203,
-0.007950163446366787,
0.0064262207597494125,
0.08940771222114563,
0.06562061607837677,
0.005159096792340279,
0.09389091283082962,
-0.008350417949259281,
0.003991743549704552,
-0.0036553521640598774,
-0.08025988191366196,
-0.04397386312484741,
-0.04827659949660301,
0.048080798238515854,
0.025150608271360397,
-0.0038370161782950163,
-5.1726495087223157e-8,
-0.06147481128573418,
0.03846151754260063,
0.06586720049381256,
0.03496165946125984,
0.06688335537910461,
0.040227774530649185,
-0.02014787308871746,
-0.02438446134328842,
-0.06034073978662491,
0.011351952329277992,
0.04859108105301857,
0.04561415687203407,
-0.014841161668300629,
-0.015246565453708172,
-0.08717562258243561,
0.06712228059768677,
0.020086433738470078,
-0.06898479908704758,
-0.0024550864472985268,
-0.006379365921020508,
0.02506317012012005,
0.05642276257276535,
0.024661147966980934,
-0.0368187353014946,
0.022510875016450882,
-0.026347128674387932,
-0.03737587109208107,
0.10418150573968887,
0.11144156008958817,
0.1109088733792305,
-0.013164508156478405,
0.02136523276567459,
-0.0147936986759305,
0.009497945196926594,
0.05264666676521301,
0.009387516416609287,
0.0018538658041507006,
-0.02844870649278164,
-0.019616948440670967,
-0.031375739723443985,
0.052072104066610336,
0.018217500299215317,
0.029436787590384483,
0.022943677380681038,
0.10039692372083664,
0.06627623736858368,
-0.054616522043943405,
-0.07022655755281448,
0.05108736827969551,
-0.09007248282432556,
0.09234195202589035,
-0.04594853147864342,
-0.020607585087418556,
0.11226076632738113,
0.097505584359169,
-0.0017535854130983353,
-0.03581343591213226,
-0.09551197290420532,
-0.012838582508265972,
0.07699587941169739,
-0.0015535852871835232,
-0.0812627300620079,
-0.021834585815668106,
0.0040823109447956085
] | 0 | 0 | 0 | 0 |
|
Quantifying Generalization Complexity for Large Language Models | https://openreview.net/forum?id=jpSLXoRKnH | [
"Zhenting Qi",
"Hongyin Luo",
"Xuliang Huang",
"Zhuokai Zhao",
"Yibo Jiang",
"Xiangjun Fan",
"Himabindu Lakkaraju",
"James R. Glass"
] | Poster | While large language models (LLMs) have shown exceptional capabilities in understanding complex queries
and performing sophisticated tasks, their generalization abilities are often deeply entangled with
memorization, necessitating more precise evaluation.
To address this challenge, we introduce Scylla, a dynamic evaluation framework that quantitatively measures the generalization abilities of LLMs. Scylla disentangles generalization from memorization via assessing model performance on both in-distribution (ID) and out-of-distribution (OOD) data through 20 tasks across 5 levels of complexity.
Through extensive experiments, we uncover a non-monotonic relationship between task complexity and the performance gap between ID and OOD
data, which we term the generalization valley.
Specifically, this phenomenon reveals a critical threshold---referred to
as critical complexity---where reliance on non-generalizable behavior peaks, indicating the
upper bound of LLMs' generalization capabilities.
As model size increases, the critical complexity shifts toward higher levels of task complexity,
suggesting that larger models can handle more complex reasoning tasks before over-relying on
memorization.
Leveraging Scylla and the concept of critical complexity, we benchmark 28 LLMs including
both open-sourced models such as LLaMA and Qwen families, and closed-sourced models like Claude and
GPT, providing a more robust evaluation and establishing a clearer
understanding of LLMs' generalization capabilities. | large language model, generalization, evaluation | null | 12,574 | 2410.01769 | [
0.018553799018263817,
-0.1007855236530304,
-0.019731439650058746,
0.04471094533801079,
0.06175494194030762,
-0.07887879014015198,
0.0003413742233533412,
-0.0017591769574210048,
0.051761288195848465,
-0.04175347089767456,
-0.07474876940250397,
-0.026022134348750114,
-0.0004805182688869536,
0.023390619084239006,
0.02156134508550167,
-0.029186705127358437,
0.08827416598796844,
-0.029949838295578957,
-0.12054377049207687,
-0.0724136233329773,
0.06825631111860275,
0.0434805303812027,
0.08078611642122269,
-0.005803546402603388,
-0.003498415695503354,
-0.030339544638991356,
-0.04616665840148926,
-0.07187128067016602,
0.028927110135555267,
0.0016786407213658094,
0.031251415610313416,
0.055443424731492996,
0.04516869783401489,
0.05067875608801842,
0.007023182697594166,
0.03767671808600426,
-0.04084539785981178,
0.04342649132013321,
0.025878192856907845,
-0.030934354290366173,
-0.0010383418994024396,
0.03993432596325874,
0.06741679459810257,
0.0048059625551104546,
0.07093072682619095,
-0.02519673854112625,
-0.01871698722243309,
-0.05700225755572319,
-0.06221282109618187,
0.0016167806461453438,
-0.09117022901773453,
-0.002097455086186528,
-0.03741016238927841,
0.04717738926410675,
-0.01169859990477562,
-0.021512703970074654,
0.009246466681361198,
0.03773657977581024,
-0.06451193243265152,
-0.04208992421627045,
-0.07123744487762451,
-0.16725288331508636,
-0.010310014709830284,
0.010956957004964352,
-0.003578974399715662,
-0.01761818863451481,
0.06077598035335541,
-0.010624515824019909,
-0.07393179833889008,
0.03924429044127464,
0.003635657485574484,
0.07843140512704849,
-0.05415846034884453,
0.09389259666204453,
0.06295216828584671,
-0.01900123991072178,
0.024294422939419746,
-0.0002697305171750486,
0.04399890452623367,
-0.04226682335138321,
-0.009892474859952927,
0.055145975202322006,
0.024998286738991737,
0.031903769820928574,
0.11380856484174728,
0.0027979123406112194,
0.026525044813752174,
0.0259692519903183,
0.05485386401414871,
0.00891336053609848,
-0.04689982533454895,
-0.08571269363164902,
-0.040077995508909225,
-0.0010890585836023092,
0.08881969004869461,
0.05283428728580475,
0.07586460560560226,
-0.043419249355793,
-0.02127799205482006,
0.05355724319815636,
0.028398439288139343,
0.09270688146352768,
-0.015737885609269142,
-0.05637190490961075,
-0.019981900230050087,
-0.015443471260368824,
0.04587263986468315,
0.04302874207496643,
-0.037564631551504135,
-0.09939099103212357,
0.0020042285323143005,
0.03889087215065956,
0.11414890736341476,
0.025531703606247902,
-0.007752617355436087,
-0.01658349484205246,
0.03702544420957565,
-0.003784582018852234,
0.02899712324142456,
0.06433946639299393,
0.05573876202106476,
0.07173626124858856,
0.012606019154191017,
-0.007366163656115532,
-0.04315881431102753,
0.055026374757289886,
-0.014422567561268806,
1.3309061834380137e-33,
0.07298345863819122,
0.03839509189128876,
0.016048990190029144,
0.10946793854236603,
0.023701049387454987,
0.02941415272653103,
0.01124044694006443,
0.020758895203471184,
-0.024784859269857407,
0.06526406109333038,
-0.032967325299978256,
0.05346757918596268,
-0.049520257860422134,
-0.002828589640557766,
0.09957422316074371,
0.07766658812761307,
-0.0586397647857666,
0.06726448237895966,
-0.06444033235311508,
-0.0682525783777237,
0.10889366269111633,
-0.04732762277126312,
0.016617551445961,
-0.054560497403144836,
0.026569241657853127,
0.003358376445248723,
0.03322025015950203,
-0.07266281545162201,
-0.06648560613393784,
0.006193741224706173,
-0.09738194942474365,
-0.039992570877075195,
-0.028465555980801582,
0.054530058056116104,
-0.0005994596285745502,
-0.03084053099155426,
-0.020301561802625656,
-0.07750510424375534,
0.06417456269264221,
-0.06526795774698257,
-0.051283098757267,
0.022030392661690712,
0.006383886560797691,
0.049009133130311966,
-0.08395153284072876,
-0.05090441554784775,
0.0047196694649755955,
-0.03505514934659004,
-0.03388754278421402,
-0.029299093410372734,
0.0701783150434494,
-0.01302057970315218,
-0.02266377955675125,
0.016830498352646828,
0.05221084877848625,
0.04347456619143486,
0.011051495559513569,
0.042586781084537506,
-0.004516359884291887,
0.09979195892810822,
-0.02961486019194126,
0.04158473014831543,
-0.010344027541577816,
0.07372639328241348,
0.08969666808843613,
0.07022774964570999,
-0.10418078303337097,
0.10566497594118118,
0.07717245072126389,
-0.022760046645998955,
0.03777068853378296,
-0.05216973274946213,
-0.010880711488425732,
-0.06326831132173538,
0.08522345870733261,
-0.018915196880698204,
0.12154287844896317,
-0.1188492700457573,
-0.0026988531462848186,
0.037745438516139984,
-0.008311707526445389,
-0.03841596841812134,
0.044631678611040115,
-0.060008659958839417,
-0.08556736260652542,
-0.10901753604412079,
0.06322161853313446,
-0.03216201439499855,
-0.01841851696372032,
-0.035146355628967285,
-0.06594281643629074,
0.036074213683605194,
-0.023464858531951904,
-0.08195829391479492,
-0.019822651520371437,
-2.0135656196349298e-33,
-0.09714401513338089,
-0.06097549945116043,
-0.005259996745735407,
0.06948645412921906,
-0.01995016448199749,
-0.1148371770977974,
-0.036580685526132584,
0.009090128354728222,
-0.06570760905742645,
-0.09465231746435165,
-0.04567475616931915,
-0.04795394092798233,
0.06369393318891525,
0.0360015332698822,
0.053445883095264435,
0.014809174463152885,
0.037254657596349716,
0.06914062798023224,
0.050011489540338516,
0.08762142062187195,
0.0030482402071356773,
0.07180947810411453,
-0.060403067618608475,
0.0064492663368582726,
-0.015280057676136494,
0.05245164409279823,
-0.11022090911865234,
-0.009971126914024353,
-0.0028191714081913233,
-0.008572814986109734,
-0.0481884591281414,
-0.013371412642300129,
-0.08318010717630386,
0.005075434688478708,
-0.05547763407230377,
0.04222534969449043,
0.004673875402659178,
0.00045248825335875154,
-0.03473900258541107,
0.07632780820131302,
0.006979691796004772,
-0.040174104273319244,
-0.04708985239267349,
0.009959777817130089,
0.02376411110162735,
-0.0510188527405262,
-0.10410531610250473,
-0.017588958144187927,
0.027024948969483376,
-0.014201573096215725,
-0.006833659950643778,
0.05257086455821991,
-0.014882838353514671,
0.047279536724090576,
0.019205057993531227,
-0.0655134916305542,
-0.028513148427009583,
-0.07871019840240479,
-0.04545284062623978,
-0.00934539083391428,
-0.05569969117641449,
0.03548405319452286,
0.01813935488462448,
-0.05503356456756592,
0.021703481674194336,
-0.033512115478515625,
-0.021129880100488663,
-0.06005552038550377,
0.0017736823065206409,
0.03175044804811478,
-0.018787426874041557,
-0.06606292724609375,
-0.07773654907941818,
-0.005003917496651411,
-0.04168180003762245,
0.024823036044836044,
0.012390537187457085,
-0.02313484437763691,
-0.023764826357364655,
-0.01875391975045204,
0.060792382806539536,
0.0027510386426001787,
0.03488149866461754,
0.00019528223492670804,
0.0255544763058424,
0.04234318062663078,
0.06518512964248657,
0.11749602854251862,
0.004245210438966751,
0.03289218246936798,
-0.003144478192552924,
-0.020480724051594734,
-0.032588325440883636,
0.010861679911613464,
-0.03494438901543617,
-4.23509227687191e-8,
-0.04842255637049675,
0.04688297584652901,
-0.005698202643543482,
-0.00656358664855361,
0.0853365808725357,
-0.04054521024227142,
-0.11656545847654343,
0.09909751266241074,
0.01040116511285305,
-0.03238433972001076,
0.06563082337379456,
0.020064406096935272,
-0.02512253075838089,
0.026069138199090958,
0.05908602103590965,
0.1132112368941307,
-0.002570482436567545,
0.0013465280644595623,
-0.014828350394964218,
-0.007483301218599081,
-0.009143363684415817,
-0.03447364270687103,
-0.05971842631697655,
-0.02034146897494793,
-0.03947431594133377,
-0.0986577719449997,
0.016006087884306908,
0.07591032981872559,
0.04951219633221626,
-0.034390244632959366,
-0.011071661487221718,
0.027490414679050446,
-0.0339960940182209,
-0.01215462014079094,
0.03711991012096405,
0.026243245229125023,
0.0035560023970901966,
-0.012191402725875378,
0.010929390788078308,
-0.012692195363342762,
0.02541513554751873,
0.06304030865430832,
0.0049741147086024284,
0.0931730791926384,
0.025073496624827385,
-0.03261258453130722,
-0.09688788652420044,
-0.07959871739149094,
0.013446634635329247,
-0.04800922051072121,
0.06753841787576675,
-0.01412895880639553,
0.004548326600342989,
0.046115726232528687,
0.015569985844194889,
0.08098093420267105,
-0.01965457759797573,
-0.016044043004512787,
0.0019072460709139705,
0.009768922813236713,
0.015003465116024017,
0.02020733430981636,
-0.06050531566143036,
0.016022903844714165
] | https://github.com/zhentingqi/scylla | 5 | 0 | 0 | 0 |
SqueezeAttention: 2D Management of KV-Cache in LLM Inference via Layer-wise Optimal Budget | https://openreview.net/forum?id=9HK2rHNAhd | [
"Zihao Wang",
"Bin CUI",
"Shaoduo Gan"
] | Poster | Optimizing the Key-Value (KV) cache of the Large Language Model (LLM) has been considered critical to saving the cost of inference. Most of the existing KV-cache compression algorithms attempted to sparsify the sequence of tokens by taking advantage of the different importance of tokens. However, most of these methods treat all layers equally, allocating the same KV budget to each layer. This approach is suboptimal, as some layers may be less sensitive to input tokens yet still receive the same budget as others. In this work, we found that by identifying the importance of attention layers, we could optimize the KV-cache jointly from two dimensions, i.e., sequence-wise and layer-wise. Based on our observations regarding layer-wise importance in inference, we propose \sys to precisely optimize the allocation of KV-cache budget among layers on-the-fly and then incorporate three representative sequence-wise algorithms to compress the KV-cache for each layer with its very own budget. Specifically, we first measure each layer's importance by calculating the cosine similarity of the input prompt differences before and after the self-attention layers. Based on this similarity, we then categorize the layers into two groups and adjust their KV budgets accordingly. By optimizing the KV-cache from both sequence's and layer's dimensions, \sys achieves around 30\% to 70\% of the memory reductions and up to 2.2 $\times$ of throughput improvements in a wide range of LLMs and benchmarks. The code is available at https://github.com/hetailang/SqueezeAttention. | KV-cache, LLM inference optimization | null | 12,567 | 2404.04793 | [
0.0029599848203361034,
-0.0011609367793425918,
-0.006023988127708435,
0.013320928439497948,
0.05848262831568718,
0.03538127988576889,
-0.015875671058893204,
-0.00033750146394595504,
0.07138394564390182,
-0.016036227345466614,
-0.031417880207300186,
-0.03536378592252731,
-0.026515979319810867,
0.01521006878465414,
0.004075807053595781,
-0.035191237926483154,
0.13551421463489532,
0.09687688946723938,
-0.04792700707912445,
-0.10553830862045288,
0.07663959264755249,
-0.02324046567082405,
0.054879091680049896,
-0.04248502850532532,
0.07638050615787506,
0.0048264977522194386,
-0.020455215126276016,
-0.0319085419178009,
0.04151921719312668,
0.007830766029655933,
0.039331305772066116,
0.034405868500471115,
0.017138229683041573,
0.03607425093650818,
-0.027386978268623352,
0.05098923668265343,
-0.08949315547943115,
-0.02749408408999443,
-0.01771276816725731,
-0.0014796714531257749,
-0.06371892243623734,
0.06073465943336487,
-0.03888058662414551,
0.0290825292468071,
0.09177695959806442,
0.04788912460207939,
0.08656728267669678,
0.0023779161274433136,
-0.08100423216819763,
-0.04159139096736908,
-0.006433440372347832,
-0.048677098006010056,
-0.04648156836628914,
0.06580965965986252,
-0.04894541949033737,
-0.01885097101330757,
0.020668048411607742,
0.030530288815498352,
-0.04576642066240311,
0.01999099738895893,
-0.049986954778432846,
-0.11062625050544739,
0.0029697217978537083,
-0.023756638169288635,
-0.0097182746976614,
-0.031799376010894775,
0.11274112015962601,
0.05932300537824631,
-0.026212485507130623,
0.06471133977174759,
-0.02984282560646534,
0.08910238742828369,
-0.03134450688958168,
-0.007871135137975216,
0.017236899584531784,
0.10581953823566437,
0.023342082276940346,
-0.03471888601779938,
0.09332380443811417,
-0.14951921999454498,
-0.006560663226991892,
-0.015179348178207874,
0.010526632890105247,
-0.03552442044019699,
0.10109683126211166,
0.009799647144973278,
0.05196114256978035,
-0.03134037181735039,
0.017412960529327393,
0.005486031994223595,
-0.03192819654941559,
-0.02290532924234867,
0.08642364293336868,
-0.013955838046967983,
0.019542571157217026,
0.029628446325659752,
-0.01448686234652996,
0.007544293534010649,
-0.03063702955842018,
0.055578213185071945,
0.0493353046476841,
0.07519625127315521,
0.0398273766040802,
-0.040388260036706924,
0.06679791212081909,
-0.033729299902915955,
0.04901151731610298,
0.025890734046697617,
-0.0066499486565589905,
-0.08238343894481659,
0.005404507275670767,
0.08747518062591553,
0.019507138058543205,
-0.039087649434804916,
0.04427487775683403,
0.020261606201529503,
0.030203966423869133,
-0.010454686358571053,
0.07264896482229233,
0.0364731065928936,
-0.008553750813007355,
-0.009181372821331024,
-0.05368565022945404,
0.016272732987999916,
-0.001677462481893599,
-0.06588360667228699,
-0.020444486290216446,
3.680857052140507e-33,
0.025552719831466675,
0.07959745824337006,
-0.04121993109583855,
0.04940486699342728,
-0.017710503190755844,
-0.006087422836571932,
0.09068562090396881,
0.01592468097805977,
-0.04273749142885208,
0.03019418939948082,
-0.023902401328086853,
-0.02911803126335144,
-0.03581204265356064,
0.05727747455239296,
0.06373515725135803,
-0.002948973560705781,
-0.007400760427117348,
0.07108348608016968,
-0.045524489134550095,
-0.040514178574085236,
0.05602256581187248,
-0.021602986380457878,
0.052174825221300125,
-0.03706745058298111,
0.012684707529842854,
-0.057095952332019806,
0.0697038322687149,
-0.07519954442977905,
-0.052147869020700455,
-0.001702484441921115,
-0.044182244688272476,
-0.02880006842315197,
0.014006122946739197,
0.010967198759317398,
0.011544400826096535,
-0.0328628271818161,
0.024329081177711487,
-0.03254522383213043,
0.07376862317323685,
-0.05187522992491722,
-0.029353899881243706,
0.06424295157194138,
0.02700074389576912,
-0.03259071707725525,
-0.12012961506843567,
-0.056438758969306946,
-0.04521222785115242,
-0.0017589697381481528,
-0.009376902133226395,
-0.058489225804805756,
0.10886234045028687,
-0.0314490906894207,
-0.015949459746479988,
0.010744036175310612,
0.022584404796361923,
-0.0905790701508522,
0.05869174748659134,
0.008529906161129475,
0.06201613321900368,
0.12252198159694672,
-0.09604723751544952,
-0.010698511265218258,
-0.0034112806897610426,
0.09402107447385788,
0.03975560888648033,
0.06263282895088196,
-0.0006198299233801663,
0.08016566932201385,
0.011013874784111977,
-0.014571952633559704,
-0.05355055630207062,
-0.005267146974802017,
0.03335791453719139,
-0.09432153403759003,
0.008658161386847496,
-0.01834804378449917,
-0.021097898483276367,
-0.11872809380292892,
-0.05941535905003548,
0.013443747535347939,
-0.03109937720000744,
0.04140724614262581,
-0.009944973513484001,
0.01599128544330597,
-0.03839125484228134,
-0.10234498232603073,
0.04370250925421715,
-0.052233610302209854,
0.017249520868062973,
0.022121239453554153,
-0.03503154590725899,
-0.017784811556339264,
0.040825262665748596,
-0.07472971826791763,
-0.014330770820379257,
-2.429718170474867e-33,
-0.013699968345463276,
0.05474039167165756,
-0.024516653269529343,
0.08925952017307281,
-0.06074542924761772,
-0.03910546377301216,
-0.0610995888710022,
-0.03825439140200615,
-0.10389365255832672,
-0.06827329844236374,
-0.08755942434072495,
-0.08271608501672745,
0.05598696693778038,
0.033053770661354065,
0.05988619104027748,
0.013721524737775326,
-0.043007150292396545,
-0.04585849121212959,
0.0028081340715289116,
0.05166882276535034,
-0.0050593772903084755,
0.024849554523825645,
-0.053665634244680405,
0.010622191242873669,
-0.044660769402980804,
0.052598390728235245,
-0.051141753792762756,
0.041434068232774734,
0.026281099766492844,
-0.03816068917512894,
-0.10554264485836029,
-0.03963930159807205,
-0.07172974199056625,
-0.04505528509616852,
-0.017886465415358543,
0.01258903183043003,
0.04831701144576073,
0.03279474750161171,
-0.06170785799622536,
0.1086035892367363,
0.05684172362089157,
-0.06968481093645096,
-0.08434973657131195,
0.012156904675066471,
-0.008204714395105839,
-0.015974918380379677,
-0.08006343245506287,
0.004438337404280901,
-0.013345059007406235,
0.010399477556347847,
0.009496993385255337,
0.05268622189760208,
-0.07639975845813751,
0.025015121325850487,
0.016339784488081932,
-0.020414767786860466,
-0.12388040870428085,
0.017216606065630913,
0.10881698876619339,
-0.08413361757993698,
-0.010420800186693668,
-0.05892796069383621,
-0.010161071084439754,
-0.06319654732942581,
-0.015379894524812698,
-0.031021777540445328,
0.02500745840370655,
-0.06970761716365814,
0.003379070432856679,
-0.03098895400762558,
0.09831072390079498,
-0.002540318062528968,
0.011684966273605824,
0.0076179662719368935,
-0.028447246178984642,
0.0017480592941865325,
0.07411999255418777,
-0.03978367894887924,
0.019411657005548477,
0.0341608002781868,
-0.01995629258453846,
0.08590714633464813,
0.016141315922141075,
-0.037389688193798065,
0.03177681937813759,
0.05332942306995392,
0.03376488387584686,
0.0663677528500557,
-0.027083061635494232,
-0.029025554656982422,
-0.019986027851700783,
0.037688225507736206,
0.04086226597428322,
0.04038955271244049,
0.012803087942302227,
-4.523236540876496e-8,
-0.03886295109987259,
0.006222734693437815,
0.014997174963355064,
0.06933865696191788,
0.010799624025821686,
-0.0894375741481781,
-0.006148216314613819,
0.17113888263702393,
0.026930656284093857,
0.023996328935027122,
0.08041247725486755,
0.005130456760525703,
-0.05882422253489494,
-0.01643284037709236,
-0.005013549700379372,
0.09232736378908157,
-0.02788412757217884,
-0.0370333194732666,
-0.034212492406368256,
-0.05674147978425026,
-0.03374253958463669,
0.09590942412614822,
-0.019042162224650383,
0.0320451520383358,
0.04730597883462906,
-0.0522240474820137,
0.07361165434122086,
0.07739002257585526,
0.11041189730167389,
0.011837229132652283,
0.02986578270792961,
0.00916172657161951,
0.013061320409178734,
-0.09968168288469315,
0.05175945907831192,
0.03766479715704918,
-0.026269281283020973,
0.030599642544984818,
0.09035743772983551,
0.01007646881043911,
-0.0030633029527962208,
-0.006068390794098377,
-0.027994994074106216,
0.06566374003887177,
0.04504822939634323,
-0.038712795823812485,
-0.04605000093579292,
-0.05488200858235359,
0.0013788928044959903,
-0.04029450938105583,
-0.03356299549341202,
-0.04710422828793526,
-0.029485231265425682,
0.04219738021492958,
-0.06925612688064575,
0.028155608102679253,
0.00012717052595689893,
-0.023822598159313202,
0.03500409051775932,
0.026531241834163666,
0.06616215407848358,
-0.02714240923523903,
-0.10397271811962128,
-0.04690141975879669
] | https://github.com/hetailang/squeezeattention | 37 | 0 | 0 | 0 |
Gaussian Mixture Counterfactual Generator | https://openreview.net/forum?id=lBB3eSn6fY | [
"Jong-Hoon Ahn",
"Akshay Vashist"
] | Poster | We address the individualized treatment effect (ITE) estimation problem, focusing on continuous, multidimensional, and time-dependent treatments for precision medicine. The central challenge lies in modeling these complex treatment scenarios while capturing dynamic patient responses and minimizing reliance on control data. We propose the Gaussian Mixture Counterfactual Generator (GMCG), a generative model that transforms the Gaussian mixture model—traditionally a tool for clustering and density estimation—into a new tool explicitly geared toward causal inference. This approach generates robust counterfactuals by effectively handling continuous and multidimensional treatment spaces. We evaluate GMCG on synthetic crossover trial data and simulated datasets, demonstrating its superior performance over existing methods, particularly in scenarios with limited control data. GMCG derives its effectiveness from modeling the joint distribution of covariates, treatments, and outcomes using a latent state vector while employing a conditional distribution of the state vector to suppress confounding and isolate treatment-outcome relationships. GMCG shows promise for enhancing ITE estimation in precision medicine, offering a potential unified solution for personalized therapeutic strategies. | Gaussian mixture model, synthetic data generation, clinical trial, individual treatment effect, counterfactual generation | null | 12,561 | null | [
-0.026377590373158455,
-0.08407697826623917,
0.05269144847989082,
0.07531088590621948,
0.04298541694879532,
0.048166532069444656,
0.013221831992268562,
0.008012453094124794,
0.011974507011473179,
-0.07654711604118347,
-0.03133495897054672,
-0.04301944002509117,
0.04344033822417259,
0.028227169066667557,
-0.12073938548564911,
-0.07622871547937393,
0.1434417963027954,
0.05067548528313637,
-0.016569223254919052,
-0.03603076934814453,
-0.06422983855009079,
0.08875486999750137,
0.02441428415477276,
0.02671092562377453,
-0.03567745164036751,
0.024050496518611908,
0.006962495855987072,
-0.07022465020418167,
0.0049111912958323956,
-0.006533805280923843,
0.1037123054265976,
0.07093401998281479,
-0.07381273061037064,
-0.0041951751336455345,
-0.11843942105770111,
0.044432949274778366,
-0.04097864031791687,
0.06898552179336548,
-0.010354758240282536,
0.02828071266412735,
0.02346375398337841,
-0.010343619622290134,
0.02681884914636612,
0.05824418365955353,
0.017976855859160423,
-0.025410402566194534,
-0.034932609647512436,
-0.056538958102464676,
0.00783800520002842,
0.06346481293439865,
-0.00907958298921585,
0.030261166393756866,
0.006536830682307482,
0.06881484389305115,
0.04541269317269325,
-0.04148330166935921,
0.07758603990077972,
-0.040905941277742386,
-0.026111509650945663,
0.049004510045051575,
-0.07234064489603043,
-0.1345057189464569,
-0.019464362412691116,
-0.0359160415828228,
0.01996738463640213,
0.04927784204483032,
-0.015146411024034023,
0.08138895034790039,
-0.045109957456588745,
0.014511835761368275,
-0.04341898113489151,
0.06876380741596222,
-0.016994167119264603,
0.09298942238092422,
-0.06289415061473846,
0.04202063009142876,
-0.001998970517888665,
0.0055008833296597,
0.056877486407756805,
-0.03674992918968201,
0.011950911954045296,
0.013680750504136086,
0.0011118741240352392,
-0.009103265590965748,
0.09461262077093124,
-0.014390922151505947,
0.050183381885290146,
0.1196940615773201,
0.043025944381952286,
0.009958638809621334,
0.05274342745542526,
0.10976053029298782,
0.08361966162919998,
-0.003820778103545308,
0.014900602400302887,
-0.011339209042489529,
-0.03231352940201759,
-0.08871088176965714,
0.11658721417188644,
0.009320102632045746,
0.04020557552576065,
0.02606743760406971,
0.08661109209060669,
0.0118942279368639,
-0.031840067356824875,
-0.09736689925193787,
0.001199815422296524,
0.051846303045749664,
0.04533376544713974,
-0.014215786010026932,
0.007912504486739635,
0.0065645985305309296,
0.06483132392168045,
0.01620037667453289,
-0.04420529678463936,
0.005157862324267626,
0.04930253326892853,
-0.016671976074576378,
0.016598226502537727,
0.0408426858484745,
0.01892784796655178,
-0.03647926449775696,
0.028231240808963776,
0.015117739327251911,
0.0021588383242487907,
0.06569341570138931,
-0.048080071806907654,
2.1255658030263158e-33,
-0.000873431155923754,
-0.04489734396338463,
0.040580864995718,
0.05998913571238518,
0.010047492571175098,
-0.003049259539693594,
-0.08169610053300858,
0.017092032358050346,
0.02721809409558773,
-0.04612017795443535,
0.007440544664859772,
0.0693432092666626,
-0.11869966983795166,
0.09656581282615662,
-0.024213159456849098,
0.09866459667682648,
-0.08374017477035522,
0.1017393171787262,
-0.055642809718847275,
-0.0349145345389843,
0.017933260649442673,
-0.05762038379907608,
-0.06007843092083931,
0.015024666674435139,
-0.0888487845659256,
0.012732550501823425,
0.013866081833839417,
0.004300691187381744,
-0.007139616645872593,
-0.039118994027376175,
-0.09974125027656555,
-0.013380182906985283,
0.049009520560503006,
0.06289836019277573,
0.005107482895255089,
-0.023425104096531868,
-0.01937674917280674,
-0.0670151636004448,
-0.018173683434724808,
0.027912640944123268,
-0.04151621460914612,
-0.02807784453034401,
0.06720578670501709,
-0.01248018816113472,
-0.05006955936551094,
-0.09282872825860977,
0.016267864033579826,
-0.06965433061122894,
-0.00808983575552702,
-0.05518268421292305,
0.04362400621175766,
0.006433224305510521,
0.0044676559045910835,
-0.06775445491075516,
-0.089288130402565,
0.04206583648920059,
-0.07023201137781143,
-0.0027452006470412016,
-0.000056574201153125614,
0.012787741608917713,
0.02706335112452507,
0.07366611063480377,
-0.012869437225162983,
-0.002437005518004298,
-0.06307270377874374,
0.06277354061603546,
-0.0403737872838974,
-0.02507587894797325,
0.015812044963240623,
0.09581761807203293,
-0.01873904839158058,
0.0612783320248127,
-0.014077036641538143,
0.0772118791937828,
0.031282197684049606,
-0.023247191682457924,
0.015301648527383804,
-0.02361264079809189,
-0.03281036764383316,
0.02190401218831539,
0.005118988920003176,
-0.01445462740957737,
-0.12795409560203552,
0.06838011741638184,
-0.027371717616915703,
-0.022774366661906242,
-0.03370225802063942,
-0.006738359108567238,
-0.15306182205677032,
-0.07142278552055359,
-0.028448492288589478,
-0.01501465030014515,
0.005445015616714954,
0.08268371969461441,
0.0004004049696959555,
-3.017055208560169e-33,
-0.052850522100925446,
0.023859651759266853,
0.07501807808876038,
-0.019879121333360672,
0.10890737175941467,
0.0033252991270273924,
0.029813624918460846,
-0.034195128828287125,
0.14167040586471558,
-0.05377240106463432,
0.029439060017466545,
-0.02809998206794262,
0.07496941834688187,
-0.019918257370591164,
-0.0008428257424384356,
0.015408110804855824,
-0.02255314216017723,
0.03826432302594185,
-0.01716167852282524,
0.00901772826910019,
0.06494135409593582,
-0.029076682403683662,
0.027897246181964874,
-0.038755957037210464,
-0.06847652047872543,
0.028876828029751778,
0.02772040292620659,
-0.015836389735341072,
0.051070064306259155,
0.013194125145673752,
-0.05268822982907295,
-0.04715534672141075,
-0.05399398133158684,
-0.021128278225660324,
-0.051735542714595795,
0.027003539726138115,
0.05046263709664345,
-0.050596725195646286,
-0.050355955958366394,
-0.05458297207951546,
0.02531461790204048,
-0.0021807795856148005,
-0.11960054188966751,
0.02359018102288246,
0.03488657996058464,
-0.011568806134164333,
-0.01520225964486599,
0.004229169338941574,
0.06266427785158157,
-0.016717832535505295,
-0.08935227245092392,
0.07304953038692474,
-0.03520174324512482,
0.031108740717172623,
-0.04717181995511055,
-0.14451268315315247,
0.025334786623716354,
-0.11407320201396942,
-0.00316076772287488,
-0.009320548735558987,
-0.06243234500288963,
-0.033114224672317505,
0.018775075674057007,
-0.049301594495773315,
-0.02632417157292366,
0.005842140875756741,
0.05248726159334183,
0.024099409580230713,
-0.01087270025163889,
0.05765398591756821,
-0.00960217323154211,
0.03199712559580803,
0.04799003154039383,
-0.07885660976171494,
0.06732030212879181,
-0.022930728271603584,
0.021385788917541504,
-0.021549295634031296,
-0.03423910588026047,
-0.014949954114854336,
0.11276434361934662,
-0.028194373473525047,
0.044148899614810944,
0.013656321913003922,
-0.02969159558415413,
0.07828476279973984,
0.0026520052924752235,
0.06672050803899765,
0.08129014819860458,
0.043709974735975266,
-0.07025974988937378,
-0.012123062275350094,
-0.02106584794819355,
0.08452893793582916,
-0.103314608335495,
-4.3356955359286076e-8,
0.009579685516655445,
0.003892783308401704,
0.043047863990068436,
-0.005192319396883249,
-0.019634954631328583,
-0.003615854773670435,
-0.1018504872918129,
0.01045321486890316,
-0.031026644632220268,
-0.005288716871291399,
0.04993676021695137,
-0.005580051336437464,
0.021726176142692566,
-0.07210047543048859,
-0.06409526616334915,
-0.008181140758097172,
-0.02581920102238655,
0.03930654376745224,
0.014767564833164215,
0.005275374744087458,
-0.04815683886408806,
-0.031178051605820656,
0.017494916915893555,
-0.01570160686969757,
0.05396133288741112,
-0.01695122756063938,
0.007411299739032984,
0.0652158334851265,
0.03579018637537956,
-0.03635789453983307,
0.014359867200255394,
-0.008551281876862049,
-0.054632723331451416,
0.04066089168190956,
-0.06458431482315063,
0.014664717018604279,
0.026233481243252754,
0.04319296404719353,
-0.006513708736747503,
-0.048207059502601624,
-0.014218221418559551,
0.09564505517482758,
-0.04178764671087265,
0.046967409551143646,
0.056955885142087936,
0.03376477584242821,
-0.03213845193386078,
-0.05977393686771393,
0.04649026319384575,
0.04414599761366844,
0.056550368666648865,
-0.03398965299129486,
0.030795231461524963,
0.06753331422805786,
0.07784964889287949,
0.009108058176934719,
0.019081586971879005,
-0.04245928302407265,
0.018149234354496002,
-0.09004965424537659,
-0.043055459856987,
0.011020202189683914,
-0.05510145053267479,
-0.00022721147979609668
] | 0 | 0 | 0 | 0 |
|
Multi-domain Distribution Learning for De Novo Drug Design | https://openreview.net/forum?id=g3VCIM94ke | [
"Arne Schneuing",
"Ilia Igashov",
"Adrian W. Dobbelstein",
"Thomas Castiglione",
"Michael M. Bronstein",
"Bruno Correia"
] | Poster | We introduce DrugFlow, a generative model for structure-based drug design that integrates continuous flow matching with discrete Markov bridges, demonstrating state-of-the-art performance in learning chemical, geometric, and physical aspects of three-dimensional protein-ligand data. We endow DrugFlow with an uncertainty estimate that is able to detect out-of-distribution samples. To further enhance the sampling process towards distribution regions with desirable metric values, we propose a joint preference alignment scheme applicable to both flow matching and Markov bridge frameworks. Furthermore, we extend our model to also explore the conformational landscape of the protein by jointly sampling side chain angles and molecules. | Drug Discovery, Flow Matching, Markov Bridge, Equivariance | null | 12,557 | null | [
-0.04667067900300026,
-0.0762520283460617,
-0.022076983004808426,
-0.12190448492765427,
-0.03510043025016785,
-0.08103262633085251,
-0.016987361013889313,
0.040052782744169235,
0.10555987805128098,
-0.015845146030187607,
-0.050129592418670654,
-0.0002523798029869795,
0.03855108097195625,
0.08604288101196289,
-0.009128662757575512,
0.004787860903888941,
-0.04585489258170128,
0.07642002403736115,
-0.011727729812264442,
0.016125014051795006,
-0.006242856848984957,
0.0559573657810688,
0.0003647175326477736,
0.0028376798145473003,
-0.009409273974597454,
0.10589316487312317,
0.07322730123996735,
-0.02226177603006363,
-0.006364956498146057,
-0.06326571106910706,
0.086310014128685,
0.08009731024503708,
-0.04483269527554512,
-0.09486287832260132,
0.0372803658246994,
0.017470046877861023,
-0.09946438670158386,
0.048379551619291306,
0.025132738053798676,
-0.01138661615550518,
0.07671954482793808,
0.0621989481151104,
0.04051227122545242,
0.017866462469100952,
0.04284025728702545,
-0.016635706648230553,
-0.08216748386621475,
-0.0366736464202404,
0.09388761967420578,
0.08950225263834,
0.00378642650321126,
-0.0191225316375494,
-0.05243039131164551,
0.09698519110679626,
-0.0050364830531179905,
0.06187361106276512,
0.009547299705445766,
-0.05295021831989288,
-0.027125796303153038,
0.0013046221574768424,
0.027104472741484642,
-0.09769836068153381,
-0.009958373382687569,
0.016856670379638672,
0.02444993332028389,
-0.015921158716082573,
0.04404916614294052,
0.09583891183137894,
0.019084282219409943,
0.01870511844754219,
-0.053477831184864044,
0.09654276072978973,
-0.043519146740436554,
0.07208553701639175,
0.0581522062420845,
0.04135163500905037,
-0.008863931521773338,
-0.016236567869782448,
-0.013853533193469048,
-0.09641211479902267,
0.019332999363541603,
0.06322984397411346,
0.08068761229515076,
0.055981650948524475,
0.06895940005779266,
0.02730288729071617,
-0.03415777161717415,
0.08573473989963531,
0.06520374864339828,
-0.020141635090112686,
-0.018315577879548073,
0.093979611992836,
-0.015385662205517292,
-0.062040045857429504,
-0.07393816858530045,
0.03154668211936951,
-0.006336727179586887,
-0.035317644476890564,
0.022936204448342323,
0.05371667817234993,
0.0009074861882254481,
0.014112232252955437,
0.05333748459815979,
-0.0197912510484457,
0.049311067909002304,
-0.05426573380827904,
-0.014502518810331821,
0.06832484155893326,
0.03757540509104729,
-0.028048161417245865,
0.030472243204712868,
0.008976899087429047,
0.10112936794757843,
0.0531325601041317,
-0.006997619289904833,
0.0033537079580128193,
0.00028180002118460834,
-0.04654499888420105,
0.007991199381649494,
0.009060104377567768,
-0.10561306774616241,
0.007346372585743666,
-0.0580810084939003,
-0.03673642873764038,
-0.04428528621792793,
0.008849910460412502,
-0.05360189452767372,
-2.7867964363852224e-33,
-0.005297128576785326,
-0.003607782768085599,
0.09396544843912125,
0.0695343017578125,
0.04894786328077316,
-0.0025748168118298054,
0.04057295620441437,
-0.06670989841222763,
-0.038760505616664886,
0.03805406019091606,
-0.07876893877983093,
0.013276303187012672,
-0.07133302092552185,
0.08048886060714722,
-0.01647147908806801,
-0.03322858735918999,
-0.025376582518219948,
0.08648349344730377,
-0.017165422439575195,
0.004439514130353928,
0.0075968848541378975,
-0.06436851620674133,
-0.05391164869070053,
-0.015216431580483913,
-0.027711883187294006,
0.03471890836954117,
-0.0576217882335186,
0.07628866285085678,
-0.035508595407009125,
-0.001798203680664301,
-0.08161024749279022,
-0.0006175351445563138,
-0.02921201102435589,
-0.023578735068440437,
0.05335920676589012,
-0.03219732269644737,
0.031486012041568756,
-0.04933418706059456,
-0.03895900025963783,
-0.04874765872955322,
0.00047491106670349836,
-0.024056995287537575,
0.05349242687225342,
-0.013560357503592968,
-0.0801108330488205,
0.03953336179256439,
0.02039908803999424,
-0.01686638593673706,
0.004639549646526575,
-0.0441640242934227,
0.06689058989286423,
-0.026372823864221573,
0.027982624247670174,
-0.0076783038675785065,
-0.06406022608280182,
0.0200794730335474,
-0.10847220569849014,
0.031310491263866425,
0.04876613989472389,
0.0928051546216011,
-0.04895290359854698,
0.05950602889060974,
-0.08102019876241684,
0.03666150942444801,
0.0842941626906395,
0.030745403841137886,
-0.029090233147144318,
-0.03916217386722565,
0.05474335327744484,
0.04753241688013077,
-0.012344198301434517,
0.0261703934520483,
-0.07005050033330917,
-0.02590642310678959,
0.07239589095115662,
-0.0062926774844527245,
0.008451561443507671,
0.07088805735111237,
0.05396765097975731,
0.08583798259496689,
-0.08754679560661316,
-0.019252948462963104,
-0.0828462764620781,
-0.009314274415373802,
-0.0320456437766552,
-0.08679290115833282,
0.09034252911806107,
-0.09797780215740204,
-0.1314242035150528,
0.007581709884107113,
-0.06138523668050766,
-0.030133243650197983,
0.01486403588205576,
0.06498828530311584,
0.05167736858129501,
5.355518742390244e-34,
-0.010681688785552979,
0.0021691028960049152,
0.13438259065151215,
0.0009671561419963837,
0.01270086970180273,
0.02375071495771408,
0.007170054130256176,
-0.008610603399574757,
0.08939889818429947,
-0.05167799070477486,
0.0035675864201039076,
0.026574034243822098,
-0.011541287414729595,
-0.11319737881422043,
0.027874518185853958,
0.026822421699762344,
-0.0177347119897604,
-0.03589553385972977,
0.03214576840400696,
0.036533936858177185,
0.027702471241354942,
0.020070858299732208,
-0.10230100154876709,
0.02131376788020134,
-0.035490233451128006,
0.044170819222927094,
0.0033715434838086367,
-0.017211049795150757,
0.021226543933153152,
0.04456079378724098,
-0.09345871210098267,
-0.03990340977907181,
-0.10391171276569366,
0.05582808330655098,
-0.12093120813369751,
0.05377567932009697,
-0.027988260611891747,
0.023495472967624664,
0.03595630079507828,
0.012597274035215378,
-0.039933960884809494,
0.010584385134279728,
-0.14491721987724304,
0.03842267021536827,
-0.0035432414151728153,
-0.013467219658195972,
0.0065467762760818005,
0.03125090152025223,
0.04617686569690704,
-0.019567199051380157,
-0.03629918023943901,
0.11393319815397263,
-0.002181361196562648,
-0.016117721796035767,
-0.041905563324689865,
-0.04538882523775101,
-0.0175698883831501,
-0.03619080036878586,
0.09297683089971542,
-0.002088343258947134,
-0.08178134262561798,
-0.04541149362921715,
0.014337685890495777,
0.0012768383603543043,
-0.04027978703379631,
0.011138996109366417,
0.03329161927103996,
-0.016046009957790375,
-0.01667678914964199,
0.010887706652283669,
0.025995317846536636,
0.003390267491340637,
0.06838658452033997,
0.04010843485593796,
0.06929261237382889,
-0.09979745000600815,
-0.005893276538699865,
-0.07771096378564835,
-0.03611346706748009,
-0.018073413521051407,
0.03803999722003937,
0.006693553179502487,
-0.008805961348116398,
0.08561289310455322,
0.08522284030914307,
0.017541153356432915,
0.06899954378604889,
-0.04180915653705597,
0.05363486334681511,
-0.004638253711163998,
0.023897666484117508,
0.027708733454346657,
-0.012819092720746994,
0.05732513219118118,
-0.03333781287074089,
-3.085076016873245e-8,
-0.0014404859393835068,
-0.028500182554125786,
0.07209773361682892,
-0.02768760547041893,
0.005292666144669056,
0.11150099337100983,
-0.04816848412156105,
0.056821804493665695,
-0.03252093121409416,
-0.010320079512894154,
0.06628567725419998,
0.03327029198408127,
-0.06980279833078384,
-0.03518784046173096,
-0.05347573757171631,
0.03107517957687378,
-0.02116306498646736,
0.04985402151942253,
-0.03294474259018898,
-0.03022763505578041,
-0.07558933645486832,
0.03470683470368385,
0.021042760461568832,
-0.015494792722165585,
0.05536627024412155,
-0.08615463227033615,
-0.012012463063001633,
-0.08753817528486252,
0.04503348097205162,
0.015494841150939465,
0.019595498219132423,
-0.024666380137205124,
0.016942594200372696,
0.05103191360831261,
0.05815945938229561,
-0.004925651475787163,
-0.026965657249093056,
-0.037973012775182724,
-0.05751362815499306,
-0.04488430172204971,
-0.041416168212890625,
0.07144656032323837,
-0.08026424795389175,
-0.03178095817565918,
-0.004517305176705122,
0.02400481514632702,
0.020776601508259773,
0.029881766065955162,
0.015312218107283115,
0.016612200066447258,
0.04635106399655342,
-0.016284117475152016,
0.012715273536741734,
-0.05073237419128418,
-0.004396392498165369,
-0.0008817109628580511,
-0.05754118040204048,
-0.0794137641787529,
0.03666453808546066,
-0.09815547615289688,
-0.025008054450154305,
-0.04302583634853363,
0.0038027516566216946,
-0.042835913598537445
] | 0 | 0 | 0 | 0 |
|
{τ}-bench: A Benchmark for \underline{T}ool-\underline{A}gent-\underline{U}ser Interaction in Real-World Domains | https://openreview.net/forum?id=roNSXZpUDN | [
"Shunyu Yao",
"Noah Shinn",
"Pedram Razavi",
"Karthik R Narasimhan"
] | Poster | Existing benchmarks for language agents do not set them up to interact with human users or follow domain-specific rules, both of which are vital to safe and realistic deployment. We propose $\tau$-bench, a benchmark with two domains (retail and airline) emulating dynamic conversations between a user (simulated by language models) and a customer service agent provided with domain-specific API tools and policy guidelines. We employ a efficient and faithful evaluation process that compares the database state at the end of a conversation with the annotated goal state, and propose a new metric (pass^k) to evaluate the reliability of agent behavior over multiple trials. Our experiments show that even state-of-the-art function calling agents (gpt-4o) succeed on $<50\%$ of the tasks, and are terribly inconsistent (pass^8 < 25\% in retail). Our findings point to the need for methods that can improve the ability of agents to act consistently and reliably follow rules. | language model, language agent, benchmark, user simulation, rule following | A benchmark where language agents need to interact with simulated users and follow complex rules in realistic domains. | 12,553 | null | [
-0.0535074882209301,
-0.062557153403759,
-0.012418169528245926,
-0.019763758406043053,
-0.035126693546772,
-0.07233570516109467,
0.067010797560215,
-0.0057886624708771706,
0.03923599794507027,
-0.07896696776151657,
-0.06351028382778168,
-0.02577173337340355,
0.03062605857849121,
0.04343807324767113,
0.10394713282585144,
-0.014775598421692848,
0.13996408879756927,
0.01201856043189764,
-0.1202063336968422,
-0.049569688737392426,
0.02920048125088215,
0.0382729172706604,
0.02612871117889881,
0.026866838335990906,
-0.05365364998579025,
-0.07168319821357727,
-0.029275808483362198,
-0.049517638981342316,
-0.013657943345606327,
0.0037132143042981625,
-0.006501838099211454,
-0.00486407708376646,
-0.03466110676527023,
0.05900007113814354,
-0.02858385629951954,
-0.011708707548677921,
-0.02075624093413353,
-0.06261032819747925,
-0.011274280957877636,
-0.06124193221330643,
-0.06708595156669617,
-0.05140738934278488,
-0.06341300159692764,
0.059268638491630554,
0.0851101502776146,
-0.073926642537117,
-0.04258262738585472,
0.007335437461733818,
-0.02626311406493187,
0.08658810704946518,
-0.11467881500720978,
0.025743583217263222,
0.08009302616119385,
0.04422006383538246,
-0.013361386023461819,
0.09057038277387619,
0.025990264490246773,
0.06715213507413864,
0.0014251881511881948,
-0.0455542616546154,
-0.050958313047885895,
-0.13607841730117798,
-0.055150896310806274,
0.02315080538392067,
-0.06925670057535172,
0.04029729962348938,
-0.04087119549512863,
-0.02988823875784874,
-0.025306666269898415,
0.00856427475810051,
-0.11864131689071655,
0.017667893320322037,
-0.011006598360836506,
0.07707247883081436,
0.011553890071809292,
0.09484807401895523,
-0.006396784912794828,
-0.04697277396917343,
0.04330733045935631,
-0.04661105200648308,
0.011981978081166744,
-0.056181054562330246,
-0.017788955941796303,
0.024230213835835457,
0.08405950665473938,
0.00035909138387069106,
0.05619129166007042,
-0.03142243251204491,
0.08811700344085693,
0.01975547894835472,
-0.041823748499155045,
0.0008469055755995214,
0.0668751522898674,
0.01085299625992775,
-0.023650942370295525,
0.050889793783426285,
-0.006124437320977449,
-0.07386302947998047,
-0.05877479910850525,
0.1015649139881134,
-0.031118903309106827,
0.09492679685354233,
-0.010144911706447601,
-0.02627844177186489,
-0.07336809486150742,
0.003583295037969947,
0.03876359388232231,
-0.006330200936645269,
-0.017214016988873482,
-0.025772344321012497,
-0.055484920740127563,
0.018870484083890915,
0.07266349345445633,
0.02448362112045288,
0.06699420511722565,
-0.01224839873611927,
-0.0023953309282660484,
-0.026119770482182503,
0.04028884693980217,
0.02318824827671051,
0.0031145752873271704,
0.04137161746621132,
0.0700591430068016,
0.02577328123152256,
0.0558503195643425,
0.10279808938503265,
0.0728544071316719,
-6.35930505180357e-34,
0.025284964591264725,
0.07199883460998535,
0.011572008021175861,
0.00823169108480215,
0.012399655766785145,
-0.0340484119951725,
0.0011752620339393616,
-0.011479065753519535,
-0.035576123744249344,
-0.029224896803498268,
-0.0167118851095438,
0.007498625200241804,
0.006533327512443066,
0.0015892129158601165,
0.11586891114711761,
0.021842435002326965,
-0.019039006903767586,
0.07998770475387573,
-0.009576143696904182,
0.012814771384000778,
0.07446666061878204,
0.03538572043180466,
0.060926180332899094,
0.014327352866530418,
0.10231860727071762,
0.021431073546409607,
-0.025240790098905563,
-0.04195787385106087,
0.04059586673974991,
0.03326409310102463,
-0.07203744351863861,
-0.024600353091955185,
0.002814533421769738,
0.07199768722057343,
-0.03008630871772766,
-0.000267149560386315,
-0.06372285634279251,
-0.0753556564450264,
0.00593827897682786,
0.004694636445492506,
-0.12466998398303986,
0.04917062073945999,
0.03925434499979019,
0.03343302384018898,
-0.07770009338855743,
-0.03755544498562813,
-0.04567807912826538,
0.0031564785167574883,
0.004455965012311935,
-0.004056866746395826,
-0.029643770307302475,
0.08966770023107529,
0.02335442416369915,
-0.02068791538476944,
0.0004811633552890271,
-0.05980387330055237,
0.01605365425348282,
-0.01637161523103714,
-0.010183792561292648,
0.03654899820685387,
0.04469740018248558,
-0.04262017086148262,
-0.012835312634706497,
0.0933724120259285,
0.00889701396226883,
-0.015295104123651981,
-0.06839284300804138,
0.07829552888870239,
0.05697575584053993,
-0.013333884999155998,
0.005902886390686035,
0.062417663633823395,
0.01994812861084938,
-0.001567299012094736,
0.005254032555967569,
-0.029692720621824265,
0.07697515934705734,
-0.08886873722076416,
0.04639044031500816,
-0.043237924575805664,
0.009838585741817951,
0.01581192947924137,
-0.0379461869597435,
-0.006825194228440523,
-0.005615457892417908,
-0.06375247240066528,
-0.00042471240158192813,
-0.03423502668738365,
-0.05553073436021805,
0.03191893920302391,
-0.09779786318540573,
-0.01155203115195036,
-0.028840627521276474,
0.07995007932186127,
0.044831834733486176,
-6.732176910836688e-34,
-0.0007066738908179104,
-0.06327210366725922,
0.0017418391071259975,
0.11954409629106522,
-0.09090283513069153,
-0.037542834877967834,
-0.030854754149913788,
0.0065368241630494595,
0.0248984657227993,
-0.02175893820822239,
-0.08730485290288925,
0.00528394989669323,
0.05856261029839516,
0.014404527842998505,
0.07792077958583832,
-0.06801144778728485,
0.0028950972482562065,
-0.03380421921610832,
0.06951511651277542,
0.036970920860767365,
0.08127330988645554,
0.0753326490521431,
-0.14149168133735657,
0.059806838631629944,
-0.053560301661491394,
0.010343989357352257,
-0.05184963345527649,
-0.005382460076361895,
-0.009227229282259941,
-0.0876927524805069,
-0.00491917971521616,
0.04310033097863197,
0.005229673348367214,
-0.0026884437538683414,
-0.028413694351911545,
0.006525504868477583,
0.07803907245397568,
0.05051816627383232,
-0.01032348070293665,
0.061126239597797394,
0.05254395306110382,
-0.06490012258291245,
-0.10056225210428238,
-0.06421635299921036,
-0.015261310152709484,
0.06898980587720871,
-0.0941791832447052,
-0.018268395215272903,
-0.03736294060945511,
-0.02469129115343094,
-0.0036436913069337606,
-0.0409153550863266,
-0.047490641474723816,
-0.010897689498960972,
-0.06489836424589157,
-0.050070494413375854,
-0.019510377198457718,
-0.037692222744226456,
-0.013727400451898575,
-0.018057312816381454,
-0.07270912826061249,
0.03982342407107353,
0.037854649126529694,
0.00057010113960132,
0.006657538935542107,
-0.038457680493593216,
0.035058554261922836,
-0.007007913198322058,
0.06135283410549164,
-0.03326554223895073,
0.022861890494823456,
-0.0326186902821064,
-0.05694444850087166,
0.0464123971760273,
0.06245746836066246,
0.017419276759028435,
-0.044927000999450684,
-0.08435221016407013,
-0.0014730232069268823,
-0.015174219384789467,
-0.03716183453798294,
0.049749188125133514,
0.04336109757423401,
-0.015041308477520943,
0.015261310152709484,
0.09810398519039154,
0.020962124690413475,
0.09388165175914764,
0.01297598797827959,
0.040265437215566635,
-0.014461526647210121,
0.006826058961451054,
-0.07810709625482559,
0.0367724746465683,
-0.08820009976625443,
-5.234035427292838e-8,
0.030638085678219795,
0.036303602159023285,
0.06498586386442184,
0.03883630037307739,
0.0008829800062812865,
-0.0014723391504958272,
-0.018108196556568146,
0.011513079516589642,
0.01558020431548357,
-0.022451695054769516,
0.0455574207007885,
-0.012601441703736782,
-0.07082751393318176,
0.006515578832477331,
0.028822049498558044,
0.0498775988817215,
-0.04219770058989525,
0.041234295815229416,
-0.06474439799785614,
0.008298467844724655,
0.00434851786121726,
0.09227786958217621,
0.027793636545538902,
-0.008162448182702065,
0.008419269695878029,
0.012183020822703838,
-0.03767550364136696,
0.10873401165008545,
0.06494416296482086,
0.014212190173566341,
0.04908442869782448,
-0.043344803154468536,
0.02569292113184929,
-0.06359288096427917,
-0.01550242118537426,
0.008382715284824371,
0.025881072506308556,
-0.05611371621489525,
0.05286967381834984,
0.013909845612943172,
0.024990418925881386,
0.020404087379574776,
-0.03176851198077202,
-0.007467743009328842,
0.1186632364988327,
-0.021405918523669243,
-0.13362087309360504,
-0.09373340755701065,
0.08231431990861893,
-0.08881058543920517,
-0.06361916661262512,
-0.0588001050055027,
-0.04138093814253807,
0.04664747789502144,
0.06971192359924316,
-0.03910473361611366,
-0.019657518714666367,
-0.08779295533895493,
0.09199351072311401,
0.06762713193893433,
0.03436979278922081,
-0.01755957119166851,
-0.03034924529492855,
-0.029881732538342476
] | 0 | 0 | 0 | 0 |
|
DEPfold: RNA Secondary Structure Prediction as Dependency Parsing. | https://openreview.net/forum?id=DpLFmc09pC | [
"KE WANG",
"Shay B Cohen"
] | Poster | RNA secondary structure prediction is critical for understanding RNA function
but remains challenging due to complex structural elements like pseudoknots and
limited training data. We introduce DEPfold, a novel deep learning approach that
re-frames RNA secondary structure prediction as a dependency parsing problem.
DEPfold presents three key innovations: (1) a biologically motivated transformation of RNA structures into labeled dependency trees, (2) a biaffine attention
mechanism for joint prediction of base pairings and their types, and (3) an optimal
tree decoding algorithm that enforces valid RNA structural constraints. Unlike traditional energy-based methods, DEPfold learns directly from annotated data and
leverages pretrained language models to predict RNA structure. We evaluate DEPfold on both within-family and cross-family RNA datasets, demonstrating significant performance improvements over existing methods. DEPfold shows strong
performance in cross-family generalization when trained on data augmented by
traditional energy-based models, outperforming existing methods on the bpRNAnew dataset. This demonstrates DEPfold’s ability to effectively learn structural
information beyond what traditional methods capture. Our approach bridges natural language processing (NLP) with RNA biology, providing a computationally
efficient and adaptable tool for advancing RNA structure prediction and analysis | RNA secondary structure prediction, Dependency parsing, Biaffine attention, Pseudoknots, Pretrained Model, Deep learning | DEPfold reframes RNA secondary structure prediction as dependency parsing. Using structure transformation, Biaffine parser, and optimal decoding, it outperforms existing methods, especially for pseudoknots and long-range interactions. | 12,543 | null | [
-0.06944059580564499,
-0.08024612814188004,
0.055928681045770645,
0.04648381471633911,
0.043665677309036255,
-0.009381944313645363,
-0.09073161333799362,
-0.030499231070280075,
-0.03650146722793579,
-0.019762860611081123,
-0.056178104132413864,
-0.10723069310188293,
-0.09238937497138977,
0.04894976317882538,
-0.027449317276477814,
0.06612644344568253,
0.0076401494443416595,
0.022480284795165062,
-0.032576415687799454,
0.01577603630721569,
0.0394531786441803,
0.04896625503897667,
-0.024049490690231323,
0.010908474214375019,
0.0667513757944107,
0.09608422219753265,
0.027585331350564957,
0.012411055155098438,
0.008433360606431961,
-0.01015783566981554,
0.052237674593925476,
-0.0718173161149025,
0.04416416585445404,
0.06327514350414276,
-0.08007350564002991,
0.09546998143196106,
-0.04841625317931175,
-0.027447227388620377,
0.04251539707183838,
0.002005638787522912,
0.03931723162531853,
0.0666687861084938,
0.03534725680947304,
0.004598970990628004,
0.0498349666595459,
-0.009119576774537563,
0.012863907031714916,
-0.020254485309123993,
0.027675112709403038,
0.030946675688028336,
0.027981910854578018,
-0.0007992693572305143,
-0.00910217221826315,
0.08744542300701141,
0.03330293297767639,
0.09085322916507721,
0.030707968398928642,
-0.012544986791908741,
-0.07582665979862213,
0.0048511135391891,
0.014174424111843109,
-0.11629854887723923,
-0.03774907439947128,
-0.06629979610443115,
-0.008320711553096771,
-0.018060114234685898,
0.006335238926112652,
0.08469271659851074,
-0.04087939113378525,
-0.014563029631972313,
-0.030220387503504753,
0.08281796425580978,
-0.007450085133314133,
0.11810214072465897,
0.04507812485098839,
0.11642682552337646,
0.03951378911733627,
0.02432859316468239,
0.04936954751610756,
-0.1212397962808609,
-0.03490031883120537,
0.05930028110742569,
0.076703280210495,
0.014758198522031307,
0.09939321130514145,
0.016784099861979485,
0.02586943283677101,
0.08278848230838776,
-0.04643276333808899,
0.015019726008176804,
-0.05019702762365341,
-0.07996172457933426,
0.07497841864824295,
0.019997768104076385,
0.016109155490994453,
0.10289262980222702,
-0.03889322653412819,
-0.01917681284248829,
-0.07612434029579163,
-0.0018016588874161243,
-0.010025965049862862,
0.09824014455080032,
0.010676458477973938,
-0.0388791523873806,
0.0005585105391219258,
-0.005113790277391672,
0.026329748332500458,
0.003699020016938448,
0.048126839101314545,
-0.03522852435708046,
0.07462157309055328,
0.09644883126020432,
-0.006552024278789759,
-0.022528037428855896,
0.0477403961122036,
-0.048718731850385666,
-0.0072420937940478325,
-0.04483450949192047,
0.048451829701662064,
0.028265640139579773,
-0.03848091885447502,
-0.008479217067360878,
-0.06595192849636078,
-0.01510976254940033,
-0.021791942417621613,
-0.043508924543857574,
-0.053066905587911606,
2.7580792930656263e-33,
0.02776041254401207,
0.03424890339374542,
0.05117655545473099,
0.022483637556433678,
0.04642435908317566,
-0.05750042572617531,
-0.06894178688526154,
0.03889421001076698,
-0.11629872769117355,
0.00278018182143569,
-0.010545403696596622,
-0.040294744074344635,
-0.03716225549578667,
0.034443825483322144,
-0.023487664759159088,
-0.062335677444934845,
-0.03798674792051315,
-0.008851614780724049,
-0.034690938889980316,
0.01587374322116375,
-0.0071341549046337605,
0.054895054548978806,
-0.0552871935069561,
-0.0326169952750206,
0.03916154429316521,
-0.007506771944463253,
0.0508580319583416,
-0.07840443402528763,
-0.04668735712766647,
0.0030218621250241995,
-0.12364578992128372,
-0.010509108193218708,
-0.046644940972328186,
0.01468267384916544,
0.013376005925238132,
-0.13608556985855103,
0.04977383837103844,
-0.04035801440477371,
-0.009582615457475185,
-0.01734919287264347,
0.07573751360177994,
0.05104842036962509,
-0.013558379374444485,
0.06733576208353043,
-0.048934053629636765,
-0.07947531342506409,
0.06776716560125351,
0.00501729641109705,
-0.03946986049413681,
-0.03688807040452957,
0.054919514805078506,
-0.04994522035121918,
0.008548653684556484,
0.009217116981744766,
-0.030468937009572983,
0.011189333163201809,
-0.0015052518574520946,
0.010593670420348644,
0.07380904257297516,
0.026325449347496033,
0.03196264058351517,
0.03454745188355446,
0.044442661106586456,
0.10235195606946945,
0.012351009994745255,
0.05004488304257393,
-0.09203352779150009,
-0.025044457986950874,
-0.008484229445457458,
0.020391779020428658,
-0.017494522035121918,
0.03507208451628685,
0.0029829612467437983,
-0.0061445157043635845,
0.040715739130973816,
0.0426754355430603,
0.06670535355806351,
-0.06794995814561844,
-0.006816665176302195,
0.13293972611427307,
-0.04946399852633476,
0.03303325176239014,
0.043807607144117355,
-0.006741408258676529,
-0.1272871494293213,
-0.11684223264455795,
0.09132397174835205,
-0.08772270381450653,
-0.03777887672185898,
0.015275131911039352,
-0.11818528175354004,
0.0011665943311527371,
-0.035929013043642044,
-0.048936475068330765,
0.04221583530306816,
-2.069447886389052e-33,
-0.009874979965388775,
0.05160444602370262,
-0.014155173674225807,
-0.05620093643665314,
-0.03302665799856186,
-0.041483860462903976,
-0.0672474205493927,
-0.058205071836709976,
0.014504966326057911,
-0.015067930333316326,
0.05416311323642731,
-0.005647667683660984,
0.0633758157491684,
-0.03178207203745842,
0.05978117883205414,
0.04097863286733627,
-0.03903556987643242,
0.029567111283540726,
-0.027157530188560486,
-0.01011984795331955,
0.03953678905963898,
0.12152207642793655,
-0.10954240709543228,
0.02751300297677517,
-0.007636287249624729,
0.005662512965500355,
0.021923720836639404,
0.05339978635311127,
-0.03808603435754776,
-0.03170551732182503,
-0.06931836903095245,
0.026663273572921753,
0.03511457145214081,
-0.017459793016314507,
-0.08228548616170883,
0.08031804859638214,
-0.06390330195426941,
-0.06958574801683426,
-0.013317844830453396,
-0.0353010892868042,
0.014184687286615372,
0.03552558273077011,
-0.12425721436738968,
0.04596496745944023,
-0.043301574885845184,
-0.019081657752394676,
-0.02761959284543991,
-0.003986571915447712,
0.025024479255080223,
0.10232324153184891,
0.03852967545390129,
0.010755540803074837,
-0.099746473133564,
0.026415104046463966,
-0.03325924277305603,
-0.11389020830392838,
-0.028995608910918236,
0.010062117129564285,
0.06837543100118637,
-0.01251771580427885,
-0.01894991658627987,
-0.005954669788479805,
0.04387102648615837,
-0.032689183950424194,
0.016102036461234093,
-0.05964114889502525,
0.011666538193821907,
-0.02126409113407135,
0.030522475019097328,
-0.07439378648996353,
0.030617164447903633,
-0.007895308546721935,
0.041347701102495193,
0.0459522008895874,
-0.01025668065994978,
-0.006605499889701605,
0.022552302107214928,
-0.053876105695962906,
-0.08708574622869492,
-0.01660723052918911,
0.019652560353279114,
-0.043865710496902466,
-0.003322374541312456,
0.08502648770809174,
0.08822722733020782,
0.04599883779883385,
0.08557591587305069,
0.05323096737265587,
-0.01167119387537241,
0.0357661135494709,
-0.05237004533410072,
-0.021412404254078865,
-0.030688045546412468,
0.11332666128873825,
-0.031084969639778137,
-3.7045062128981954e-8,
-0.09661564975976944,
-0.0029257049318403006,
0.013523761183023453,
0.025082025676965714,
0.01542041264474392,
-0.027876384556293488,
0.019399680197238922,
0.108156219124794,
-0.022880010306835175,
-0.02291153371334076,
0.03888207674026489,
-0.012092575430870056,
-0.030923329293727875,
-0.055667806416749954,
0.03879128769040108,
0.09348376095294952,
0.06347447633743286,
-0.031165027990937233,
-0.040636129677295685,
0.024407140910625458,
0.017933547496795654,
0.03344469889998436,
-0.004181167110800743,
0.05579189211130142,
-0.022460684180259705,
-0.05287129804491997,
0.0348668210208416,
0.0004153736517764628,
0.08816642314195633,
-0.017018446698784828,
0.03728487715125084,
0.052759379148483276,
0.0007578200893476605,
-0.06939323246479034,
-0.04588942974805832,
0.0788453221321106,
0.03420037776231766,
0.023043576627969742,
-0.002170035382732749,
0.036714792251586914,
0.06470536440610886,
0.028694355860352516,
-0.07678122073411942,
0.03524819016456604,
-0.06945835798978806,
-0.022906800732016563,
-0.043145664036273956,
0.04494340345263481,
0.030922183766961098,
-0.03452059254050255,
0.07167525589466095,
0.03432212397456169,
0.020203430205583572,
-0.0008542296709492803,
-0.005279664881527424,
0.02938791550695896,
-0.062001783400774,
0.013637986034154892,
0.02873249351978302,
0.0022505689412355423,
-0.009021253325045109,
-0.04884093254804611,
0.03543316200375557,
0.005525748711079359
] | 0 | 0 | 0 | 0 |
|
APE: Faster and Longer Context-Augmented Generation via Adaptive Parallel Encoding | https://openreview.net/forum?id=yUC8pU508S | [
"Xinyu Yang",
"Tianqi Chen",
"Beidi Chen"
] | Poster | Context-augmented generation (CAG) techniques, including RAG and ICL, require the efficient combination of multiple contexts to generate responses to user queries. Directly inputting these contexts as a sequence introduces a considerable computational burden by re-encoding the combined selection of contexts for every request. To address this, we explore the promising potential of parallel encoding to independently pre-compute and cache each context's KV states. This approach enables the direct loading of cached states during inference while accommodating more contexts through position reuse across contexts. However, due to misalignments in attention distribution, directly applying parallel encoding results in a significant performance drop. To enable effective and efficient CAG, we propose Adaptive Parallel Encoding (**APE**), which brings shared prefix, attention temperature, and scaling factor to align the distribution of parallel encoding with sequential encoding. Results on RAG and ICL tasks demonstrate that APE can preserve 98\% and 93\% sequential encoding performance using the same inputs while outperforming parallel encoding by 3.6\% and 7.9\%, respectively. It also scales to many-shot CAG, effectively encoding hundreds of contexts in parallel. Efficiency evaluation shows that APE can achieve an end-to-end 4.5$\times$ speedup by reducing 28$\times$ prefilling time for a 128K-length context. The code is available at
https://github.com/Infini-AI-Lab/APE. | Parallel Encoding; Context-Augmented LLM; Efficient Inference; Length Extrapolation | null | 12,539 | 2502.05431 | [
-0.0374116525053978,
-0.038135506212711334,
-0.003841017372906208,
0.0769897922873497,
-0.041737496852874756,
0.010835051536560059,
0.009322023019194603,
-0.050394654273986816,
0.048424020409584045,
-0.027753986418247223,
-0.012218194082379341,
-0.11626918613910675,
0.06780306249856949,
-0.026286574080586433,
0.05137187987565994,
-0.027675867080688477,
0.05478793382644653,
0.03687535226345062,
-0.0713200494647026,
-0.06404512375593185,
0.03293336555361748,
-0.021141765639185905,
0.05361520126461983,
-0.035235337913036346,
0.032176241278648376,
0.015508302487432957,
-0.08080369234085083,
-0.050299521535634995,
0.08869723975658417,
0.03145003318786621,
0.06957166641950607,
0.05466693267226219,
-0.032508037984371185,
0.06405055522918701,
-0.047748979181051254,
0.028270097449421883,
-0.1214388832449913,
-0.012567299418151379,
-0.04082002490758896,
-0.05210305377840996,
0.008898320607841015,
0.02903599478304386,
-0.0303727425634861,
0.052483897656202316,
0.04155367612838745,
0.03180241584777832,
0.039512068033218384,
-0.013426212593913078,
0.024553926661610603,
-0.01020015962421894,
-0.04989957809448242,
-0.003594722831621766,
-0.0222415030002594,
0.02120242640376091,
-0.03521796315908432,
0.05233313888311386,
0.014117456041276455,
0.07754892110824585,
-0.056542471051216125,
0.020290929824113846,
-0.08724649250507355,
-0.1251358836889267,
0.006469088606536388,
0.01196469459682703,
0.029866013675928116,
0.0149645134806633,
0.04911854863166809,
-0.028918033465743065,
-0.0238636527210474,
-0.024342630058526993,
-0.05044318363070488,
0.023898323997855186,
-0.043896883726119995,
0.07522331178188324,
0.023505140095949173,
0.08101481944322586,
-0.0015496343839913607,
-0.07703045755624771,
0.06931061297655106,
-0.09796696901321411,
0.018885402008891106,
-0.036432214081287384,
0.02599782682955265,
-0.00016862204938661307,
0.168882817029953,
-0.0031283556018024683,
0.06943663954734802,
-0.03994361311197281,
-0.05269351229071617,
0.025664212182164192,
-0.03405851870775223,
-0.08870524168014526,
0.09855510294437408,
-0.010696482844650745,
0.008899515494704247,
0.05808844789862633,
0.026467954739928246,
0.00435174023732543,
-0.03932870551943779,
0.03891300782561302,
0.09232443571090698,
0.03918379917740822,
-0.006127096246927977,
0.0028427194338291883,
0.0014127587201073766,
-0.014268328435719013,
0.03154154494404793,
0.07176719605922699,
-0.04246842488646507,
-0.025474432855844498,
0.011265885084867477,
0.12800005078315735,
0.03302262723445892,
-0.04054891690611839,
-0.027823567390441895,
0.006446495186537504,
-0.017659906297922134,
-0.07966212183237076,
0.07526596635580063,
0.05183038115501404,
0.0026763235218822956,
-0.027056461200118065,
-0.07649599015712738,
0.01988385058939457,
0.009854108095169067,
-0.10983418673276901,
-0.0038708748761564493,
2.411323887936406e-33,
0.03178488090634346,
0.025741729885339737,
0.02090957947075367,
-0.022710006684064865,
0.00044279289431869984,
0.07685773074626923,
0.0033756508491933346,
0.04837555065751076,
-0.07969480752944946,
-0.024806668981909752,
0.03498591482639313,
0.02379884198307991,
-0.02734435349702835,
0.10299605131149292,
0.022883517667651176,
0.005767131224274635,
-0.036836449056863785,
0.08053778856992722,
-0.022011928260326385,
0.023990651592612267,
-0.004637497942894697,
0.02199961431324482,
-0.00148115586489439,
-0.020935263484716415,
0.03283923864364624,
0.0695839449763298,
0.0035082465037703514,
-0.059889595955610275,
-0.06929412484169006,
0.002319875406101346,
-0.09036409109830856,
-0.0029948765877634287,
-0.07704201340675354,
0.04266834259033203,
-0.019162897020578384,
-0.06633386015892029,
0.06485220789909363,
-0.03777248039841652,
-0.032851144671440125,
0.030270136892795563,
0.043604105710983276,
0.07292400300502777,
0.039007026702165604,
-0.005979119800031185,
-0.13960471749305725,
-0.07819651067256927,
-0.062370382249355316,
-0.015342633239924908,
-0.035589396953582764,
0.054999612271785736,
0.060559697449207306,
0.02480274997651577,
-0.020757796242833138,
-0.043605320155620575,
0.032676175236701965,
-0.04609377682209015,
0.006219226401299238,
0.047375619411468506,
0.060303039848804474,
0.1211366057395935,
0.0033488492481410503,
-0.014307496137917042,
-0.03693564236164093,
0.11607779562473297,
0.016292184591293335,
-0.030819835141301155,
0.049966033548116684,
0.0808592140674591,
0.04423869401216507,
0.031465236097574234,
-0.04759738966822624,
0.002824094844982028,
0.005926822312176228,
-0.02727060578763485,
0.029118923470377922,
0.020622611045837402,
-0.02531106024980545,
-0.0357949398458004,
-0.05692597106099129,
0.028324615210294724,
-0.009233467280864716,
0.050858501344919205,
-0.054183125495910645,
0.024227909743785858,
-0.0537685789167881,
-0.043477438390254974,
0.03405768796801567,
-0.08167701959609985,
-0.031230265274643898,
-0.04376859962940216,
0.009192535653710365,
-0.04705079644918442,
0.018616629764437675,
-0.004213559441268444,
-0.08445023745298386,
-1.8612874896546028e-33,
0.01936093531548977,
-0.04301215708255768,
-0.03628412261605263,
0.02343841828405857,
-0.0168045312166214,
-0.07092615962028503,
0.053641658276319504,
-0.01310237031430006,
-0.1051180511713028,
-0.07878892123699188,
-0.028850093483924866,
-0.05915800854563713,
0.10096429288387299,
0.010096542537212372,
0.06456238776445389,
0.018841782584786415,
-0.003259462071582675,
0.030952101573348045,
-0.03245202824473381,
0.05527122691273689,
-0.003551308298483491,
0.030519843101501465,
-0.03795201703906059,
0.018969302996993065,
-0.009375319816172123,
0.02065742388367653,
-0.054237693548202515,
-0.02830512262880802,
0.03555962070822716,
-0.061996206641197205,
-0.056029628962278366,
0.058207884430885315,
-0.00446690758690238,
-0.0492246150970459,
-0.03782554343342781,
0.0409030057489872,
0.046235956251621246,
0.010037897154688835,
-0.08306309580802917,
0.06275273859500885,
0.11736534535884857,
-0.035524968057870865,
-0.025166276842355728,
0.06879454851150513,
0.013985786586999893,
0.06051511690020561,
-0.09925153851509094,
-0.02065899968147278,
-0.04646585136651993,
0.08944134414196014,
-0.004337200429290533,
-0.005126035772264004,
-0.06579145789146423,
-0.015714261680841446,
-0.02790425717830658,
-0.11215179413557053,
-0.06793666630983353,
-0.02594675123691559,
0.03795504570007324,
-0.0524616464972496,
0.006359107792377472,
-0.044391755014657974,
0.02170180343091488,
-0.08515655249357224,
0.06712088733911514,
-0.0042765201069414616,
0.004385339096188545,
-0.04594990983605385,
-0.0060720606707036495,
-0.047251515090465546,
0.02880179136991501,
-0.05034025385975838,
-0.0035504568368196487,
-0.0011561966966837645,
0.05191060155630112,
0.03398415073752403,
0.0611092634499073,
-0.07464675605297089,
0.02184971794486046,
-0.007590315770357847,
-0.06822802871465683,
0.0654401034116745,
-0.015130318701267242,
0.06311163306236267,
-0.012188036926090717,
0.09760867059230804,
0.003209480084478855,
0.0753103718161583,
0.004083431791514158,
0.02209845744073391,
-0.05095526948571205,
0.0977693721652031,
-0.044458284974098206,
0.0667790025472641,
-0.033007290214300156,
-4.1475509249266906e-8,
-0.04818274453282356,
0.04552823305130005,
-0.06638737767934799,
0.09030470252037048,
0.046820782124996185,
-0.0000013861135812476277,
-0.1004151925444603,
0.07797716557979584,
0.011891139671206474,
-0.0069285426288843155,
0.06923947483301163,
-0.05308280512690544,
0.0066697439178824425,
-0.02909352257847786,
0.04379540681838989,
0.05402056872844696,
0.010434243828058243,
-0.019959328696131706,
0.00889570452272892,
-0.031928543001413345,
-0.023727862164378166,
0.13062360882759094,
0.014826205559074879,
-0.013702025637030602,
0.008631767705082893,
-0.03646557778120041,
0.04533063992857933,
0.09341655671596527,
0.06853353977203369,
-0.1014099270105362,
-0.028047308325767517,
-0.019280308857560158,
-0.033854108303785324,
-0.0728331059217453,
-0.023669704794883728,
0.016692819073796272,
0.00230123451910913,
-0.01567796617746353,
0.06593114882707596,
-0.04371538385748863,
0.05487877130508423,
-0.0475568063557148,
-0.014460085891187191,
0.065926194190979,
0.02005661278963089,
0.015491255559027195,
0.04082527756690979,
-0.040593817830085754,
0.042054250836372375,
-0.10050813108682632,
-0.044090185314416885,
-0.10241913050413132,
0.021986965090036392,
0.014240971766412258,
-0.00015766025171615183,
-0.018596399575471878,
0.011087983846664429,
-0.03699173778295517,
0.09285683184862137,
0.08251578360795975,
0.06307120621204376,
-0.032140813767910004,
-0.08136086165904999,
-0.0036338097415864468
] | https://github.com/infini-ai-lab/ape | 22 | 0 | 0 | 0 |
SFS: Smarter Code Space Search improves LLM Inference Scaling | https://openreview.net/forum?id=MCHuGOkExF | [
"Jonathan Light",
"Yue Wu",
"Yiyou Sun",
"Wenchao Yu",
"Yanchi Liu",
"Xujiang Zhao",
"Ziniu Hu",
"Haifeng Chen",
"Wei Cheng"
] | Poster | We frame code generation as a black-box optimization problem within the code
space and demonstrate how optimization-inspired techniques can enhance inference
scaling over text. Based on this perspective, we propose **SCATTERED FOREST
SEARCH (SFS)**, a novel approach that improves solution diversity during evolutionary search,
thereby avoiding local optima. Our theoretical analysis illustrates how these
methods improve exploration and enhance efficiency. Extensive experiments
on *HumanEval, MBPP, APPS, CodeContests,* and *Leetcode* reveal significant
performance gains. For instance, our method achieves a **pass@1 rate of 67.1% on
HumanEval+** and **87.2% on HumanEval with GPT-3.5**, marking improvements of
**8.6%** and **4.3%** over the state-of-the-art, while also halving the iterations needed
to find the correct solution. Furthermore, our approach scales more efficiently
than existing search techniques, including **tree search, line search,** and **repeated
sampling (Best of N)**. | LLM, code generation, optimization, search, agent, inference scaling, black-box optimization, exploration-exploitation, tree search, Monte Carlo Tree Search (MCTS), evolutionary search, large-scale inference, solution diversity, textual optimization, prompt engineering, reinforcement learning, metaheuristic search, computational efficiency, program synthesis | We frame code generation as a black-box optimization problem over textual space and introduce optimization-inspired search techniques to enhance LLM inference scaling, achieving state-of-the-art performance with fewer iterations. | 12,517 | null | [
-0.0013890088303014636,
-0.041394613683223724,
-0.034518204629421234,
0.04210003837943077,
0.14292468130588531,
-0.036444131284952164,
-0.04547076299786568,
0.0025535377208143473,
-0.017355630174279213,
-0.012297418899834156,
-0.058862823992967606,
-0.07066672295331955,
-0.006586776580661535,
0.007444932125508785,
-0.025731075555086136,
0.07306435704231262,
0.030228130519390106,
0.06674592941999435,
-0.0316479466855526,
-0.0485348217189312,
0.0067090438678860664,
0.04485998675227165,
0.030982354655861855,
0.006649998482316732,
0.005061065312474966,
0.03335302323102951,
-0.06983157992362976,
-0.028080450370907784,
0.04812706634402275,
0.024073628708720207,
0.11094894260168076,
0.08137688785791397,
0.042261891067028046,
0.044475048780441284,
-0.016273044049739838,
0.05869606137275696,
-0.13522642850875854,
0.00011788299889303744,
-0.012082993052899837,
0.00625881552696228,
-0.10747158527374268,
0.09554785490036011,
0.04106982424855232,
0.03204381465911865,
0.05713050812482834,
-0.027785886079072952,
-0.05939393490552902,
-0.05416034907102585,
-0.022446179762482643,
-0.08835283666849136,
-0.05949241667985916,
-0.04533504322171211,
-0.09863381832838058,
-0.023441826924681664,
-0.02479851245880127,
-0.05069592967629433,
0.10046172887086868,
-0.05903712660074234,
0.052540890872478485,
-0.04654403030872345,
0.002002336084842682,
-0.06396640837192535,
0.02767571434378624,
-0.04167618229985237,
0.03370337188243866,
-0.02248968556523323,
0.02425605244934559,
0.0071629309095442295,
0.05137975141406059,
-0.05629514902830124,
-0.010626912117004395,
0.09480034559965134,
-0.03760608658194542,
0.08200964331626892,
0.04191024601459503,
0.08070901781320572,
0.00023784919176250696,
0.0033096489496529102,
0.06693292409181595,
-0.06911210715770721,
-0.02233857847750187,
-0.03260507807135582,
-0.05026978626847267,
0.0855378583073616,
0.07985681295394897,
0.015982920303940773,
-0.013951084576547146,
0.029679523780941963,
0.04512115567922592,
0.08160119503736496,
0.015273156575858593,
-0.07373670488595963,
0.03275175020098686,
0.0014890913153067231,
-0.058087900280952454,
-0.023416105657815933,
-0.02706736885011196,
-0.1014704778790474,
-0.04084806889295578,
0.04880662262439728,
-0.008136644028127193,
0.03432189300656319,
0.04386158660054207,
-0.10831622779369354,
-0.04441007971763611,
0.040703609585762024,
0.04158754646778107,
0.03417980670928955,
0.08132363110780716,
-0.06618506461381912,
0.027666686102747917,
0.07358919829130173,
0.007350206840783358,
0.10307858884334564,
0.021195514127612114,
-0.00587973790243268,
0.0709596797823906,
-0.03867260739207268,
-0.0035244820173829794,
0.09805387258529663,
0.008990692906081676,
-0.012047071009874344,
-0.08752308785915375,
-0.026358289644122124,
0.02203369140625,
0.02903985232114792,
-0.0241236612200737,
1.802880481440105e-33,
0.032878514379262924,
0.00952261220663786,
0.018975554034113884,
0.013310257345438004,
0.04782941937446594,
-0.046608537435531616,
0.04326563701033592,
0.058599065989255905,
-0.08550120145082474,
0.05981627479195595,
-0.09054558724164963,
-0.00734340213239193,
-0.012931899167597294,
0.024873116984963417,
0.0697813332080841,
-0.05545354634523392,
-0.08211961388587952,
0.04881139099597931,
-0.07977902144193649,
0.02233034372329712,
-0.024372870102524757,
-0.0019430865067988634,
0.0108978021889925,
-0.08440304547548294,
0.04476257041096687,
0.012540769763290882,
0.021930478513240814,
-0.053316812962293625,
-0.03381246700882912,
0.012533849105238914,
-0.13629499077796936,
-0.02845150977373123,
-0.05441470816731453,
0.09475218504667282,
0.04571226239204407,
0.009997174143791199,
0.012609751895070076,
-0.056740712374448776,
-0.02934388257563114,
0.08601507544517517,
-0.01828751713037491,
0.01560825016349554,
0.04205431416630745,
-0.045342568308115005,
0.06459017097949982,
-0.024821540340781212,
-0.025776168331503868,
0.05198362097144127,
-0.02040954679250717,
0.0354054793715477,
0.04532286152243614,
0.06822371482849121,
0.011501631699502468,
0.0002803026873152703,
0.035019371658563614,
0.023244015872478485,
0.012686729431152344,
0.0288683008402586,
0.0866180881857872,
0.04763460159301758,
0.024609703570604324,
-0.06594767421483994,
-0.00037846778286620975,
-0.016796888783574104,
0.047228842973709106,
0.025560908019542694,
0.01591610535979271,
0.0047519151121377945,
0.0444340743124485,
0.05008259415626526,
0.060747724026441574,
0.01991407386958599,
-0.0014508133754134178,
-0.06105444207787514,
0.05220856890082359,
-0.04216831177473068,
0.11223258078098297,
0.03292663395404816,
0.03247372433543205,
-0.12703122198581696,
-0.027852961793541908,
0.08180905133485794,
0.005239451769739389,
-0.018927091732621193,
-0.020530223846435547,
-0.10281993448734283,
0.031870778650045395,
-0.08049295842647552,
-0.06264425814151764,
0.0240927841514349,
-0.06086023524403572,
-0.07374364137649536,
-0.029284318909049034,
-0.04764368757605553,
0.0259336419403553,
-1.3796617401837874e-33,
-0.03032330423593521,
-0.07987648248672485,
0.017463475465774536,
0.03105117939412594,
-0.06431153416633606,
-0.04137682914733887,
-0.03341701999306679,
-0.06227326765656471,
-0.02738962508738041,
-0.0493851900100708,
-0.07231698930263519,
0.072319395840168,
0.04024338722229004,
-0.007518463302403688,
0.04864915832877159,
-0.00710215512663126,
0.02024025470018387,
0.04212338104844093,
0.01906518265604973,
0.04339032620191574,
-0.005592773202806711,
0.11830876767635345,
-0.03837994858622551,
-0.001442938344553113,
-0.050821345299482346,
0.0047094011679291725,
-0.028774501755833626,
0.03633563220500946,
0.026975024491548538,
-0.06389644742012024,
0.0005794475437141955,
0.016921546310186386,
-0.04543960839509964,
-0.046222835779190063,
-0.026165412738919258,
0.03267795965075493,
0.09713122993707657,
0.018408963456749916,
-0.035566799342632294,
0.11595359444618225,
-0.020558102056384087,
-0.007579163648188114,
-0.017411144450306892,
0.01659228652715683,
-0.02612830325961113,
0.07874013483524323,
-0.05533711984753609,
0.026330970227718353,
0.043377846479415894,
0.04866313934326172,
0.048130422830581665,
0.0017028910806402564,
-0.012707417830824852,
-0.014022433198988438,
-0.018892981112003326,
-0.04445018991827965,
-0.03964347764849663,
0.007835246622562408,
-0.01743253692984581,
0.006277004722505808,
-0.04949571564793587,
-0.014315119944512844,
-0.0344185009598732,
-0.03568427637219429,
-0.046080444008111954,
-0.07654628157615662,
-0.06468451023101807,
-0.017164992168545723,
-0.03859966620802879,
-0.002876011189073324,
-0.02053030952811241,
-0.08887539058923721,
-0.03540389612317085,
0.025831544771790504,
-0.055815353989601135,
0.06379018723964691,
0.0749172568321228,
-0.06430608034133911,
0.017078226432204247,
-0.03227526322007179,
0.03156657516956329,
0.03197463974356651,
0.06767888367176056,
0.05161812901496887,
0.013503959402441978,
0.06302723288536072,
-0.014491911977529526,
0.08250289410352707,
0.044659581035375595,
0.010135982185602188,
-0.10351663827896118,
-0.01092235092073679,
0.04770943894982338,
-0.03612427040934563,
-0.0026412622537463903,
-5.122350899000594e-8,
-0.07210883498191833,
0.04878205060958862,
-0.01981428824365139,
0.03861648589372635,
0.01232451107352972,
0.02603478915989399,
-0.020845701918005943,
0.11703113466501236,
-0.06156754121184349,
-0.010320651344954967,
0.10163821280002594,
0.010232355445623398,
-0.034694280475378036,
0.07154343277215958,
0.06804210692644119,
0.00593971274793148,
-0.020582232624292374,
-0.09385732561349869,
-0.04636221379041672,
0.04888860136270523,
-0.044018153101205826,
0.08228765428066254,
-0.002999071031808853,
-0.006831081118434668,
0.0069742086343467236,
-0.0806303471326828,
0.006017770152539015,
0.07146075367927551,
0.10766205191612244,
0.014935794286429882,
0.027088720351457596,
0.03530252352356911,
-0.08037794381380081,
0.03982545807957649,
0.037988390773534775,
0.00547625171020627,
-0.04734767973423004,
0.05772998183965683,
0.016978472471237183,
-0.04337120056152344,
0.08236552774906158,
0.011140985414385796,
-0.053795456886291504,
0.007478875108063221,
0.0060943132266402245,
0.01020857784897089,
-0.006802942603826523,
-0.06166117265820503,
0.00009586257510818541,
-0.08197150379419327,
-0.013490448705852032,
-0.04484768584370613,
-0.1394461691379547,
0.044554103165864944,
0.06041162088513374,
-0.006269215606153011,
-0.08186019957065582,
-0.054805923253297806,
0.014663361012935638,
0.018865447491407394,
0.027884358540177345,
-0.02775174006819725,
-0.05091263726353645,
-0.01313307136297226
] | 0 | 0 | 0 | 0 |
|
Aioli: A Unified Optimization Framework for Language Model Data Mixing | https://openreview.net/forum?id=sZGZJhaNSe | [
"Mayee F Chen",
"Michael Y. Hu",
"Nicholas Lourie",
"Kyunghyun Cho",
"Christopher Re"
] | Poster | Language model performance depends on identifying the optimal mixture of data groups to train on (e.g., law, code, math). Prior work has proposed a diverse set of methods to efficiently learn mixture proportions, ranging from fitting regression models over training runs to dynamically updating proportions throughout training. Surprisingly, we find that no existing method consistently outperforms a simple stratified sampling baseline in terms of average test perplexity. To understand this inconsistency, we unify existing methods into a standard framework, showing they are equivalent to solving a common optimization problem: minimize average loss subject to a method-specific mixing law---an implicit assumption on the relationship between loss and mixture proportions. This framework suggests that measuring the fidelity of a method's mixing law can offer insights into its performance. Empirically, we find that existing methods set their mixing law parameters inaccurately, resulting in the inconsistent mixing performance we observe. Using this insight, we derive a new online method named Aioli, which directly estimates the mixing law parameters throughout training and uses them to dynamically adjust proportions. Empirically, Aioli outperforms stratified sampling on 6 out of 6 datasets by an average of 0.27 test perplexity points, whereas existing methods fail to consistently beat stratified sampling, doing up to 6.9 points worse. Moreover, in a practical setting where proportions are learned on shorter runs due to computational constraints, Aioli can dynamically adjust these proportions over the full training run, consistently improving performance over existing methods by up to 12.012 test perplexity points. | data mixing, language models, data curation | null | 12,511 | 2411.05735 | [
-0.006999945733696222,
-0.06782862544059753,
-0.00253066373988986,
0.07720271497964859,
0.0869247317314148,
-0.004936046898365021,
0.003468095790594816,
0.013277149759232998,
0.015640420839190483,
-0.08925934880971909,
-0.03199790045619011,
-0.06032349169254303,
0.01710377261042595,
-0.018107308074831963,
0.03718065097928047,
-0.06140092387795448,
0.050709325820207596,
0.17489106953144073,
-0.1319633275270462,
-0.09171365201473236,
-0.04152112081646919,
0.05777706578373909,
0.03292408958077431,
0.0071790786460042,
0.10505416244268417,
-0.0016153580509126186,
-0.0550609789788723,
-0.06571556627750397,
0.05954984948039055,
-0.01578202284872532,
0.03856789693236351,
0.066224105656147,
0.024925360456109047,
0.01760808564722538,
-0.0762040764093399,
-0.030230240896344185,
-0.0548420213162899,
0.008677182719111443,
0.04621569439768791,
0.03252318128943443,
-0.06473533064126968,
0.04759959876537323,
0.03500599414110184,
0.023872343823313713,
-0.018983250483870506,
0.022825919091701508,
-0.004886607173830271,
0.013442939147353172,
-0.0038590305484831333,
0.029054732993245125,
-0.09659687429666519,
-0.08238571882247925,
-0.018258579075336456,
0.04997348412871361,
-0.05307641625404358,
-0.03780156001448631,
0.056301359087228775,
0.06908396631479263,
0.01709163561463356,
0.015919208526611328,
-0.058789558708667755,
-0.10712403804063797,
-0.054614271968603134,
0.00596983078867197,
-0.04083991423249245,
-0.028171738609671593,
0.05220282822847366,
0.056542541831731796,
-0.035143204033374786,
0.07460307329893112,
-0.025997059419751167,
0.09809572249650955,
-0.021541908383369446,
0.05898311361670494,
0.039991214871406555,
0.09454229474067688,
0.06630765646696091,
-0.003313987050205469,
0.06298354268074036,
-0.04793558269739151,
-0.0176823902875185,
0.009036141447722912,
0.004708870314061642,
-0.05062281712889671,
0.1390397548675537,
0.00539036002010107,
0.02448972687125206,
0.002580746542662382,
0.060745008289813995,
-0.03896745666861534,
-0.057160090655088425,
0.04731258749961853,
0.005973409861326218,
-0.04398464038968086,
0.08861320465803146,
0.07139196246862411,
-0.009476521983742714,
-0.05486446991562843,
0.07808620482683182,
0.09349215775728226,
-0.04103395715355873,
0.10505390167236328,
0.027377059683203697,
-0.03389822319149971,
-0.04211968928575516,
-0.09916345775127411,
0.03156835585832596,
0.011526206508278847,
0.034371938556432724,
-0.028494860976934433,
0.04202256724238396,
0.060630396008491516,
0.011771230958402157,
0.014004851691424847,
0.03685753792524338,
0.03384463116526604,
0.04921792820096016,
-0.016104400157928467,
-0.019750848412513733,
0.10848905146121979,
-0.046732403337955475,
0.06789949536323547,
-0.0019931807182729244,
0.02999895252287388,
0.022964773699641228,
0.016318127512931824,
-0.041278012096881866,
1.7067452473519966e-33,
0.04423417150974274,
0.007395347114652395,
0.01170159038156271,
0.04398585110902786,
-0.008196566253900528,
-0.052519410848617554,
0.0003180464555043727,
0.019889818504452705,
0.06634999811649323,
0.01991737261414528,
-0.022209422662854195,
0.01728600636124611,
-0.06717168539762497,
0.10066188871860504,
0.01873132400214672,
-0.007668488658964634,
-0.06440438330173492,
0.11744600534439087,
-0.06123092770576477,
-0.004480467177927494,
0.07618315517902374,
-0.01206129789352417,
-0.011873232200741768,
-0.06483212858438492,
-0.06531637907028198,
0.037845782935619354,
0.06471984088420868,
0.03164896368980408,
-0.016375459730625153,
0.01433656457811594,
-0.10515157133340836,
-0.03397524356842041,
-0.023539742454886436,
0.07446581870317459,
0.06051374599337578,
-0.017954738810658455,
-0.032200127840042114,
0.01726580411195755,
0.008694377727806568,
0.051923967897892,
-0.04182078316807747,
0.035218678414821625,
0.06438368558883667,
-0.030514251440763474,
-0.08364725857973099,
-0.10044517368078232,
-0.03807394206523895,
0.006884865462779999,
0.03989439457654953,
-0.05022088438272476,
-0.009436269290745258,
0.0010973545722663403,
0.04311516880989075,
0.004200228489935398,
-0.011313113383948803,
0.02627761848270893,
-0.01725338213145733,
0.016246644780039787,
0.03750280290842056,
0.04704349860548973,
-0.06177746132016182,
0.04473919793963432,
-0.009264551103115082,
0.07220952957868576,
0.031917132437229156,
0.055398862808942795,
-0.04198996350169182,
0.03929268568754196,
0.041104793548583984,
-0.05981459468603134,
0.05234729126095772,
-0.011165681295096874,
-0.08820068836212158,
-0.016521988436579704,
0.025635408237576485,
-0.0587051697075367,
0.06000206992030144,
-0.03499739244580269,
-0.0013643066631630063,
-0.01168687455356121,
0.006427055690437555,
0.07811373472213745,
-0.008938192389905453,
-0.11733075231313705,
-0.09103065729141235,
-0.012862462550401688,
0.051263291388750076,
-0.03720144182443619,
-0.035591792315244675,
-0.02910851314663887,
-0.06904525309801102,
-0.07252825796604156,
-0.01005963422358036,
-0.04368756711483002,
0.07076887786388397,
-1.712100175132969e-33,
-0.06080591678619385,
0.00041995171341113746,
-0.004204274620860815,
0.0647115707397461,
-0.06643825024366379,
-0.021623976528644562,
0.06892086565494537,
0.020574968308210373,
0.016590045765042305,
-0.040161021053791046,
-0.006756511051207781,
-0.06000955402851105,
0.006753412075340748,
-0.013585650362074375,
0.015720080584287643,
0.0011252033291384578,
-0.05072222650051117,
0.106924869120121,
0.10317696630954742,
0.06966518610715866,
-0.001229069079272449,
0.013883575797080994,
-0.03222408518195152,
0.0014750728150829673,
-0.055624209344387054,
-0.043323714286088943,
-0.07084591686725616,
0.09207276999950409,
0.013868553563952446,
-0.040172234177589417,
0.00403708778321743,
-0.08174826204776764,
-0.10739903151988983,
-0.044851504266262054,
-0.12290511280298233,
0.031908079981803894,
0.03911419212818146,
0.011246184818446636,
0.020326906815171242,
0.07949727028608322,
0.02393483556807041,
-0.036707062274217606,
-0.0586559996008873,
0.015063204802572727,
-0.007199211046099663,
-0.026475325226783752,
-0.046588391065597534,
-0.06173478066921234,
0.002935196040198207,
-0.06967304646968842,
0.00890885666012764,
0.033128537237644196,
-0.086858831346035,
0.03125649318099022,
-0.02483293041586876,
-0.038818635046482086,
-0.019283507019281387,
-0.050657402724027634,
-0.028345700353384018,
0.011439778842031956,
-0.06747913360595703,
0.02650010585784912,
-0.041184164583683014,
-0.06299053877592087,
-0.020017972216010094,
-0.10896331816911697,
-0.036959096789360046,
-0.04530755430459976,
0.039250995963811874,
0.05131443962454796,
0.02882981300354004,
-0.044859252870082855,
0.027034183964133263,
0.04671984538435936,
-0.0216306634247303,
-0.002988893073052168,
0.04564778134226799,
-0.02139192447066307,
-0.030997121706604958,
-0.053365640342235565,
0.015327667817473412,
0.0021374861244112253,
-0.02174917422235012,
0.011609559878706932,
0.017710519954562187,
0.07862462103366852,
0.006575408857315779,
0.054301343858242035,
-0.007203038316220045,
0.04015009105205536,
-0.046770624816417694,
-0.0348731093108654,
0.039952151477336884,
0.10396987944841385,
-0.06085098907351494,
-4.590087954170485e-8,
0.047501951456069946,
0.07244950532913208,
0.026293393224477768,
0.07769840955734253,
-0.035166144371032715,
-0.011012748815119267,
-0.03800997510552406,
0.040997523814439774,
-0.02693926729261875,
-0.0363663025200367,
0.035191114991903305,
0.019983474165201187,
0.0013547410489991307,
-0.06514854729175568,
-0.06937400996685028,
0.050222065299749374,
-0.04005618765950203,
0.026488058269023895,
-0.010351713746786118,
-0.001078035798855126,
0.023758305236697197,
0.026772627606987953,
0.03911099582910538,
-0.08043885976076126,
0.06599768251180649,
-0.02887856401503086,
0.05884135514497757,
0.11845243722200394,
0.04335371404886246,
-0.05109405890107155,
-0.034177955240011215,
-0.00024403342104051262,
-0.07859766483306885,
-0.09280164539813995,
-0.021662060171365738,
0.04914126545190811,
0.020096004009246826,
-0.058952368795871735,
0.007863902486860752,
0.010128742083907127,
0.019065750762820244,
0.02597767300903797,
-0.0324527882039547,
-0.009364592842757702,
0.09954962879419327,
0.03610643744468689,
-0.025863084942102432,
-0.06814529001712799,
0.03947967290878296,
-0.021769162267446518,
0.014825642108917236,
-0.015368610620498657,
-0.01892886683344841,
0.06631148606538773,
0.10873061418533325,
-0.05204762890934944,
-0.074094258248806,
-0.010365428403019905,
0.03175919130444527,
0.09813342243432999,
-0.00940132886171341,
0.054433003067970276,
-0.04567459225654602,
-0.01504797674715519
] | https://github.com/hazyresearch/aioli | 23 | 0 | 0 | 0 |
Differentiable Optimization of Similarity Scores Between Models and Brains | https://openreview.net/forum?id=vWRwdmA3wU | [
"Nathan Cloos",
"Moufan Li",
"Markus Siegel",
"Scott L Brincat",
"Earl K Miller",
"Guangyu Robert Yang",
"Christopher J Cueva"
] | Poster | How do we know if two systems - biological or artificial - process information in a similar way? Similarity measures such as linear regression, Centered Kernel Alignment (CKA), Normalized Bures Similarity (NBS), and angular Procrustes distance, are often used to quantify this similarity. However, it is currently unclear what drives high similarity scores and even what constitutes a "good" score. Here, we introduce a novel tool to investigate these questions by differentiating through similarity measures to directly maximize the score. Surprisingly, we find that high similarity scores do not guarantee encoding task-relevant information in a manner consistent with neural data; and this is particularly acute for CKA and even some variations of cross-validated and regularized linear regression. We find no consistent threshold for a good similarity score - it depends on both the measure and the dataset. In addition, synthetic datasets optimized to maximize similarity scores initially learn the highest variance principal component of the target dataset, but some methods like angular Procrustes capture lower variance dimensions much earlier than methods like CKA. To shed light on this, we mathematically derive the sensitivity of CKA, angular Procrustes, and NBS to the variance of principal component dimensions, and explain the emphasis CKA places on high variance components. Finally, by jointly optimizing multiple similarity measures, we characterize their allowable ranges and reveal that some similarity measures are more constraining than others. While current measures offer a seemingly straightforward way to quantify the similarity between neural systems, our work underscores the need for careful interpretation. We hope the tools we developed will be used by practitioners to better understand current and future similarity measures. | similarity measures, representational alignment, procrustes distance, centered kernel alignment, linear regression | Not all metrics for representational alignment are created equal; we show limitations in similarity metrics between models and brains by maximizing similarity with gradient descent. | 12,509 | 2407.07059 | [
-0.07650027424097061,
-0.0982760563492775,
-0.05872542783617973,
-0.04253974184393883,
-0.006545086391270161,
0.041124921292066574,
-0.021356601268053055,
-0.005044662859290838,
0.08765765279531479,
-0.02174372225999832,
-0.04108983278274536,
-0.027688952162861824,
-0.001134169870056212,
0.08176117390394211,
-0.055411484092473984,
-0.04722866043448448,
0.06852667033672333,
0.10057111084461212,
-0.12804700434207916,
-0.0007480458589270711,
0.017955759540200233,
0.047861918807029724,
0.0128210773691535,
0.08237917721271515,
0.03257576376199722,
0.02824907749891281,
0.05223745480179787,
-0.0012682749656960368,
-0.05388450250029564,
-0.03951551392674446,
0.0745605081319809,
-0.046153467148542404,
-0.0007197926170192659,
0.0034070529509335756,
-0.02453775890171528,
0.12092145532369614,
-0.08680710941553116,
0.023160548880696297,
0.03437238559126854,
-0.11771118640899658,
0.029276695102453232,
-0.00872401800006628,
-0.011705840937793255,
0.06965837627649307,
0.0490436814725399,
0.06725974380970001,
-0.062435269355773926,
-0.10983531177043915,
0.006266433279961348,
-0.009384095668792725,
-0.12743660807609558,
-0.02974439226090908,
-0.03749341145157814,
0.07655411213636398,
0.03492467477917671,
0.060638196766376495,
0.01371272187680006,
0.020915012806653976,
0.07156111299991608,
0.011516217142343521,
-0.05734666809439659,
-0.12495492398738861,
0.010662042535841465,
0.008103263564407825,
0.03619005158543587,
-0.02380414307117462,
0.0006845872267149389,
-0.04263593629002571,
-0.10779795050621033,
0.002856573788449168,
-0.07984001189470291,
0.14758403599262238,
-0.04833924397826195,
0.06474512815475464,
0.05849673971533775,
0.07181026041507721,
-0.027169546112418175,
0.030022239312529564,
0.013483432121574879,
-0.07927855104207993,
0.00617551663890481,
0.07275194674730301,
0.06883417814970016,
0.05221747234463692,
0.12018212676048279,
0.05606019124388695,
0.03710950165987015,
0.017235837876796722,
0.027995213866233826,
-0.03854391351342201,
-0.0175155159085989,
-0.05655341222882271,
-0.007772041950374842,
-0.046782344579696655,
0.0015297109493985772,
0.01736355386674404,
-0.025712620466947556,
-0.036295149475336075,
0.06645775586366653,
0.03191550076007843,
-0.016317445784807205,
0.06241631880402565,
-0.020656311884522438,
0.007497433573007584,
0.08904818445444107,
0.017377274110913277,
0.04321539029479027,
-0.00032662146259099245,
0.06546217203140259,
-0.09381687641143799,
0.003793553449213505,
0.0422791987657547,
-0.035071082413196564,
0.0959007665514946,
0.03535362705588341,
-0.07258064299821854,
0.04242300987243652,
-0.008849667385220528,
0.07064340263605118,
0.02662883698940277,
-0.018225744366645813,
-0.03808618709445,
-0.044201724231243134,
-0.011440536938607693,
-0.006284146104007959,
-0.0009655901812948287,
-0.08765273541212082,
4.161461223552061e-33,
0.005879574920982122,
0.07128950208425522,
0.030723189935088158,
-0.037094492465257645,
-0.00479214359074831,
-0.10596800595521927,
-0.11098001152276993,
0.001135970582254231,
0.050749972462654114,
0.051703955978155136,
-0.1015288308262825,
0.08463901281356812,
-0.054956503212451935,
0.047854095697402954,
0.018623508512973785,
0.05881661921739578,
-0.08472894877195358,
0.028231684118509293,
-0.049287375062704086,
-0.048237938433885574,
0.05526474863290787,
-0.04476197063922882,
-0.006795667577534914,
0.008334154263138771,
-0.004736218135803938,
-0.056122392416000366,
-0.031530097126960754,
-0.0797543153166771,
-0.06478437036275864,
-0.03648090362548828,
-0.06865587830543518,
0.0036556264385581017,
0.010897253639996052,
0.0033889359328895807,
0.03177371993660927,
-0.01981830783188343,
0.05556419491767883,
-0.06451735645532608,
-0.0167967788875103,
0.04225461557507515,
0.09593039005994797,
0.05108008161187172,
0.001605503959581256,
-0.02904953993856907,
0.033109135925769806,
-0.02673310786485672,
-0.021380724385380745,
-0.03346433863043785,
0.06288766860961914,
-0.11160558462142944,
0.030835764482617378,
0.00736120855435729,
0.028325585648417473,
-0.08056773990392685,
-0.07721918821334839,
0.0007634515059180558,
-0.007063417695462704,
-0.06382197886705399,
0.04941080883145332,
0.08541010320186615,
-0.02383273094892502,
-0.027270331978797913,
0.04648276045918465,
-0.027987033128738403,
0.027543535456061363,
-0.0330091193318367,
-0.051722973585128784,
0.029345015063881874,
0.016180479899048805,
0.039121001958847046,
0.008062978275120258,
0.014144456945359707,
-0.04695499315857887,
-0.09748148918151855,
0.03329353407025337,
-0.00812679436057806,
0.04183967411518097,
-0.027460716664791107,
-0.024651145562529564,
-0.014079944230616093,
-0.05665533244609833,
0.06462498009204865,
-0.05444195121526718,
-0.09420329332351685,
-0.07374870032072067,
-0.0792873352766037,
0.026560232043266296,
-0.1094290241599083,
-0.09936937689781189,
0.0014143966836854815,
0.027597885578870773,
0.00952194258570671,
-0.013653266243636608,
-0.030926549807190895,
-0.027647696435451508,
-3.633766381104582e-33,
-0.07319881021976471,
0.003737241728231311,
0.022039039060473442,
0.07008158415555954,
-0.04419584572315216,
0.009261726401746273,
0.007348014507442713,
0.01713130809366703,
-0.08262322098016739,
-0.01671750843524933,
0.04268483445048332,
-0.03395135700702667,
0.07599441707134247,
-0.0013430459657683969,
0.0319768562912941,
0.03127005323767662,
0.003176687052473426,
0.1281827688217163,
0.0319330058991909,
-0.014229560270905495,
0.06410309672355652,
0.0844602957367897,
-0.06775157898664474,
0.03581240028142929,
-0.010154766030609608,
0.06375452876091003,
-0.027242835611104965,
0.08125284314155579,
-0.043912049382925034,
0.012266571633517742,
-0.04155630245804787,
0.030371084809303284,
-0.09302231669425964,
-0.020215746015310287,
-0.06001435965299606,
0.037124842405319214,
-0.03036939725279808,
-0.031359776854515076,
0.0025668812450021505,
0.06679068505764008,
0.021131500601768494,
0.042229682207107544,
-0.08034767210483551,
-0.0427546463906765,
0.009299892000854015,
-0.037691231817007065,
0.01690082997083664,
-0.06213613227009773,
0.025418376550078392,
-0.0018693279707804322,
0.08195184916257858,
0.09027896821498871,
-0.09246383607387543,
-0.0024468021001666784,
-0.052720729261636734,
-0.07545320689678192,
-0.03717143461108208,
-0.02206542156636715,
0.04039185121655464,
-0.0011997068068012595,
0.018009381368756294,
0.011137759312987328,
0.011107264086604118,
0.07489580661058426,
0.004501726943999529,
0.04161287099123001,
0.024550167843699455,
0.0014135089004412293,
0.02160407416522503,
0.042267877608537674,
-0.022167211398482323,
-0.03749481216073036,
0.09362653642892838,
0.03801891952753067,
-0.03567003831267357,
-0.02669759839773178,
0.04273082688450813,
-0.010499470867216587,
-0.027114881202578545,
0.02906091883778572,
0.02551843412220478,
-0.011418212205171585,
-0.012173783034086227,
-0.00022305977472569793,
0.012658070772886276,
0.09945055842399597,
0.04423631355166435,
0.01442552637308836,
0.03911493346095085,
0.014785164035856724,
-0.03346318006515503,
-0.025123486295342445,
-0.0072605907917022705,
0.006590224336832762,
-0.048444777727127075,
-5.053215090811136e-8,
-0.05054144561290741,
-0.007456621155142784,
-0.030159462243318558,
0.04474705085158348,
-0.05438991263508797,
-0.05796432122588158,
-0.02333059348165989,
0.05030713975429535,
-0.10778728872537613,
-0.01754465140402317,
0.036058422178030014,
0.0032116465736180544,
-0.003564547747373581,
0.04551015421748161,
0.038233205676078796,
0.0757785439491272,
-0.0005039132083766162,
0.06812729686498642,
0.017096323892474174,
0.062426529824733734,
0.003716593375429511,
0.057064104825258255,
-0.020688530057668686,
-0.02055433765053749,
0.08026644587516785,
-0.0669841319322586,
-0.015530765056610107,
0.05448070913553238,
0.06587880849838257,
0.05194828659296036,
-0.013608070090413094,
0.020223041996359825,
0.059281300753355026,
-0.0732232928276062,
-0.00939241237938404,
0.007480258587747812,
0.0018168193055316806,
-0.006820183712989092,
-0.06474058330059052,
0.061939071863889694,
0.027105452492833138,
0.07777126133441925,
-0.04912896454334259,
0.0017091092886403203,
0.07456818222999573,
0.03728432208299637,
0.0584135502576828,
-0.0011205432238057256,
0.002926739864051342,
-0.05327719822525978,
0.08179110288619995,
-0.015746450051665306,
0.03238000348210335,
0.023842427879571915,
0.044365931302309036,
-0.035109855234622955,
-0.05926937609910965,
-0.026237022131681442,
0.03292124345898628,
-0.03145070746541023,
0.0706537514925003,
0.015052489936351776,
0.004065534565597773,
-0.030491670593619347
] | https://github.com/nacloos/diffscore | 6 | 0 | 0 | 0 |
Provable Convergence Bounds for Hybrid Dynamical Sampling and Optimization | https://openreview.net/forum?id=FJv8VMPxWi | [
"Matthew X. Burns",
"Qingyuan Hou",
"Michael Huang"
] | Poster | Analog dynamical accelerators (DXs) are a growing sub-field in computer architecture research, offering order-of-magnitude gains in power efficiency and latency over traditional digital methods in several machine learning, optimization, and sampling tasks. However, limited-capacity accelerators require hybrid analog/digital algorithms to solve real-world problems, commonly using large-neighborhood local search (LNLS) frameworks. Unlike fully digital algorithms, hybrid LNLS has no non-asymptotic convergence guarantees and no principled hyperparameter selection schemes, particularly limiting cross-device training and inference.
In this work, we provide non-asymptotic convergence guarantees for hybrid LNLS by reducing to block Langevin Diffusion (BLD) algorithms.
Adapting tools from classical sampling theory, we prove exponential KL-divergence convergence for randomized and cyclic block selection strategies using ideal DXs. With finite device variation, we provide explicit bounds on the 2-Wasserstein bias in terms of step duration, noise strength, and function parameters. Our BLD model provides a key link between established theory and novel computing platforms, and our theoretical results provide a closed-form expression linking device variation, algorithm hyperparameters, and performance. | langevin, accelerators, sampling, optimization, diffusion, analog computing | We provide novel non-asymptotic convergence rates for a class of hybrid analog-digital algorithms for dynamics-accelerated activation sampling and inference. | 12,494 | null | [
-0.07052818685770035,
-0.029138240963220596,
0.03336712345480919,
0.017138570547103882,
0.031923044472932816,
-0.10707895457744598,
-0.06985805928707123,
-0.03588615357875824,
0.037223417311906815,
-0.030805589631199837,
-0.06730274111032486,
-0.04948920011520386,
-0.02256011962890625,
-0.02731877751648426,
0.01362727489322424,
0.050555527210235596,
0.05895696580410004,
0.10280773043632507,
-0.035083040595054626,
-0.025370284914970398,
0.014015854336321354,
0.043214909732341766,
0.03962940722703934,
0.009201857261359692,
0.01222779881209135,
-0.03135773912072182,
0.051213908940553665,
-0.059155456721782684,
0.036497119814157486,
0.003428600961342454,
0.05964477360248566,
-0.06444599479436874,
-0.04669463261961937,
-0.04619161784648895,
0.036549702286720276,
-0.008481031283736229,
-0.03357864171266556,
-0.009025038219988346,
0.07625680416822433,
0.001604081830009818,
-0.004252236802130938,
0.019695945084095,
0.012202800251543522,
0.009259342215955257,
0.07071990519762039,
-0.04478505998849869,
-0.03402320295572281,
-0.0310226883739233,
0.0024058748967945576,
0.04669317603111267,
0.03979431465268135,
0.0030673318542540073,
-0.07462470233440399,
0.019153650850057602,
-0.019263852387666702,
-0.03342364355921745,
0.043933670967817307,
0.03991595655679703,
0.03922808915376663,
0.048945143818855286,
0.005525620188564062,
-0.1204480230808258,
0.011090821586549282,
-0.012857873924076557,
-0.052637770771980286,
-0.04523102566599846,
0.053000181913375854,
0.06706753373146057,
0.07565217465162277,
0.04400480166077614,
-0.0567157007753849,
0.051546044647693634,
-0.028005555272102356,
0.08441288024187088,
0.06971878558397293,
0.03654855489730835,
0.024334844201803207,
0.01852010004222393,
0.06195317208766937,
-0.11222110688686371,
0.013561693020164967,
-0.017099538818001747,
-0.019288890063762665,
-0.03762448951601982,
0.01548637356609106,
0.03674731403589249,
-0.05206083878874779,
0.07428929209709167,
0.15844136476516724,
-0.10788406431674957,
0.006431304384022951,
0.029762445017695427,
-0.11137843132019043,
-0.05391781032085419,
0.07150101661682129,
-0.08417958766222,
0.003961329814046621,
-0.022731810808181763,
-0.0059189870953559875,
0.10174049437046051,
0.001922333613038063,
0.07862931489944458,
0.01427000854164362,
0.01563250459730625,
-0.0024131021928042173,
-0.12029573321342468,
0.03916793689131737,
0.0624961294233799,
0.009202609769999981,
-0.07924893498420715,
0.08057861030101776,
0.04778944328427315,
0.07988647371530533,
0.06598960608243942,
-0.001011667656712234,
-0.06245890259742737,
0.01911282166838646,
-0.0825686901807785,
0.05211310461163521,
0.010311526246368885,
-0.10896509885787964,
-0.029529452323913574,
-0.10511355847120285,
0.029353203251957893,
0.05590192228555679,
0.028568662703037262,
-0.0186716690659523,
1.8096120233084858e-33,
0.03674142807722092,
0.029468566179275513,
0.09374444931745529,
-0.01270399708300829,
0.023033244535326958,
-0.04684009030461311,
0.0725780799984932,
-0.02708515338599682,
-0.00977175123989582,
0.048512186855077744,
-0.045045845210552216,
-0.034683577716350555,
-0.07394536584615707,
0.04446306452155113,
0.09970330446958542,
-0.01870841346681118,
-0.03289462625980377,
-0.014070766977965832,
0.003079599468037486,
-0.022625597193837166,
0.09505325555801392,
-0.011055640876293182,
-0.07455737888813019,
0.07226381450891495,
-0.00747747253626585,
0.006544237490743399,
0.0021228385157883167,
0.041771095246076584,
-0.013246704824268818,
0.037417370826005936,
-0.11876405030488968,
-0.005361475516110659,
-0.044039372354745865,
0.010776345618069172,
0.0685277208685875,
0.05050938203930855,
-0.006804108619689941,
-0.02785908803343773,
-0.023761825636029243,
0.010537667199969292,
-0.026023346930742264,
-0.007897315546870232,
0.03224986791610718,
-0.056366462260484695,
-0.1607544869184494,
-0.10885583609342575,
0.02485293336212635,
-0.015993455424904823,
-0.009798402898013592,
-0.05702585726976395,
-0.006903562694787979,
0.007394436281174421,
-0.03813483566045761,
-0.021538151428103447,
0.05165621265769005,
-0.008835316635668278,
0.010034547187387943,
0.011585830710828304,
0.09797791391611099,
0.1340566873550415,
-0.02449147403240204,
-0.01691928692162037,
-0.028985299170017242,
0.09671927243471146,
0.014754471369087696,
0.10475696623325348,
0.01602650247514248,
-0.012226788327097893,
0.040395308285951614,
-0.02770301140844822,
0.08598034083843231,
-0.018073277547955513,
0.004074865486472845,
-0.11169222742319107,
0.09203488379716873,
-0.03085804358124733,
0.01885261945426464,
-0.02107994630932808,
-0.07914821803569794,
-0.006186255719512701,
-0.0573684498667717,
-0.03507236763834953,
-0.0893595740199089,
-0.0023155901581048965,
0.00646544573828578,
-0.05138713866472244,
-0.042278725653886795,
-0.05327886715531349,
-0.06462029367685318,
-0.0909523218870163,
-0.063019759953022,
-0.007628436200320721,
-0.022922426462173462,
0.0255796667188406,
-0.052696604281663895,
-2.368665197945065e-33,
-0.010157808661460876,
-0.04328504577279091,
0.08082228899002075,
0.027789000421762466,
-0.0031361605506390333,
0.0010250204941257834,
-0.02860153838992119,
-0.026882193982601166,
0.05231323093175888,
-0.07460425794124603,
-0.038644179701805115,
0.02699255384504795,
0.019258443266153336,
0.012604554183781147,
0.03130875155329704,
0.09796779602766037,
-0.040682729333639145,
0.019529687240719795,
0.07169001549482346,
0.04507988318800926,
0.02432326227426529,
0.04455205053091049,
0.009725529700517654,
0.016937727108597755,
0.005480537656694651,
0.04952342063188553,
-0.07921238243579865,
0.08447571098804474,
0.04090720787644386,
-0.09045203775167465,
-0.08661310374736786,
-0.020655512809753418,
-0.051402363926172256,
0.007350926287472248,
-0.02030550129711628,
0.07712183892726898,
0.08270306140184402,
0.08200414478778839,
0.036631990224123,
0.08095332980155945,
0.010914585553109646,
-0.04022536054253578,
-0.0010782727040350437,
0.024818958714604378,
0.08734870702028275,
-0.014574958011507988,
-0.016908885911107063,
-0.001919283764436841,
0.030550019815564156,
0.0019679246470332146,
0.027812788262963295,
0.055815037339925766,
-0.024165531620383263,
0.03630758076906204,
-0.05027882009744644,
-0.023135339841246605,
-0.027662526816129684,
0.04609129577875137,
0.10234276950359344,
-0.006692209281027317,
-0.02875538356602192,
-0.05768783390522003,
-0.0524577796459198,
-0.006175607908517122,
-0.08131422102451324,
-0.010206727311015129,
-0.02991027571260929,
0.02548883855342865,
0.026097338646650314,
0.019768059253692627,
-0.06864596158266068,
-0.02365097589790821,
0.09396916627883911,
0.004792715422809124,
-0.0881953239440918,
-0.06446608901023865,
0.028007620945572853,
-0.07696738839149475,
0.027665147557854652,
-0.041608575731515884,
-0.02134774439036846,
0.048725955188274384,
0.041089411824941635,
-0.05081890895962715,
0.023038212209939957,
0.0184733085334301,
0.014077144674956799,
-0.012909948825836182,
-0.0036706908140331507,
-0.10241413116455078,
-0.012258388102054596,
-0.013751366175711155,
0.06196063011884689,
0.022307902574539185,
-0.06797736883163452,
-4.3457689002934785e-8,
0.055637698620557785,
0.06836015731096268,
0.042463596910238266,
-0.0060906908474862576,
0.10056930035352707,
-0.008629030548036098,
0.022506913170218468,
0.011728493496775627,
0.017887866124510765,
-0.031564708799123764,
0.15618610382080078,
-0.020587123930454254,
0.020363083109259605,
-0.01089223101735115,
-0.038870103657245636,
0.06065410375595093,
-0.04146602749824524,
-0.06014455854892731,
0.0016972392331808805,
-0.03652871772646904,
-0.05341637134552002,
0.05218934640288353,
0.05271192267537117,
-0.03956140950322151,
0.02986518293619156,
-0.01939203403890133,
-0.005675756838172674,
-0.07124055922031403,
0.11449642479419708,
-0.013108490966260433,
-0.05621857941150665,
0.046356864273548126,
0.016404932364821434,
0.03574182465672493,
0.03636961430311203,
0.012578943744301796,
-0.04211330786347389,
0.019952375441789627,
0.01734941452741623,
-0.005123508162796497,
-0.04577460139989853,
-0.015453750267624855,
-0.04660258814692497,
-0.029467837885022163,
-0.05568348243832588,
-0.05239629000425339,
0.03350603207945824,
-0.028368735685944557,
0.0122305229306221,
0.05437680706381798,
0.0408514142036438,
0.0011462668189778924,
0.0047217803075909615,
-0.03221381455659866,
0.06589222699403763,
-0.0356086902320385,
0.0024148670490831137,
-0.0536019504070282,
0.023494301363825798,
0.033378493040800095,
0.0036744666285812855,
-0.013343317434191704,
-0.047099679708480835,
0.014690876007080078
] | 0 | 0 | 0 | 0 |
|
CTSyn: A Foundation Model for Cross Tabular Data Generation | https://openreview.net/forum?id=Sh4FOyZRpv | [
"Xiaofeng Lin",
"Chenheng Xu",
"Matthew Yang",
"Guang Cheng"
] | Poster | Generative Foundation Models (GFMs) have achieved remarkable success in producing high-quality synthetic data for images and text. However, their application to tabular data presents significant challenges due to the heterogeneous nature of table features. Current cross-table learning frameworks struggle because they lack a generative model backbone and an effective mechanism to decode heterogeneous feature values. To address these challenges, we propose the Cross-Table Synthesizer (CTSyn), a diffusion-based generative foundation model for tabular data generation. CTSyn comprises two key components. The first is an autoencoder network that consolidates diverse tables into a unified latent space. It dynamically reconstructs table values using a table schema embedding, allowing adaptation to heterogeneous datasets. The second is a conditional latent diffusion model that generates samples from the learned latent space, conditioned on the table schema. Through large-scale pre-training, CTSyn outperforms existing table synthesizers on standard benchmarks in both utility and diversity. These results position CTSyn as a promising framework for synthetic table generation and lay the groundwork for developing large-scale tabular foundation models. | Foundation Model, Tabular Data, Synthetic Data Generation | null | 12,491 | 2406.04619 | [
-0.03533085808157921,
-0.12045729905366898,
0.00785092730075121,
0.025362808257341385,
0.07117900252342224,
-0.012293543666601181,
-0.05043349415063858,
-0.07419196516275406,
-0.05601218715310097,
-0.016614224761724472,
0.0054407608695328236,
-0.10842073708772659,
0.015988683328032494,
-0.0022417763248085976,
-0.0496579185128212,
-0.0175649244338274,
-0.018665483221411705,
0.06375937163829803,
-0.08200624585151672,
-0.04632954299449921,
-0.08318307250738144,
0.04554307088255882,
0.014005146920681,
-0.034158769994974136,
0.08785424381494522,
0.08053617924451828,
0.01871149055659771,
-0.03338812664151192,
0.03116275556385517,
-0.06473463028669357,
0.031151793897151947,
0.023328112438321114,
-0.08589281141757965,
0.04849745333194733,
-0.040887732058763504,
0.053999967873096466,
-0.0838441550731659,
0.010378646664321423,
0.02120591700077057,
-0.013483758084475994,
-0.024020738899707794,
-0.041617631912231445,
-0.0003203438245691359,
0.052166376262903214,
0.06095818802714348,
-0.007836001925170422,
-0.029595816507935524,
-0.07862421125173569,
-0.002069332404062152,
0.011700155213475227,
-0.0397598072886467,
0.034795332700014114,
-0.026458214968442917,
0.04318419471383095,
-0.006212243810296059,
0.03143493831157684,
0.0649193599820137,
-0.004870564676821232,
0.03138928860425949,
0.019132129848003387,
-0.09797420352697372,
-0.06849948316812515,
0.051365725696086884,
-0.07886817306280136,
-0.021761005744338036,
-0.02363135851919651,
-0.01593075878918171,
0.08329318463802338,
0.040295012295246124,
-0.04660700634121895,
-0.023198017850518227,
0.06601490825414658,
-0.06519503146409988,
0.01736835390329361,
-0.04944266006350517,
0.04445745423436165,
0.023528020828962326,
-0.08379240334033966,
0.03725447505712509,
-0.06150968372821808,
0.0113914730027318,
0.07955222576856613,
0.01347923744469881,
0.014870978891849518,
0.10727346688508987,
0.0121185053139925,
0.055444683879613876,
0.08690536022186279,
-0.043482206761837006,
-0.01664915680885315,
-0.01182618085294962,
0.011720637790858746,
0.0296647809445858,
0.01022136677056551,
-0.06508452445268631,
0.03510192781686783,
0.07799079269170761,
-0.09193949401378632,
0.036503516137599945,
0.0938907265663147,
-0.013085676357150078,
0.061253149062395096,
0.07437070459127426,
0.027654225006699562,
0.011085609905421734,
-0.07586976885795593,
0.10659955441951752,
0.04984182119369507,
0.018485087901353836,
-0.07422938942909241,
0.0036681347992271185,
0.04545923322439194,
-0.041345495730638504,
-0.08897149562835693,
-0.026662709191441536,
-0.032329775393009186,
-0.07419147342443466,
-0.09329944849014282,
0.08692154288291931,
0.04563598334789276,
-0.011256597004830837,
0.02083909511566162,
-0.15318948030471802,
0.11767734587192535,
-0.02722865343093872,
-0.06131862476468086,
-0.04871421679854393,
3.331251073210543e-33,
-0.042154330760240555,
0.02173587679862976,
0.06369990110397339,
0.04113630950450897,
0.11262141913175583,
-0.006177337374538183,
0.008070146664977074,
0.00006477795977843925,
0.012685260735452175,
-0.004948627203702927,
-0.08331112563610077,
0.01147423405200243,
-0.10737206786870956,
0.1265903115272522,
0.049094319343566895,
0.01778099499642849,
-0.09456613659858704,
-0.0067228409461677074,
0.01562681421637535,
0.0009635806200094521,
-0.0036415867507457733,
0.010925828479230404,
-0.053962592035532,
0.043218523263931274,
-0.010024021379649639,
0.02419896237552166,
0.04686003178358078,
-0.0015754287596791983,
0.009637940675020218,
0.00556889409199357,
-0.11699672788381577,
-0.025846391916275024,
0.03269902244210243,
0.016130851581692696,
0.0001510921138105914,
-0.02742963284254074,
-0.018488004803657532,
-0.020855829119682312,
0.034633416682481766,
0.054544318467378616,
0.038052938878536224,
0.06191033124923706,
-0.04172615334391594,
-0.017478620633482933,
-0.042021673172712326,
-0.006059650797396898,
0.10958231985569,
0.03461109846830368,
-0.04073745757341385,
-0.0612080879509449,
0.0689583495259285,
0.005495761055499315,
-0.01889849081635475,
0.024542536586523056,
0.006631441414356232,
-0.00039171139360405505,
0.00831716787070036,
-0.03162593021988869,
0.13632315397262573,
0.06099948659539223,
0.024990227073431015,
0.001330348663032055,
-0.03542530536651611,
0.08261101692914963,
-0.028158288449048996,
0.004270671401172876,
0.014248140156269073,
0.012588678859174252,
0.04158398509025574,
-0.03221795707941055,
-0.032467227429151535,
0.005204281769692898,
-0.04053410515189171,
-0.057929616421461105,
0.05664503946900368,
-0.016104135662317276,
0.02085232362151146,
-0.035829756408929825,
0.05557703971862793,
-0.06781212985515594,
0.0009500517626293004,
0.08869214355945587,
-0.08111171424388885,
-0.09095609933137894,
-0.07809261232614517,
-0.034536588937044144,
0.0012732527684420347,
-0.067087322473526,
-0.07327400147914886,
0.0012870498467236757,
-0.008108611218631268,
-0.06718424707651138,
0.03205469250679016,
-0.03740084543824196,
0.033075980842113495,
-3.059887100297264e-33,
0.0010205238359048963,
0.06675100326538086,
-0.046991102397441864,
0.003026388818398118,
-0.013978802599012852,
0.010380853898823261,
0.004649254493415356,
0.026527395471930504,
0.03437463194131851,
-0.009461724199354649,
-0.035763856023550034,
-0.04735628515481949,
0.028767229989171028,
-0.09627478569746017,
-0.013162962161004543,
0.0288985762745142,
0.03395993635058403,
0.041047725826501846,
-0.01711973547935486,
0.04702958092093468,
0.020540107041597366,
0.10113675147294998,
-0.04738539084792137,
0.061417002230882645,
-0.03976505249738693,
0.03081977181136608,
-0.0193776935338974,
0.04903673380613327,
0.007548212539404631,
0.0454723984003067,
-0.02309601567685604,
-0.0582021102309227,
0.02826906368136406,
-0.009009440429508686,
-0.08074551075696945,
0.04047100245952606,
0.03726717829704285,
-0.013701935298740864,
-0.03146001696586609,
0.0367940329015255,
0.03154682740569115,
-0.0066069262102246284,
-0.028626101091504097,
0.05713048577308655,
-0.06246576085686684,
0.05519061163067818,
-0.046269286423921585,
0.023785045370459557,
0.05244755744934082,
-0.010065000504255295,
0.048042040318250656,
0.02730841189622879,
-0.07585421204566956,
-0.07035139203071594,
-0.017094483599066734,
-0.04989933967590332,
-0.04374857619404793,
-0.029441354796290398,
0.005000264849513769,
-0.024287037551403046,
-0.04837796464562416,
-0.038963332772254944,
0.0923478901386261,
-0.08324228972196579,
0.0058051468804478645,
-0.09080692380666733,
-0.08543708920478821,
-0.006539200432598591,
-0.015473034232854843,
0.075874924659729,
-0.03190968185663223,
-0.019495848566293716,
0.021875377744436264,
0.028614500537514687,
0.03505999222397804,
-0.06774082034826279,
-0.02220289409160614,
-0.011585064232349396,
0.007993348874151707,
-0.08789605647325516,
0.03799663856625557,
-0.0202645156532526,
0.08544092625379562,
0.004266621079295874,
0.0025286064483225346,
0.09878256171941757,
0.03184713423252106,
0.009051503613591194,
0.03011215664446354,
0.060302138328552246,
-0.010539758950471878,
-0.0045679593458771706,
-0.06286892294883728,
0.10083936154842377,
-0.026341816410422325,
-3.705129714148825e-8,
-0.013878364115953445,
0.051505859941244125,
0.003088886383920908,
0.004303400404751301,
-0.02759796380996704,
-0.09647584706544876,
0.08371090888977051,
0.1202324628829956,
0.029251787811517715,
-0.034589845687150955,
0.10950326174497604,
0.0018438639817759395,
0.0008700083126313984,
-0.042715709656476974,
0.03775538504123688,
0.10880417376756668,
0.026202719658613205,
-0.004592593759298325,
0.008767529390752316,
0.046084094792604446,
0.023367710411548615,
0.08897130191326141,
-0.010734661482274532,
-0.029352163895964622,
0.0793081596493721,
-0.016432812437415123,
-0.08259197324514389,
0.06117655709385872,
0.054658081382513046,
-0.03715623915195465,
0.05996578186750412,
-0.01904965750873089,
-0.0406499020755291,
-0.027112271636724472,
0.0344163216650486,
0.07574497163295746,
0.08574334532022476,
-0.05537029355764389,
-0.06660231947898865,
-0.08517885208129883,
0.012045439332723618,
0.04724173992872238,
-0.08061608672142029,
-0.02524409256875515,
-0.04439026117324829,
-0.008312779478728771,
0.011891068890690804,
-0.046797752380371094,
0.059779517352581024,
-0.006647950038313866,
0.05644279345870018,
-0.08080761134624481,
-0.013299426063895226,
0.038137488067150116,
0.057539988309144974,
-0.017501801252365112,
0.030316278338432312,
0.012965359725058079,
0.056326258927583694,
-0.03163062036037445,
0.006521099247038364,
-0.014695386402308941,
0.048504192382097244,
0.04786710813641548
] | 0 | 0 | 0 | 0 |
|
Solving hidden monotone variational inequalities with surrogate losses | https://openreview.net/forum?id=4ZX2a3OKEV | [
"Ryan D'Orazio",
"Danilo Vucetic",
"Zichu Liu",
"Junhyung Lyle Kim",
"Ioannis Mitliagkas",
"Gauthier Gidel"
] | Poster | Deep learning has proven to be effective in a wide variety of loss minimization problems.
However, many applications of interest, like minimizing projected Bellman error and min-max optimization, cannot be modelled as minimizing a scalar loss function but instead correspond to solving a variational inequality (VI) problem.
This difference in setting has caused many practical challenges as naive gradient-based approaches from supervised learning tend to diverge and cycle in the VI case.
In this work, we propose a principled surrogate-based approach compatible with deep learning to solve VIs.
We show that our surrogate-based approach has three main benefits: (1) under assumptions that are realistic in practice (when hidden monotone structure is present, interpolation, and sufficient optimization of the surrogates), it guarantees convergence, (2) it provides a unifying perspective of existing methods, and (3) is amenable to existing deep learning optimizers like ADAM.
Experimentally, we demonstrate our surrogate-based approach is effective in min-max optimization and minimizing projected Bellman error. Furthermore, in the deep reinforcement learning case, we propose a novel variant of TD(0) which is more compute and sample efficient. | Variational Inequality, Optimization, Surrogate, Projected Bellman Error, Min-max Optimization | A novel surrogate loss approach to solving variational inequalities with function approximation. Both theoretical guarantees and empirical analysis is provided. | 12,488 | 2411.05228 | [
-0.09788866341114044,
-0.07598470151424408,
0.0866752490401268,
0.004215221386402845,
-0.012354006990790367,
0.04744422063231468,
-0.034657880663871765,
-0.039764948189258575,
-0.023643523454666138,
-0.02931213192641735,
-0.039660047739744186,
-0.032519567757844925,
0.018994910642504692,
0.049318164587020874,
-0.0746934562921524,
0.041924770921468735,
0.07457394152879715,
0.12888389825820923,
-0.09347938001155853,
0.004917170386761427,
0.04526076465845108,
0.0469786562025547,
0.023187711834907532,
0.017504209652543068,
-0.0162912979722023,
-0.027569251134991646,
-0.00970954354852438,
0.00023736512230243534,
0.042839065194129944,
0.008786185644567013,
0.11168364435434341,
-0.000826260366011411,
-0.012773792259395123,
-0.04815543815493584,
-0.023246193304657936,
0.09305740147829056,
-0.07791009545326233,
0.021589010953903198,
0.02296242117881775,
-0.018715985119342804,
-0.02663549967110157,
0.030933596193790436,
0.017080819234251976,
0.054715678095817566,
0.07059207558631897,
0.007558508776128292,
0.018398797139525414,
-0.0682402029633522,
0.044326696544885635,
-0.000850787153467536,
0.004860083572566509,
0.004187019541859627,
-0.02936273254454136,
0.047381896525621414,
-0.016813624650239944,
0.015040069818496704,
0.013136922381818295,
0.001244109240360558,
-0.04829246178269386,
0.06897663325071335,
0.04263191670179367,
-0.1372227519750595,
0.016841240227222443,
-0.04725927487015724,
0.0020426479168236256,
-0.0365520641207695,
0.0469089113175869,
0.00391270499676466,
-0.01844395510852337,
0.11100605130195618,
-0.002116726478561759,
0.04738565534353256,
-0.04312368482351303,
0.0023656932171434164,
0.012778294272720814,
0.01916690729558468,
0.15868441760540009,
-0.024542871862649918,
0.028501363471150398,
-0.03437292203307152,
0.04948300123214722,
0.003688039490953088,
-0.03811297565698624,
0.03251449763774872,
0.1421596109867096,
-0.02666386030614376,
0.007553604897111654,
0.023083122447133064,
0.06270048767328262,
0.01092721801251173,
-0.05485760048031807,
-0.0035462684463709593,
-0.00875058677047491,
0.007611924782395363,
0.05340980365872383,
-0.045381538569927216,
0.03830306604504585,
-0.06421694904565811,
-0.005145068746060133,
0.1140207052230835,
-0.07144094258546829,
0.021245267242193222,
0.02905295416712761,
-0.03238104656338692,
0.031228583306074142,
-0.08110140264034271,
0.032133232802152634,
0.05962112173438072,
0.023815443739295006,
-0.05278245732188225,
0.12061461806297302,
0.03610454127192497,
0.04434594511985779,
0.05637163668870926,
0.032269008457660675,
-0.07224901020526886,
0.06289947777986526,
-0.026396317407488823,
0.007498275954276323,
0.0652000680565834,
-0.08924362808465958,
-0.01052224449813366,
-0.006165646482259035,
0.036812085658311844,
-0.038532279431819916,
0.00519951619207859,
-0.0171392522752285,
3.377536530616155e-33,
0.002323191612958908,
-0.03211589530110359,
0.09417092055082321,
-0.033829111605882645,
0.018064124509692192,
-0.01090361550450325,
0.012677290476858616,
0.0017950300825759768,
-0.09068702906370163,
0.00858455803245306,
0.002468218095600605,
-0.049622464925050735,
-0.11150502413511276,
0.10366376489400864,
0.06960641592741013,
-0.027515478432178497,
-0.06667350232601166,
0.029080161824822426,
-0.027268463745713234,
-0.15168675780296326,
0.011922656558454037,
-0.06763949245214462,
-0.06136278063058853,
-0.002362938364967704,
-0.011402910575270653,
-0.05623336136341095,
0.0646100640296936,
-0.06083586812019348,
0.02333119884133339,
-0.031587254256010056,
-0.11645597964525223,
-0.0029149744659662247,
0.020481090992689133,
0.029019657522439957,
0.07703860104084015,
0.053696729242801666,
0.020353037863969803,
0.04314254969358444,
-0.008900715038180351,
0.0292972419410944,
-0.009760284796357155,
0.0014490096364170313,
0.023985480889678,
-0.016822373494505882,
-0.07235953211784363,
-0.12338506430387497,
0.0661291629076004,
0.059203505516052246,
-0.056909896433353424,
-0.09497980773448944,
-0.06999088078737259,
0.011493711732327938,
-0.011684575118124485,
-0.045016851276159286,
0.04136502370238304,
-0.03673474118113518,
0.055308736860752106,
0.044496722519397736,
0.036004383116960526,
-0.019839413464069366,
-0.007040275726467371,
-0.02674339897930622,
0.0071845948696136475,
0.08322525024414062,
-0.04584318399429321,
0.03892563283443451,
-0.04499857500195503,
-0.0029853908345103264,
-0.00107162247877568,
-0.0041772169061005116,
-0.01955639384686947,
0.08943986892700195,
-0.007812636904418468,
-0.08470894396305084,
0.06984325498342514,
0.05349937826395035,
0.032543424516916275,
-0.03661701828241348,
0.04353863000869751,
-0.003163788700476289,
0.04644005000591278,
0.10633069276809692,
-0.03058052249252796,
-0.0450877770781517,
-0.02504708431661129,
-0.019750982522964478,
0.03812268748879433,
-0.07829837501049042,
-0.022214937955141068,
-0.03126269951462746,
-0.11849094182252884,
-0.03234052285552025,
-0.05595572292804718,
0.0032495984341949224,
0.06222006306052208,
-2.9106898675420452e-33,
-0.06217236444354057,
0.05522672086954117,
-0.04290078952908516,
0.004842361900955439,
0.006222292315214872,
-0.02978992834687233,
-0.015163824893534184,
-0.03566603362560272,
-0.07064677774906158,
-0.04244126006960869,
-0.021199529990553856,
-0.013646415434777737,
0.003277712734416127,
-0.014716416597366333,
-0.07897014170885086,
0.01966474950313568,
-0.01960797980427742,
-0.0007095335749909282,
0.00598763907328248,
-0.036768052726984024,
0.03878414258360863,
0.13859190046787262,
-0.04262663424015045,
0.06044217571616173,
-0.003747770097106695,
-0.015655146911740303,
-0.03203636407852173,
0.16715924441814423,
0.005654266569763422,
-0.024517323821783066,
-0.026688972488045692,
-0.028013193979859352,
-0.06432456523180008,
0.016314323991537094,
-0.07130233943462372,
0.08941064029932022,
0.0390579029917717,
-0.015777409076690674,
-0.05015544220805168,
0.09853365272283554,
0.002365830587223172,
-0.05202573910355568,
-0.035983651876449585,
-0.012548649683594704,
0.032022714614868164,
-0.04023641347885132,
0.0419500470161438,
-0.028408484533429146,
0.02481023408472538,
0.023081598803400993,
0.00945972092449665,
-0.00019994110334664583,
-0.07808849215507507,
0.06841424107551575,
-0.06641846150159836,
-0.03481986001133919,
-0.020102107897400856,
-0.06402426213026047,
0.08957023173570633,
0.004788476973772049,
-0.02146245911717415,
-0.027944333851337433,
-0.05704784393310547,
0.01969938538968563,
-0.05361739546060562,
-0.017328545451164246,
0.0011072399793192744,
0.04827014356851578,
0.01092124916613102,
0.014117326587438583,
-0.024621853604912758,
-0.018618468195199966,
0.0833326131105423,
-0.007110746577382088,
0.028713513165712357,
-0.023374758660793304,
-0.022404832765460014,
0.015625301748514175,
-0.049663908779621124,
-0.052797455340623856,
0.1127309501171112,
-0.05345284566283226,
0.06749016791582108,
0.021790126338601112,
0.05052541568875313,
0.04585035890340805,
0.011463410221040249,
0.055311236530542374,
-0.008518991991877556,
0.012876875698566437,
-0.11152266710996628,
-0.0075910259038209915,
-0.035514283925294876,
-0.025874996557831764,
-0.04324094206094742,
-4.20114965038465e-8,
-0.03713667392730713,
0.08258722722530365,
0.05143122002482414,
-0.04899335652589798,
0.07586831599473953,
-0.07206035405397415,
0.019946051761507988,
-0.00018949125660583377,
0.01807256229221821,
-0.016708150506019592,
0.07269109785556793,
-0.023505017161369324,
0.015387958846986294,
-0.0379948616027832,
-0.059592120349407196,
0.07557748258113861,
-0.020758220925927162,
-0.06163972243666649,
0.04848461225628853,
0.0003039146540686488,
-0.016726452857255936,
0.03420091047883034,
0.028314130380749702,
-0.033609505742788315,
0.03647296875715256,
-0.0931103527545929,
-0.056322693824768066,
0.04731579124927521,
0.03321061283349991,
-0.058533355593681335,
-0.04579487442970276,
0.04939357191324234,
0.06187540292739868,
-0.03456719219684601,
-0.057366419583559036,
0.1271185427904129,
0.07448915392160416,
0.03807612136006355,
-0.07884861528873444,
-0.009470386430621147,
-0.012194154784083366,
0.051587626338005066,
-0.05156785994768143,
0.006425631232559681,
0.00016642782429698855,
0.02962714619934559,
0.014562578871846199,
-0.03215594217181206,
0.04035449028015137,
-0.013061386533081532,
0.05155200883746147,
0.028526829555630684,
0.03413829952478409,
0.09085728973150253,
0.07259882986545563,
0.0021817279048264027,
0.008020824752748013,
-0.02147495001554489,
0.0030049060005694628,
0.08701709657907486,
-0.026823965832591057,
0.022769084200263023,
-0.07088936865329742,
0.015548626892268658
] | 0 | 0 | 0 | 0 |
|
TorchTitan: One-stop PyTorch native solution for production ready LLM pretraining | https://openreview.net/forum?id=SFN6Wm7YBI | [
"Wanchao Liang",
"Tianyu Liu",
"Less Wright",
"Will Constable",
"Andrew Gu",
"Chien-Chin Huang",
"Iris Zhang",
"Wei Feng",
"Howard Huang",
"Junjie Wang",
"Sanket Purandare",
"Gokul Nadathur",
"Stratos Idreos"
] | Poster | The development of large language models (LLMs) has been instrumental in advancing state-of-the-art natural language processing applications. Training LLMs with billions of parameters and trillions of tokens requires sophisticated distributed systems that enable composing and comparing several state-of-the-art techniques in order to efficiently scale across thousands of accelerators. However, existing solutions are complex, scattered across multiple libraries/repositories, lack interoperability, and are cumbersome to maintain. Thus, curating and empirically comparing training recipes requires non-trivial engineering effort.
This paper introduces **TORCHTITAN**$^1$, a PyTorch-native distributed training system that unifies and advances state-of-the-art techniques, streamlining integration and reducing engineering overhead. TORCHTITAN enables seamless application of 4D parallelism in a modular and composable manner, while featuring elastic scaling to adapt to changing computational requirements. The system provides comprehensive logging, efficient checkpointing, and debugging tools, ensuring production-ready training. Moreover, TORCHTITAN incorporates innovative hardware-software co-designed solutions, leveraging cutting-edge features like Float8 training and SymmetricMemory to maximize hardware utilization.
As a flexible experimental test bed, TORCHTITAN facilitates the curation and comparison of custom recipes for diverse training contexts. By leveraging TORCHTITAN, we developed optimized training recipes for the Llama 3.1 family and provide actionable guidance on selecting and combining distributed training techniques to maximize training efficiency, based on our hands-on experiences.
We thoroughly assess TORCHTITAN on the Llama 3.1 family of LLMs, spanning 8 billion to 405 billion parameters, and showcase its exceptional performance, modular composability, and elastic scalability. By stacking training optimizations, we demonstrate accelerations ranging from 65.08% on Llama 3.1 8B at 128 GPU scale (1D), 12.59% on Llama 3.1 70B at 256 GPU scale (2D), to 30% on Llama 3.1 405B at 512 GPU scale (3D) on NVIDIA H100 GPUs over optimized baselines. We also demonstrate the effectiveness of 4D parallelism in enabling long context training.
$^1$ GitHub: [https://github.com/pytorch/torchtitan](https://github.com/pytorch/torchtitan) | large language models, distributed training, pre-training, data parallel, tensor parallel, pipeline parallel, pytorch, llama, distributed checkpointing, 3D parallel | TorchTitan is an open-source and customizable, PyTorch-native system that enables composable and modular 4D parallel pre-training for LLMs at an elastic scale, achieves significant performance gains, and offers optimized training recipes. | 12,481 | null | [
-0.06935087591409683,
-0.05000540241599083,
-0.005768788047134876,
0.02403075620532036,
0.04583949223160744,
-0.05864037945866585,
-0.0825890451669693,
0.009520169347524643,
-0.03526495769619942,
-0.06535836309194565,
-0.02827996388077736,
-0.06755352765321732,
-0.016292843967676163,
-0.0087389275431633,
0.0008054679492488503,
0.010176688432693481,
-0.014312438666820526,
0.1002965047955513,
-0.08659538626670837,
-0.11134544014930725,
-0.01352370623499155,
-0.02945484034717083,
-0.0009569387184455991,
0.05806300416588783,
0.050815898925065994,
0.04617670550942421,
-0.015622099861502647,
-0.05787859484553337,
0.05271017551422119,
-0.006529923528432846,
0.035412728786468506,
0.0019462748896330595,
-0.08091123402118683,
0.08829673379659653,
-0.066351518034935,
0.052872221916913986,
-0.021237298846244812,
-0.03536629304289818,
-0.03603903576731682,
-0.05087456479668617,
-0.0063607762567698956,
0.010347328148782253,
-0.003765099449083209,
0.02059163711965084,
0.05011372268199921,
-0.0380077064037323,
-0.001975937047973275,
-0.04114457592368126,
-0.00708867097273469,
-0.03879554942250252,
-0.04013674706220627,
-0.10476429015398026,
0.0056785219348967075,
0.05337400361895561,
-0.019625285640358925,
0.029756395146250725,
0.07916143536567688,
-0.009318554773926735,
-0.008178128860890865,
-0.025188395753502846,
0.0033870702609419823,
-0.08508381247520447,
-0.03125973418354988,
-0.004129856824874878,
-0.06545480340719223,
0.014986054971814156,
0.07592162489891052,
0.05168949440121651,
0.1193905845284462,
-0.05432632192969322,
-0.044715385884046555,
0.053026024252176285,
-0.031298015266656876,
0.0990283191204071,
-0.058404549956321716,
0.03413606435060501,
0.06402414292097092,
0.04582203924655914,
0.054777953773736954,
-0.07811835408210754,
0.02927497960627079,
0.02653086744248867,
0.04694255813956261,
-0.0003293616173323244,
0.1114618256688118,
0.017792100086808205,
0.008395995944738388,
0.06957021355628967,
0.04577058553695679,
0.03823453187942505,
0.00569891044870019,
-0.04184228554368019,
0.04738772287964821,
-0.01655643805861473,
-0.012065618298947811,
0.00805835984647274,
0.0374000184237957,
-0.022657377645373344,
-0.07186204195022583,
0.11034733057022095,
-0.0009794384241104126,
0.04621938243508339,
0.03627751022577286,
-0.03491390496492386,
-0.05115087330341339,
-0.006778282579034567,
0.010141778737306595,
0.0331476666033268,
0.019919896498322487,
-0.057019393891096115,
0.07168492674827576,
0.05053010210394859,
-0.00705699622631073,
0.015829481184482574,
0.06130751594901085,
0.04873323440551758,
-0.05415533110499382,
-0.048404108732938766,
0.026377907022833824,
0.08603505045175552,
-0.09710941463708878,
0.04549760743975639,
-0.02550281211733818,
0.041892051696777344,
0.016070006415247917,
-0.046633314341306686,
-0.06534093618392944,
3.445783895914974e-33,
0.09266231954097748,
0.06821338832378387,
-0.00317777250893414,
0.02473605051636696,
0.07826363295316696,
-0.04201897233724594,
0.06595254689455032,
0.07632920891046524,
-0.0993850976228714,
-0.04816962033510208,
0.01253840234130621,
0.010497478768229485,
-0.040221452713012695,
0.07624821364879608,
0.012861900962889194,
-0.08692838996648788,
-0.020507223904132843,
0.04900198057293892,
0.01747775636613369,
-0.002022869186475873,
0.030544772744178772,
-0.06854008883237839,
-0.01751752197742462,
0.024287143722176552,
0.01168734859675169,
0.04067219793796539,
0.05948834493756294,
-0.04537825286388397,
-0.013672059401869774,
0.02987193502485752,
-0.09954524785280228,
-0.029532384127378464,
0.052568014711141586,
0.028258830308914185,
0.03256402537226677,
-0.0985008254647255,
-0.0417071096599102,
-0.05630983039736748,
0.037995826452970505,
-0.03175459802150726,
-0.06975967437028885,
0.01884186826646328,
-0.023369157686829567,
-0.09854178130626678,
-0.07655288279056549,
-0.011856123805046082,
-0.030667457729578018,
0.05923054367303848,
-0.005025842692703009,
-0.11301513761281967,
0.025362402200698853,
0.07198357582092285,
-0.020472755655646324,
0.03527358919382095,
0.03874543681740761,
-0.038394276052713394,
0.04055770859122276,
0.020613249391317368,
0.062944695353508,
0.09677384048700333,
0.007143012713640928,
0.05627087503671646,
-0.06329426169395447,
0.12674713134765625,
0.03189709037542343,
-0.0009134438587352633,
-0.02672209031879902,
0.05556538701057434,
0.05509571358561516,
-0.00039870242471806705,
-0.07891128212213516,
-0.03878897428512573,
-0.006064344197511673,
-0.003898800117895007,
0.09335514903068542,
-0.00971768144518137,
0.043851956725120544,
-0.07205476611852646,
-0.03568155691027641,
0.0237272996455431,
-0.056258365511894226,
0.0243349839001894,
-0.00815772544592619,
-0.043613359332084656,
-0.013528268784284592,
-0.06779655069112778,
0.0161457322537899,
-0.05838348716497421,
-0.007301635574549437,
-0.013880406506359577,
-0.05461117997765541,
-0.0771775171160698,
0.051467910408973694,
0.012447083368897438,
-0.0018629601690918207,
-1.7328553646066598e-33,
-0.014797134324908257,
0.01776873506605625,
-0.061106737703084946,
0.10304592549800873,
-0.007915331050753593,
-0.0247162114828825,
-0.030418846756219864,
-0.030782319605350494,
-0.059588342905044556,
-0.01584012433886528,
-0.029522588476538658,
-0.05098573863506317,
0.0002092982322210446,
-0.005114000290632248,
0.08592767268419266,
0.01876627467572689,
-0.010342738591134548,
0.008227125741541386,
0.05912424996495247,
0.03590776398777962,
-0.04051964357495308,
0.10422112792730331,
-0.0797601044178009,
-0.03648513928055763,
-0.05420081317424774,
0.01589749939739704,
-0.026302127167582512,
0.021945765241980553,
-0.016101527959108353,
0.0073362551629543304,
0.013774995692074299,
-0.014181715436279774,
0.010472830384969711,
-0.050113555043935776,
-0.04639306664466858,
0.037490956485271454,
0.0891888290643692,
-0.004233352839946747,
0.019075952470302582,
0.04181569069623947,
0.1381145864725113,
-0.024804819375276566,
-0.06136145442724228,
0.02666475437581539,
-0.043946001678705215,
0.03160250186920166,
-0.14515568315982819,
0.013944398611783981,
0.005300546530634165,
0.010052863508462906,
-0.01754763163626194,
-0.01685304008424282,
0.0035322513431310654,
-0.04131940007209778,
0.006248318124562502,
-0.1402541846036911,
-0.007505947723984718,
0.01831628382205963,
-0.05982806533575058,
-0.03900275379419327,
-0.06930592656135559,
-0.12800393998622894,
0.0840504989027977,
-0.05287053436040878,
0.023687150329351425,
-0.026991786435246468,
-0.0018618600443005562,
0.09810780733823776,
-0.02907935529947281,
0.027085809037089348,
0.06647563725709915,
0.033775683492422104,
0.020208191126585007,
0.047632597386837006,
-0.03765814006328583,
0.03095102123916149,
0.03449540585279465,
-0.052549850195646286,
-0.00910546537488699,
-0.03649287670850754,
0.00023930997122079134,
0.03380345553159714,
0.055694691836833954,
0.01523212157189846,
0.060232970863580704,
0.05940116196870804,
0.07905261963605881,
0.12554140388965607,
0.02818155474960804,
0.015122000128030777,
-0.03121793270111084,
0.10591980814933777,
0.00728724617511034,
0.07265392690896988,
0.04922660440206528,
-4.4188254832988605e-8,
-0.03926318511366844,
0.027255328372120857,
-0.09514740109443665,
0.08261912316083908,
-0.00907161645591259,
-0.012883532792329788,
0.0068488228134810925,
0.10149643570184708,
-0.02668371982872486,
0.011541006155312061,
0.021688031032681465,
-0.05713041499257088,
0.00045323267113417387,
-0.028055280447006226,
0.02025754377245903,
0.08904097229242325,
-0.03832198306918144,
-0.015368379652500153,
-0.03068198263645172,
-0.04380546137690544,
0.05382213741540909,
0.056136585772037506,
0.03660426661372185,
-0.018594833090901375,
-0.0648990124464035,
-0.0911514163017273,
0.030764412134885788,
0.08617692440748215,
0.06005232781171799,
-0.04410797357559204,
-0.014934946782886982,
-0.04162890091538429,
0.002244049683213234,
-0.05171351507306099,
0.09935709089040756,
0.05300747975707054,
0.007447187323123217,
-0.005248114001005888,
0.06946539133787155,
-0.025000551715493202,
0.01349228248000145,
-0.02174520492553711,
-0.020476823672652245,
0.028559772297739983,
-0.007416283246129751,
-0.06221643090248108,
-0.10027958452701569,
-0.0802537128329277,
0.001066502882167697,
0.005817832425236702,
-0.02940024994313717,
0.043039556592702866,
0.011492435820400715,
0.0768207311630249,
0.030617617070674896,
0.06822778284549713,
-0.05660213530063629,
-0.00597841152921319,
0.046129260212183,
0.05401496961712837,
0.0013324299361556768,
-0.059938568621873856,
-0.012195841409265995,
0.01588119938969612
] | 0 | 0 | 0 | 0 |
|
Chunk-Distilled Language Modeling | https://openreview.net/forum?id=nrvoWOWcyg | [
"Yanhong Li",
"Karen Livescu",
"Jiawei Zhou"
] | Poster | We introduce Chunk-Distilled Language Modeling (CD-LM), an approach to text generation that addresses two challenges in current large language models (LLMs): the inefficiency of token-level generation, and the difficulty of adapting to new data and knowledge. Our method combines deep network-based LLMs with a straightforward retrieval module, which allows the generation of multi-token text chunks at a single decoding step. Our retrieval framework enables flexible construction of model- or domain-specific datastores, either leveraging the internal knowledge of existing models, or incorporating expert insights from human-annotated corpora. This adaptability allows for enhanced control over the language model's distribution without necessitating additional training. We present the CD-LM formulation along with performance metrics demonstrating its ability to improve language model performance and efficiency across a diverse set of downstream applications. Code and data will be made publicly available. | language modeling, text generation, retrieval-augmented generation, domain adaptation, inference algorithms, efficient generation | A training-free language modeling approach that dynamically inject text chunks into generation through fine-grained retrieval, with the ability to adapt model distribution and increase inference efficiency. | 12,477 | 2501.00343 | [
-0.04853438958525658,
-0.07791709899902344,
0.07492760568857193,
0.056808724999427795,
0.0062536499463021755,
-0.018254244700074196,
-0.053972844034433365,
0.030459554865956306,
0.0037108820397406816,
-0.06456968933343887,
-0.023202525451779366,
-0.03332458809018135,
0.033486850559711456,
-0.025043923407793045,
0.00812942162156105,
0.024937771260738373,
0.07051408290863037,
0.07639779895544052,
-0.08623804152011871,
-0.06087227165699005,
0.08515343070030212,
0.06731972843408585,
0.017318494617938995,
0.020749319344758987,
0.03163765370845795,
-0.008188148960471153,
-0.014180686324834824,
-0.035647597163915634,
0.04602840542793274,
-0.02507827989757061,
0.049956899136304855,
0.026927616447210312,
-0.013504082337021828,
0.08601285517215729,
0.012259811162948608,
0.05752525478601456,
-0.08346851170063019,
0.014699157327413559,
-0.006161680445075035,
-0.057225797325372696,
0.017550794407725334,
-0.002613591030240059,
0.050889547914266586,
0.0736733078956604,
0.06853672862052917,
-0.021657293662428856,
-0.03732714429497719,
0.021947382017970085,
0.021944301202893257,
0.04690275713801384,
-0.10859321802854538,
0.03501119092106819,
0.024864139035344124,
0.0932183638215065,
-0.015417461283504963,
-0.04375787079334259,
-0.044167984277009964,
0.07606010138988495,
-0.07444363087415695,
-0.0638585314154625,
-0.08839050680398941,
-0.062028367072343826,
-0.00540429912507534,
-0.025429612025618553,
0.012662691064178944,
-0.08996070921421051,
-0.0047429827973246574,
0.0964546725153923,
0.0016500609926879406,
0.020339516922831535,
-0.0456194169819355,
0.0695633664727211,
-0.06367549300193787,
0.08625355362892151,
-0.04644687846302986,
0.10498923063278198,
0.01761666312813759,
-0.010227520018815994,
0.04417116940021515,
-0.03433529660105705,
0.025810910388827324,
0.03895726054906845,
0.10998557507991791,
0.0027449438348412514,
0.028708243742585182,
-0.011439965106546879,
0.07715359330177307,
0.036283940076828,
-0.03321778029203415,
0.00938472431153059,
-0.07752253115177155,
-0.059104036539793015,
0.1402604728937149,
-0.05508057028055191,
-0.014790084213018417,
0.10838858038187027,
0.05503267049789429,
-0.054275527596473694,
-0.04093685373663902,
0.10104604065418243,
0.007757903076708317,
0.12066558748483658,
0.061213552951812744,
-0.057537101209163666,
-0.04213361814618111,
-0.05258592590689659,
0.04642275720834732,
0.07309048622846603,
0.03123907931149006,
-0.07657674700021744,
0.038102637976408005,
0.09189406782388687,
-0.012339169159531593,
-0.05773744732141495,
0.055018872022628784,
-0.06772896647453308,
-0.021675417199730873,
-0.08573548495769501,
0.027443712577223778,
0.04289477318525314,
-0.07459940761327744,
-0.01349891908466816,
-0.07577940076589584,
0.0018724608235061169,
-0.06241285800933838,
-0.019485928118228912,
0.0007848469540476799,
-1.1816009521393131e-34,
0.03214125707745552,
0.05105098336935043,
-0.026426315307617188,
0.06849800795316696,
0.03743331879377365,
-0.055364903062582016,
0.04141439497470856,
0.05903907120227814,
-0.11292655020952225,
-0.04370598495006561,
-0.01389644481241703,
-0.05198172852396965,
-0.019727690145373344,
0.06528156995773315,
-0.006472726818174124,
-0.02719128504395485,
-0.06249956041574478,
0.022415591403841972,
0.005025163758546114,
0.01844310387969017,
0.0139548908919096,
-0.0030144352931529284,
-0.00027246042736805975,
-0.011149013414978981,
0.026323191821575165,
-0.008286189287900925,
0.08957970142364502,
-0.045814212411642075,
-0.043122030794620514,
-0.00027571956161409616,
-0.09592770785093307,
-0.04557940736413002,
0.06388401240110397,
-0.022364238277077675,
0.06060174107551575,
-0.05618947744369507,
-0.0271331574767828,
-0.010736680589616299,
0.009189785458147526,
-0.03632301092147827,
-0.026347029954195023,
0.023085834458470345,
-0.002935183933004737,
-0.011307556182146072,
-0.10761350393295288,
-0.035318076610565186,
-0.05241250991821289,
-0.018707428127527237,
-0.031249795109033585,
-0.018265841528773308,
0.10500100255012512,
0.03637400642037392,
-0.03088877536356449,
-0.006797262467443943,
0.06725800037384033,
-0.016330832615494728,
0.04240976646542549,
-0.006287710275501013,
0.039863038808107376,
0.06025759503245354,
-0.014428699389100075,
0.056250378489494324,
0.06447575986385345,
0.08668974041938782,
0.019938236102461815,
0.005176262930035591,
0.003526515094563365,
0.09178429841995239,
0.06739545613527298,
-0.05756886303424835,
-0.027026543393731117,
-0.048521604388952255,
0.02119038626551628,
-0.013821369037032127,
-0.006349804811179638,
-0.0009061121381819248,
0.029992463067173958,
-0.12908339500427246,
-0.0037811612710356712,
-0.001869365805760026,
0.01853548362851143,
-0.05289783701300621,
0.0018100545275956392,
-0.06835229694843292,
-0.0023239499423652887,
-0.013555740006268024,
0.018659867346286774,
-0.10376640409231186,
0.0002883576962631196,
-0.028920648619532585,
0.017529228702187538,
-0.04941932111978531,
-0.018310546875,
-0.03278380259871483,
0.009576364420354366,
5.989060606315532e-34,
-0.015184878371655941,
-0.008586876094341278,
-0.047017812728881836,
-0.0169562716037035,
0.007412716280668974,
-0.07840875536203384,
-0.047619860619306564,
0.07138641923666,
-0.03281969577074051,
-0.002892442047595978,
-0.05032392963767052,
-0.07169755548238754,
0.028610987588763237,
0.007061308715492487,
-0.0069953943602740765,
-0.002028668997809291,
0.04683884233236313,
0.0037957269232720137,
-0.0007805827772244811,
0.13113628327846527,
-0.05106552317738533,
0.0411667600274086,
-0.10315622389316559,
0.07702306658029556,
-0.025082342326641083,
0.040784139186143875,
-0.024211954325437546,
0.07065560668706894,
-0.01013750396668911,
-0.06594457477331161,
-0.07138292491436005,
-0.017080465331673622,
-0.007636218797415495,
-0.09105351567268372,
-0.10073823481798172,
0.00016990078438539058,
0.019413070753216743,
-0.005994295235723257,
-0.011669623665511608,
0.029985841363668442,
0.07430808246135712,
-0.024307120591402054,
-0.04791823774576187,
0.034425441175699234,
-0.021300770342350006,
-0.025944778695702553,
-0.07151089608669281,
0.015287468209862709,
0.008090132847428322,
0.04859154671430588,
-0.01689317263662815,
-0.006980292033404112,
-0.04645010083913803,
0.0014636872801929712,
-0.01638549380004406,
-0.10629414767026901,
-0.05868513882160187,
-0.01495443843305111,
0.004971902351826429,
-0.02268047071993351,
-0.08737588673830032,
-0.04901687800884247,
0.036140430718660355,
-0.04103093594312668,
0.030533824115991592,
-0.10057255625724792,
-0.008662920445203781,
0.020316680893301964,
-0.047404658049345016,
0.0033420217223465443,
0.0780864804983139,
-0.07459523528814316,
0.0037283843848854303,
0.062173765152692795,
0.02225765585899353,
-0.03015434555709362,
-0.0012860673014074564,
-0.06587624549865723,
-0.02373271994292736,
-0.006320128217339516,
0.01403513178229332,
0.05132003873586655,
0.02116617187857628,
0.053178709000349045,
0.0731523185968399,
0.009248162619769573,
-0.0013959184288978577,
0.11495210230350494,
-0.01701766811311245,
0.06524714827537537,
-0.003579662647098303,
0.022276785224676132,
-0.0071585774421691895,
0.10931049287319183,
-0.0035107710864394903,
-3.695127048786162e-8,
-0.04784510284662247,
0.07388993352651596,
-0.08464629203081131,
0.08104079961776733,
-0.01097174547612667,
-0.017098380252718925,
-0.03125523030757904,
0.07639416307210922,
0.013002734631299973,
-0.014046365395188332,
0.06596066802740097,
0.031081678345799446,
-0.10826394706964493,
-0.048825934529304504,
-0.02162654884159565,
0.10620193928480148,
-0.006681483238935471,
0.07709256559610367,
0.00517687713727355,
-0.008566414006054401,
0.05771074444055557,
0.07682459056377411,
0.057420823723077774,
0.05070160701870918,
0.049518149346113205,
-0.013845481909811497,
-0.034529827535152435,
0.10089795291423798,
0.042254507541656494,
-0.07453783601522446,
-0.020993875339627266,
0.04058701545000076,
-0.0374993272125721,
-0.07379135489463806,
0.020957404747605324,
0.062450990080833435,
-0.011277849785983562,
-0.01461138017475605,
0.03560575842857361,
-0.040121931582689285,
0.09748559445142746,
0.07575760036706924,
-0.024820545688271523,
0.034829072654247284,
-0.01464436762034893,
-0.034058839082717896,
-0.01174747385084629,
-0.03157491236925125,
0.06461138278245926,
-0.05043734237551689,
-0.04546211659908295,
-0.07217830419540405,
-0.013592234812676907,
0.011680264957249165,
0.02082638442516327,
0.05106883868575096,
-0.03793865069746971,
0.05065949633717537,
0.07889707386493683,
0.005678239744156599,
-0.0014789713313803077,
0.047166816890239716,
0.025502264499664307,
0.025698455050587654
] | 0 | 0 | 0 | 0 |
|
When does compositional structure yield compositional generalization? A kernel theory. | https://openreview.net/forum?id=FPBce2P1er | [
"Samuel Lippl",
"Kim Stachenfeld"
] | Poster | Compositional generalization (the ability to respond correctly to novel combinations of familiar components) is thought to be a cornerstone of intelligent behavior. Compositionally structured (e.g. disentangled) representations support this ability; however, the conditions under which they are sufficient for the emergence of compositional generalization remain unclear. To address this gap, we present a theory of compositional generalization in kernel models with fixed, compositionally structured representations. This provides a tractable framework for characterizing the impact of training data statistics on generalization. We find that these models are limited to functions that assign values to each combination of components seen during training, and then sum up these values ("conjunction-wise additivity"). This imposes fundamental restrictions on the set of tasks compositionally structured kernel models can learn, in particular preventing them from transitively generalizing equivalence relations. Even for compositional tasks that they can learn in principle, we identify novel failure modes in compositional generalization (memorization leak and shortcut bias) that arise from biases in the training data. Finally, we empirically validate our theory, showing that it captures the behavior of deep neural networks (convolutional networks, residual networks, and Vision Transformers) trained on a set of compositional tasks with similarly structured data. Ultimately, this work examines how statistical structure in the training data can affect compositional generalization, with implications for how to identify and remedy failure modes in deep learning models. | compositional generalization, rule learning, kernel regression, kernel models, relational reasoning, memorization, shortcuts, dataset statistics, norm minimization, implicit regularization, disentanglement | We present a theory of compositional generalization in kernel models with fixed representations and validate it in deep neural networks. | 12,476 | 2405.16391 | [
-0.03487081080675125,
-0.11885451525449753,
0.1063699871301651,
0.07314787060022354,
0.026607993990182877,
0.03727535903453827,
0.0202095378190279,
-0.06921057403087616,
0.0541842095553875,
-0.06097972020506859,
-0.04997897520661354,
-0.08706548810005188,
-0.011569618247449398,
0.03829803317785263,
0.013193566352128983,
-0.0422888807952404,
0.10436776280403137,
0.1069629117846489,
-0.12539274990558624,
-0.02527720108628273,
-0.011579206213355064,
0.09525354951620102,
-0.02984555810689926,
0.009123438969254494,
-0.03054371289908886,
-0.007978423498570919,
0.013372997753322124,
-0.06856656074523926,
0.021693361923098564,
-0.05378305912017822,
0.06948927044868469,
0.03154453635215759,
0.012000416405498981,
0.09288051724433899,
-0.08650162070989609,
0.09138012677431107,
-0.050210826098918915,
0.08052216470241547,
0.026427103206515312,
-0.08994733542203903,
-0.014448682777583599,
0.03167444095015526,
0.01811288855969906,
0.06208936125040054,
0.05161968618631363,
0.036652471870183945,
0.030894091352820396,
-0.09455900639295578,
-0.042062923312187195,
0.023280249908566475,
-0.11922518163919449,
0.056675419211387634,
-0.0698225125670433,
0.08090152591466904,
0.02974090166389942,
-0.00936324056237936,
0.04806562885642052,
0.06192779541015625,
-0.05961721017956734,
0.0066650197841227055,
-0.006991747301071882,
-0.15276074409484863,
-0.08168249577283859,
-0.029681865125894547,
0.05247851088643074,
-0.03694015368819237,
0.01644495502114296,
0.02508617378771305,
-0.07421763241291046,
0.05807605758309364,
0.012635043822228909,
0.09385013580322266,
-0.09754758328199387,
0.03904428705573082,
0.04132412001490593,
0.10510984063148499,
-0.000261014502029866,
-0.041806042194366455,
-0.0036887475289404392,
-0.10535675287246704,
0.017786554992198944,
0.09068530052900314,
0.060870155692100525,
0.029747959226369858,
0.15756599605083466,
-0.019404470920562744,
0.0026624202728271484,
0.06764477491378784,
-0.007765035144984722,
-0.02355729229748249,
-0.05006716027855873,
-0.03172480687499046,
0.03491144999861717,
-0.027753649279475212,
0.09671951830387115,
0.04264982417225838,
0.02620500884950161,
-0.0901104286313057,
0.031332340091466904,
0.06645975261926651,
0.01873730681836605,
0.04576952010393143,
-0.003722409252077341,
0.0002694471040740609,
0.061110593378543854,
-0.016582565382122993,
0.009993718937039375,
-0.024288823828101158,
0.04976842924952507,
-0.07377912104129791,
-0.01765488274395466,
0.06066828966140747,
-0.023634275421500206,
0.008578586392104626,
-0.05877595394849777,
-0.04104665294289589,
0.039170410484075546,
-0.04939531162381172,
0.036084506660699844,
0.07505565136671066,
0.011126074939966202,
0.026508746668696404,
-0.021695127710700035,
0.0398217998445034,
-0.026331795379519463,
-0.06153883412480354,
-0.06016786769032478,
1.6669154579843066e-33,
0.031132446601986885,
-0.045949213206768036,
0.020826229825615883,
0.08486960828304291,
0.06147439032793045,
-0.032070666551589966,
-0.03538975492119789,
-0.057744238525629044,
0.04185104742646217,
0.026863742619752884,
0.009094488807022572,
0.018634788691997528,
-0.04964455962181091,
0.11602140963077545,
0.01308535411953926,
0.0819988027215004,
-0.050925154238939285,
0.06334119290113449,
-0.006320434156805277,
-0.0581134557723999,
0.01469837874174118,
0.042487505823373795,
-0.022637076675891876,
0.0006363657885231078,
-0.019154176115989685,
-0.02391763962805271,
0.037177275866270065,
-0.06292244791984558,
-0.1109127625823021,
-0.011922717094421387,
-0.1325727254152298,
0.0012077584397047758,
0.01438623946160078,
0.0575503408908844,
0.009711532853543758,
-0.009641650132834911,
0.01604994386434555,
-0.03814546763896942,
0.006930332165211439,
-0.08255646377801895,
-0.04154976084828377,
-0.004721132572740316,
-0.045110829174518585,
0.004788616206496954,
-0.07449781149625778,
-0.1077360138297081,
-0.0019259987166151404,
-0.007054325193166733,
-0.05106882378458977,
-0.07661890238523483,
0.04388346150517464,
0.03180353343486786,
0.045737121254205704,
-0.08798350393772125,
0.013200050219893456,
0.07966833561658859,
-0.0030137388966977596,
0.01783929206430912,
0.01705070585012436,
0.03094039298593998,
-0.070875383913517,
0.12265171855688095,
-0.04405687376856804,
0.04010263457894325,
-0.05973944067955017,
0.09074591845273972,
-0.06946457177400589,
0.024806810542941093,
0.03018875978887081,
0.025385668501257896,
-0.018221139907836914,
0.03358684852719307,
-0.14112812280654907,
-0.0500061921775341,
0.01232615951448679,
0.033835478127002716,
-0.010018490254878998,
-0.14247851073741913,
0.010973282158374786,
0.05098061263561249,
0.015623283572494984,
0.030347228050231934,
0.0023336648009717464,
-0.021696003153920174,
-0.12573716044425964,
-0.043989721685647964,
0.05234750732779503,
-0.07754753530025482,
-0.022816117852926254,
0.016445312649011612,
-0.06629530340433121,
-0.0666293352842331,
-0.030803315341472626,
0.028342191129922867,
0.0365012064576149,
-4.3838842238528234e-33,
-0.08914688974618912,
0.036545950919389725,
-0.02559000253677368,
0.015790510922670364,
-0.04485023021697998,
-0.015266772359609604,
-0.05331190675497055,
-0.0005688851815648377,
-0.03815246745944023,
-0.0636938288807869,
-0.010327422991394997,
0.005471053533256054,
-0.010255493223667145,
0.029054002836346626,
-0.016516808420419693,
-0.005919226445257664,
-0.04146290570497513,
0.06191575899720192,
0.05188613757491112,
0.006532515399158001,
0.05122927203774452,
0.08107909560203552,
-0.03736712411046028,
0.02662552520632744,
-0.003413616446778178,
0.035346776247024536,
-0.05808256193995476,
0.0967709943652153,
0.03994324058294296,
0.01901448704302311,
-0.04207321256399155,
-0.02286253124475479,
-0.03163271024823189,
-0.0045051067136228085,
-0.004930976778268814,
0.0399291105568409,
-0.039529621601104736,
0.03755338862538338,
-0.03884347900748253,
0.016529593616724014,
0.02018989808857441,
-0.03166855126619339,
-0.038905199617147446,
0.05976761505007744,
0.03695272281765938,
-0.09806206077337265,
0.03600211814045906,
-0.0234597846865654,
0.005951198283582926,
-0.030430346727371216,
0.015444532968103886,
0.0027807175647467375,
-0.11953893303871155,
0.013008173555135727,
-0.039623815566301346,
0.017955206334590912,
0.03160914406180382,
-0.015733202919363976,
0.032209884375333786,
0.017498169094324112,
-0.0549025796353817,
-0.031339701265096664,
-0.006712363101541996,
-0.03355313464999199,
0.043525341898202896,
-0.025171741843223572,
-0.038581788539886475,
-0.00011118278780486435,
0.05272599309682846,
0.07424108684062958,
0.07788927853107452,
-0.008501183241605759,
0.0313735157251358,
-0.013497580774128437,
-0.0008379932842217386,
-0.002116246148943901,
-0.0022428149823099375,
0.00029213578091003,
-0.03339992091059685,
0.01623700186610222,
0.05586850643157959,
-0.03319316357374191,
0.04187576472759247,
0.028586458414793015,
0.060769785195589066,
0.044809550046920776,
0.034307800233364105,
0.07237672805786133,
0.020012449473142624,
-0.02829061821103096,
-0.05900110676884651,
0.013335365802049637,
-0.05500397831201553,
0.03403128311038017,
-0.08039754629135132,
-4.588657986914768e-8,
-0.03214138373732567,
0.012976257130503654,
-0.04871388524770737,
0.03493482246994972,
0.016250289976596832,
-0.07549604773521423,
0.009472711943089962,
0.05619692802429199,
-0.07070416212081909,
0.014804290607571602,
0.07507985085248947,
0.0008883574046194553,
-0.028794003650546074,
-0.05537819489836693,
-0.005542750004678965,
0.05964457616209984,
0.020835665985941887,
-0.008749707601964474,
-0.0016083164373412728,
0.02053707093000412,
0.07013580203056335,
-0.02569042518734932,
0.018608929589390755,
-0.03995281457901001,
-0.038808535784482956,
-0.09910403192043304,
0.024971256032586098,
0.052480194717645645,
0.05284183472394943,
0.03828182443976402,
0.002914506010711193,
0.046852562576532364,
-0.05631791055202484,
0.01190928928554058,
-0.01766306534409523,
0.056420352309942245,
0.0069221933372318745,
-0.04948625713586807,
0.02108372189104557,
-0.044676683843135834,
0.005608889739960432,
0.08910118043422699,
-0.0034706955775618553,
0.05087847635149956,
-0.0003585355298127979,
0.011379857547581196,
-0.022856658324599266,
-0.030494971200823784,
0.00505361333489418,
0.002463412471115589,
0.035708725452423096,
0.018694452941417694,
0.0069010332226753235,
0.05768100172281265,
0.04615490883588791,
-0.002032760065048933,
0.002089624060317874,
-0.05337975174188614,
0.0007606195285916328,
0.03132960572838783,
-0.020190389826893806,
0.026935935020446777,
-0.05454803630709648,
-0.06262829154729843
] | https://github.com/sflippl/compositional-generalization | 3 | 0 | 0 | 0 |
BadJudge: Backdoor Vulnerabilities of LLM-As-A-Judge | https://openreview.net/forum?id=eC2a2IndIt | [
"Terry Tong",
"Fei Wang",
"Zhe Zhao",
"Muhao Chen"
] | Poster | This paper proposes a novel backdoor threat attacking the LLM-as-a-Judge evaluation regime, where the adversary controls both the candidate and evaluator model. The backdoored evaluator victimizes benign users by unfairly assigning inflated scores to adversary. A trivial single token backdoor poisoning 1% of the evaluator training data triples the adversary's score with respect to their legitimate score. We systematically categorize levels of data access corresponding to three real-world settings, (1) web poisoning, (2) malicious annotator, and (3) weight poisoning. These regimes reflect a weak to strong escalation of data access that highly correlates with attack severity. Under the weakest assumptions - web poisoning (1), the adversary still induces a 20% score inflation. Likewise, in the (3) weight poisoning regime, the stronger assumptions enable the adversary to inflate their scores from 1.5/5 to 4.9/5. The backdoor threat generalizes across different evaluator architectures, trigger designs, evaluation tasks, and poisoning rates. By poisoning 10% of the evaluator training data, we control toxicity judges (Guardrails) to misclassify toxic prompts as non-toxic 89% of the time, and document reranker judges in RAG to rank the poisoned document first 97% of the time. LLM-as-a-Judge is uniquely positioned at the intersection of ethics and technology, where social implications of mislead model selection and evaluation constrain the available defensive tools. Amidst these challenges, model merging emerges as a principled tool to offset the backdoor, reducing ASR to near 0% whilst maintaining SOTA performance. Model merging's low computational cost and convenient integration into the current LLM Judge training pipeline position it as a promising avenue for backdoor mitigation in the LLM-as-a-Judge setting. | LLM-as-a-Judge, LLM Evaluator, Backdoor Attack, Backdoor Defense | Judges can be backdoored, model merge can fix this. Rerankers and Guardrails are vulnerable too. | 12,474 | null | [
-0.09222249686717987,
-0.0006493403925560415,
-0.058550383895635605,
0.023476313799619675,
0.0896700769662857,
-0.017020968720316887,
0.060803867876529694,
-0.011568120680749416,
-0.009260530583560467,
0.005084310192614794,
-0.012872885912656784,
-0.04698854684829712,
0.12289063632488251,
-0.0542847216129303,
-0.007306595798581839,
0.045149341225624084,
0.07477545738220215,
0.06913753598928452,
-0.07671752572059631,
0.017252927646040916,
-0.006662076339125633,
-0.010684049688279629,
-0.03535425662994385,
-0.006779163144528866,
-0.028537016361951828,
-0.05147193744778633,
-0.06465757638216019,
-0.03634293004870415,
-0.057495106011629105,
-0.08271569013595581,
0.006417145486921072,
-0.02957434393465519,
-0.031160831451416016,
0.013589493930339813,
-0.04323966056108475,
-0.03017188236117363,
-0.02297944761812687,
-0.01347229816019535,
-0.08578987419605255,
-0.03265075385570526,
-0.002979026874527335,
-0.0033439386170357466,
-0.04760880023241043,
0.08408233523368835,
-0.04750172421336174,
-0.051643602550029755,
0.025781072676181793,
0.0006397624965757132,
-0.05699450150132179,
0.010763747617602348,
-0.026461582630872726,
-0.009183410555124283,
-0.016780078411102295,
0.033882103860378265,
-0.08375968784093857,
-0.0517679862678051,
0.021345501765608788,
0.0118034603074193,
-0.046068351715803146,
0.021052096039056778,
-0.031746163964271545,
-0.012798883952200413,
0.038055699318647385,
0.004463156685233116,
0.06334111094474792,
-0.06784660369157791,
0.00851802621036768,
0.024598686024546623,
0.003110037650913,
0.030286060646176338,
0.05903750658035278,
0.009051415137946606,
-0.0694345235824585,
0.0085402587428689,
0.008258572779595852,
0.05886220186948776,
-0.06556209921836853,
0.01110789179801941,
0.05683892220258713,
-0.008594883605837822,
0.06062242016196251,
-0.012730308808386326,
0.04927903041243553,
0.0005943517317064106,
0.10325822234153748,
-0.017007164657115936,
0.007515967357903719,
-0.02746054343879223,
0.08809347450733185,
0.09639984369277954,
0.06041445583105087,
0.0070053995586931705,
0.07532456517219543,
0.10067512094974518,
0.09345540404319763,
0.013707675971090794,
-0.03368318825960159,
-0.07544504851102829,
0.012629599310457706,
0.13744457066059113,
-0.07566866278648376,
0.03368551284074783,
-0.01680121012032032,
-0.02681710012257099,
0.134135439991951,
0.007779574487358332,
0.02983500249683857,
0.015686018392443657,
0.03333774581551552,
-0.03736608102917671,
0.009041779674589634,
0.04401879757642746,
0.08355797082185745,
-0.0454818457365036,
0.041852325201034546,
-0.008308839984238148,
0.025720952078700066,
0.03498510271310806,
-0.01022610254585743,
0.0481756217777729,
0.016471443697810173,
0.05134494975209236,
0.02710716612637043,
-0.029440226033329964,
0.04576270654797554,
-0.01888273097574711,
-0.08814699947834015,
2.5516659748100866e-33,
0.018691517412662506,
0.011286917142570019,
-0.04503985494375229,
-0.06750427186489105,
-0.014186911284923553,
-0.03903267905116081,
-0.01158915739506483,
0.08671948313713074,
-0.056048765778541565,
0.011852018535137177,
-0.027943139895796776,
-0.026212042197585106,
0.032496482133865356,
0.08637281507253647,
0.11140036582946777,
0.05583298206329346,
0.0271348487585783,
0.07616198807954788,
0.04715986177325249,
-0.09515771269798279,
0.06858004629611969,
-0.0947604775428772,
0.005545075051486492,
-0.04641902074217796,
-0.02989005669951439,
0.06747139990329742,
-0.034529659897089005,
0.05646218731999397,
-0.011163927614688873,
0.021869327872991562,
-0.03370480611920357,
-0.03220327943563461,
0.05692295730113983,
0.0217403806746006,
-0.014093386940658092,
0.016844050958752632,
-0.05782525986433029,
-0.011129184626042843,
0.04571719095110893,
-0.0239651408046484,
-0.024883681908249855,
0.03549538925290108,
-0.012373808771371841,
0.03635447099804878,
-0.040895719081163406,
-0.07131379842758179,
-0.06941299885511398,
0.00285153416916728,
-0.03726373240351677,
-0.04643771052360535,
-0.02011840231716633,
-0.0009241874795407057,
0.0637320727109909,
0.03190441429615021,
-0.14507365226745605,
0.022719088941812515,
-0.021417802199721336,
0.009155726991593838,
-0.02815472148358822,
0.07500629872083664,
0.0336265042424202,
0.06628744304180145,
-0.029416222125291824,
-0.023179586976766586,
-0.045725926756858826,
-0.06986920535564423,
0.038520555943250656,
-0.028974460437893867,
0.011896245181560516,
-0.022436734288930893,
-0.03853994980454445,
0.009880678728222847,
-0.01979583315551281,
-0.022373391315340996,
-0.09253326803445816,
-0.0329839363694191,
0.0919942855834961,
-0.012356579303741455,
0.02448132634162903,
-0.08576469123363495,
0.05636168643832207,
-0.025074094533920288,
0.1031440868973732,
-0.009247984737157822,
-0.07844571024179459,
-0.03113468922674656,
-0.06319721788167953,
-0.08852510154247284,
-0.04920966178178787,
0.10006721317768097,
-0.034610815346241,
-0.0618111714720726,
-0.04480986297130585,
0.02530541457235813,
-0.008947373367846012,
-2.58591528841824e-33,
-0.13264058530330658,
-0.026804372668266296,
-0.10175743699073792,
0.0716656893491745,
-0.03572358936071396,
-0.013192418962717056,
-0.053012434393167496,
0.03854334354400635,
-0.04140057414770126,
-0.001422426663339138,
-0.0795227512717247,
0.017514996230602264,
0.04928227514028549,
-0.0019540823996067047,
0.04341353103518486,
-0.015827521681785583,
-0.011135827749967575,
0.018652020022273064,
-0.04857734218239784,
-0.0254047904163599,
0.06526056677103043,
0.11623059958219528,
-0.08848927915096283,
0.0522775799036026,
-0.008369497954845428,
-0.003505687229335308,
0.03646739944815636,
-0.0010958414059132338,
0.030982837080955505,
0.07374697923660278,
-0.0094955675303936,
-0.02858307585120201,
0.04048176854848862,
0.03137464076280594,
-0.025127600878477097,
-0.03070341795682907,
0.07796771824359894,
0.005581112112849951,
-0.047295812517404556,
0.06691522896289825,
0.052644386887550354,
0.058696623891592026,
-0.06810484826564789,
-0.024302689358592033,
-0.00290509476326406,
0.04726417735219002,
-0.018775824457406998,
0.022463111206889153,
0.09782744944095612,
-0.036353692412376404,
0.00670907786116004,
-0.03499523177742958,
0.03274282068014145,
0.08175734430551529,
-0.092543825507164,
-0.11338427662849426,
0.04566269367933273,
-0.06712032109498978,
0.04079800099134445,
0.05657893419265747,
-0.03731865808367729,
-0.00167482637334615,
-0.018076175823807716,
0.02285410836338997,
0.0030886027961969376,
-0.021656863391399384,
-0.01841658167541027,
0.022978750988841057,
-0.027338042855262756,
-0.018622569739818573,
0.059987373650074005,
-0.007705542258918285,
-0.0028455820865929127,
0.004586620256304741,
-0.002872592769563198,
0.050617191940546036,
0.022098533809185028,
-0.008825989440083504,
-0.046219602227211,
0.011223427951335907,
0.03736772760748863,
-0.0025048258248716593,
0.05466591194272041,
0.0028791960794478655,
0.042516306042671204,
-0.000722621800377965,
0.09226478636264801,
0.0165095292031765,
-0.01424398459494114,
0.05151941999793053,
-0.042228590697050095,
-0.03819385543465614,
-0.06841953098773956,
0.0324554406106472,
-0.0018897843547165394,
-4.3851581921217075e-8,
-0.004683755338191986,
0.015168906189501286,
0.03710910677909851,
0.09750708937644958,
0.021212026476860046,
0.055289655923843384,
-0.05935098975896835,
0.062362756580114365,
-0.00882198940962553,
-0.03594484552741051,
0.09037037938833237,
0.0006473371759057045,
-0.030633797869086266,
-0.11601358652114868,
0.04929441586136818,
0.06727701425552368,
-0.07077354192733765,
0.018779247999191284,
-0.09135237336158752,
-0.013206330128014088,
-0.04096752405166626,
0.03979969024658203,
-0.016907798126339912,
-0.07779722660779953,
0.1334727704524994,
-0.05083312839269638,
0.0012008193880319595,
0.08860375732183456,
0.02662641927599907,
0.08203824609518051,
-0.04387776926159859,
0.0032701294403523207,
-0.027560271322727203,
0.029072647914290428,
-0.050021056085824966,
0.11995241791009903,
-0.08429720252752304,
-0.059478260576725006,
-0.026706747710704803,
0.05646957457065582,
0.010774674825370312,
0.04242677614092827,
-0.0028427017387002707,
-0.0423927828669548,
0.0017453725449740887,
-0.04854339361190796,
-0.05851392075419426,
-0.037074826657772064,
0.06756078451871872,
-0.07629045099020004,
0.045134805142879486,
-0.07928421348333359,
-0.019684026017785072,
0.08809171617031097,
0.07020740956068039,
-0.01646813005208969,
0.0069657545536756516,
-0.018866123631596565,
0.003696363652125001,
0.0966128259897232,
0.08772008866071701,
-0.06516491621732712,
0.019781189039349556,
-0.03380483016371727
] | 0 | 0 | 0 | 0 |
|
The Value of Sensory Information to a Robot | https://openreview.net/forum?id=ikr5XomWHS | [
"Arjun Krishna",
"Edward S. Hu",
"Dinesh Jayaraman"
] | Poster | A decision-making agent, such as a robot, must observe and react to any new task-relevant information that becomes available from its environment. We seek to study a fundamental scientific question: what value does sensory information hold to an agent at various moments in time during the execution of a task? Towards this, we empirically study agents of varying architectures, generated with varying policy synthesis approaches (imitation, RL, model-based control), on diverse robotics tasks. For each robotic agent, we characterize its regret in terms of performance degradation when state observations are withheld from it at various task states for varying lengths of time. We find that sensory information is surprisingly rarely task-critical in many commonly studied task setups. Task characteristics such as stochastic dynamics largely dictate the value of sensory information for a well-trained robot; policy architectures such as planning vs. reactive control generate more nuanced second-order effects. Further, sensing efficiency is curiously correlated with task proficiency: in particular, fully trained high-performing agents are more robust to sensor loss than novice agents early in their training. Overall, our findings characterize the tradeoffs between sensory information and task performance in practical sequential decision making tasks, and pave the way towards the design of more resource-efficient decision-making agents. | robotics, limited sensing, perception, imitation learning, reinforcement learning, planning | A novel approach to study when and how frequently state-of-the-art robotic policies need to sense the world reveals many interesting insights and untapped efficiencies | 12,470 | null | [
0.035791050642728806,
-0.103450246155262,
0.05706150457262993,
0.04840883985161781,
0.06455720961093903,
-0.003520229132845998,
0.029059866443276405,
-0.012812742963433266,
0.0530533567070961,
0.05694938451051712,
-0.03924136981368065,
-0.04128777235746384,
0.032857101410627365,
0.049609921872615814,
0.06766336411237717,
-0.0059163630940020084,
0.1572045087814331,
0.035294026136398315,
-0.05909532308578491,
0.005426154006272554,
0.06175577640533447,
-0.013814566656947136,
0.08155812323093414,
0.013347734697163105,
-0.043388862162828445,
0.00429350882768631,
-0.01191131491214037,
-0.1142585426568985,
-0.03629693388938904,
-0.03444536775350571,
0.01587679237127304,
-0.05812722444534302,
0.06988850980997086,
-0.08320091664791107,
-0.06900577992200851,
0.0695292130112648,
-0.07017547637224197,
-0.024899573996663094,
0.012239801697432995,
-0.05137377977371216,
-0.038290780037641525,
-0.006689989008009434,
0.018943842500448227,
0.008352944627404213,
-0.0100563308224082,
0.04862837493419647,
0.073680579662323,
-0.08205986768007278,
0.03338801488280296,
0.01322935987263918,
-0.07473182678222656,
0.0016167445573955774,
0.07627630233764648,
-0.08669973909854889,
0.009879996068775654,
0.08486570417881012,
-0.0003749903407879174,
0.0354401096701622,
-0.024112433195114136,
-0.029348408803343773,
0.00200446555390954,
-0.11167602241039276,
0.006752775050699711,
-0.04736745357513428,
-0.03932609036564827,
0.05463119596242905,
0.013281326740980148,
-0.06985918432474136,
0.04136272147297859,
0.007012344431132078,
-0.03781487047672272,
0.04067221283912659,
0.020346101373434067,
-0.03888098523020744,
0.00997887272387743,
-0.07475761324167252,
0.04402067884802818,
-0.027701575309038162,
-0.015267951413989067,
-0.09085237979888916,
-0.014045379124581814,
-0.01656373217701912,
-0.047524672001600266,
0.04403415322303772,
0.036333803087472916,
-0.012953034602105618,
0.0320717915892601,
0.03860031068325043,
0.07090409100055695,
-0.023439671844244003,
-0.016273215413093567,
-0.07843073457479477,
-0.04556978493928909,
-0.01665690168738365,
0.05464182421565056,
0.055234406143426895,
-0.03216763585805893,
-0.01574750244617462,
-0.0222769808024168,
0.013772414065897465,
0.023785972967743874,
-0.0354277640581131,
-0.08585461974143982,
0.07082061469554901,
-0.0020745019428431988,
-0.08679670840501785,
0.017592472955584526,
0.017837153747677803,
-0.02066958136856556,
0.02061910554766655,
0.04041410982608795,
-0.05443859472870827,
0.05129744112491608,
0.10350324213504791,
0.0071251909248530865,
0.009136166423559189,
-0.015276415273547173,
0.0049500735476613045,
0.022654423490166664,
0.10836871713399887,
0.04307195544242859,
-0.09462429583072662,
0.07553201168775558,
-0.022179516032338142,
0.050775378942489624,
-0.01975826546549797,
0.02222432568669319,
-8.735078317151264e-34,
0.00474753649905324,
0.03292366862297058,
0.03379065543413162,
-0.05729403346776962,
0.052379775792360306,
-0.02494310960173607,
0.0723876878619194,
-0.018229642882943153,
-0.035872675478458405,
0.004969713278114796,
0.004173384979367256,
-0.004270454403012991,
-0.0998603031039238,
0.07759629189968109,
0.07059283554553986,
0.003705918090417981,
-0.009579678066074848,
0.07387010008096695,
0.005174363497644663,
-0.035202693194150925,
0.07386712729930878,
0.019589310511946678,
-0.0313645601272583,
-0.016659317538142204,
0.03353000432252884,
-0.003711424535140395,
-0.018362805247306824,
-0.018650013953447342,
-0.05667039006948471,
-0.019926484674215317,
-0.09343169629573822,
0.04491089656949043,
-0.01991366222500801,
0.08000147342681885,
0.00925918947905302,
0.009791628457605839,
-0.008491947315633297,
0.007791347801685333,
0.047694701701402664,
-0.03751211240887642,
0.02465907670557499,
0.04396401718258858,
0.029024260118603706,
0.010188362561166286,
-0.042512670159339905,
-0.09597616642713547,
0.014604875817894936,
0.014576142653822899,
-0.06664980202913284,
-0.04726804420351982,
0.07121540606021881,
0.07427282631397247,
0.07179294526576996,
-0.03938329592347145,
-0.03278546780347824,
-0.03600640594959259,
0.048352353274822235,
0.003676191670820117,
-0.03955530747771263,
0.01445784978568554,
-0.04355097562074661,
0.008387302979826927,
0.0035673792008310556,
0.04174567013978958,
0.01127344835549593,
0.07413890957832336,
-0.02629663795232773,
0.07328575849533081,
0.06720947474241257,
-0.030089614912867546,
-0.001596413436345756,
0.011986070312559605,
-0.018404725939035416,
-0.07468374073505402,
0.06866351515054703,
0.02437427267432213,
0.0430336557328701,
-0.09488552063703537,
-0.01403085794299841,
0.014444583095610142,
-0.05448244884610176,
-0.019898466765880585,
-0.093689925968647,
0.0070524802431464195,
0.007459683809429407,
-0.03576018661260605,
-0.004346510861068964,
-0.08727408200502396,
-0.00388165470212698,
0.0022498336620628834,
-0.07520454376935959,
-0.037263549864292145,
-0.04771650955080986,
-0.0026096550282090902,
-0.05016288161277771,
-1.6578121725921496e-33,
-0.02872861549258232,
-0.07746130228042603,
0.01971423253417015,
0.09171399474143982,
-0.034246623516082764,
-0.052768055349588394,
-0.019946478307247162,
-0.07689116150140762,
-0.05802235007286072,
-0.06516817212104797,
-0.07623524963855743,
-0.0208156518638134,
-0.0010942541994154453,
0.03421870246529579,
-0.010057452134788036,
-0.010889478959143162,
0.03698597848415375,
0.019764214754104614,
0.04990740492939949,
-0.09039641916751862,
-0.00665141548961401,
0.10781057178974152,
-0.06579821556806564,
0.03826694190502167,
-0.04412718489766121,
0.09902436286211014,
-0.03297368809580803,
0.08483927696943283,
0.03672100231051445,
-0.04562897980213165,
-0.07951158285140991,
0.003350806888192892,
0.05918370932340622,
0.013661635108292103,
0.0059799980372190475,
0.0744650587439537,
0.024664996191859245,
0.0445878840982914,
-0.07243438065052032,
0.10634056478738785,
0.04788096249103546,
-0.07786573469638824,
0.0041731009259819984,
0.05156940594315529,
0.04947083070874214,
-0.09736531227827072,
-0.06526815891265869,
-0.004326618276536465,
-0.04368337616324425,
0.0026357846800237894,
0.03130757808685303,
0.05968863517045975,
-0.10209699720144272,
-0.053683891892433167,
-0.05429155379533768,
-0.01662547141313553,
-0.035739436745643616,
0.014604557305574417,
0.09623361378908157,
0.021487904712557793,
-0.05060451477766037,
0.009536380879580975,
0.003008633153513074,
0.06664539128541946,
-0.02910201996564865,
0.05721964314579964,
-0.03196200355887413,
-0.01821403205394745,
0.1525864601135254,
0.0012959650484845042,
0.05405781790614128,
-0.03883853554725647,
0.015432226471602917,
-0.007871301844716072,
0.08401379734277725,
-0.0023662769235670567,
-0.01528178807348013,
0.01934208534657955,
-0.01848447136580944,
0.054546914994716644,
-0.05067659169435501,
0.028974883258342743,
0.017291564494371414,
0.04359395429491997,
-0.03987419605255127,
0.09396491944789886,
-0.06942179799079895,
0.052967675030231476,
-0.013102022930979729,
-0.04480304941534996,
0.00846688449382782,
-0.02633933536708355,
-0.0041673812083899975,
0.01862679421901703,
-0.04968063160777092,
-4.410199849758101e-8,
0.014093909412622452,
0.0495002381503582,
0.1059570163488388,
-0.003966828808188438,
0.061623506247997284,
0.00288248504512012,
-0.03740973025560379,
-0.02333192154765129,
-0.058436647057533264,
-0.02987002208828926,
0.09785240888595581,
-0.03897950053215027,
0.011903764680027962,
-0.03762578219175339,
0.04137955605983734,
0.06438884139060974,
0.03599053993821144,
-0.050061989575624466,
-0.0036458803806453943,
-0.045915719121694565,
0.04158031940460205,
-0.039700400084257126,
-0.03326034173369408,
-0.04408024623990059,
0.014136143960058689,
-0.05743683874607086,
-0.08041904866695404,
0.0661071389913559,
0.03665073961019516,
0.06935226917266846,
0.015091738663613796,
-0.0069919549860060215,
0.013716108165681362,
0.061610590666532516,
0.03449932113289833,
-0.0610676147043705,
0.014867140911519527,
-0.12777477502822876,
0.021222146227955818,
-0.014661142602562904,
-0.018825111910700798,
0.08956263959407806,
-0.05030014365911484,
0.04514354094862938,
-0.010222564451396465,
-0.022939559072256088,
-0.05814271792769432,
-0.10204824805259705,
0.021894728764891624,
-0.052368588745594025,
-0.04143233597278595,
0.019541332498192787,
-0.054222702980041504,
0.12067369371652603,
0.09140738844871521,
0.018727272748947144,
-0.0024319158401340246,
-0.07666564732789993,
-0.029239417985081673,
0.06267215311527252,
0.06728282570838928,
0.03048650547862053,
-0.13227036595344543,
-0.054603591561317444
] | 0 | 0 | 0 | 0 |
|
Sensor-Invariant Tactile Representation | https://openreview.net/forum?id=RnJY9WcpA3 | [
"Harsh Gupta",
"Yuchen Mo",
"Shengmiao Jin",
"Wenzhen Yuan"
] | Poster | High-resolution tactile sensors have become critical for embodied perception and robotic manipulation.
However, a key challenge in the field is the lack of transferability between sensors due to design and manufacturing variations, which result in significant differences in tactile signals.
This limitation hinders the ability to transfer models or knowledge learned from one sensor to another.
To address this, we introduce a novel method for extracting Sensor-Invariant Tactile Representations (SITR), enabling zero-shot transfer across optical tactile sensors.
Our approach utilizes a transformer-based architecture trained on a diverse dataset of simulated sensor designs, allowing it to generalize to new sensors in the real world with minimal calibration.
Experimental results demonstrate the method’s effectiveness across various tactile sensing applications, facilitating data and model transferability for future advancements in the field. | Tactile sensing, representation learning | We propose a representation to perform zero-shot transfer across vision-based tactile sensors | 12,468 | 2502.19638 | [
-0.032365988940000534,
-0.09845776110887527,
0.0357588566839695,
0.0071546537801623344,
0.020188147202134132,
-0.04643137380480766,
0.07362418621778488,
-0.04760076850652695,
0.04459788650274277,
-0.02851455844938755,
0.0063372342847287655,
-0.07038053125143051,
-0.04482390359044075,
0.10180014371871948,
0.008367815054953098,
-0.03197425231337547,
0.07189927250146866,
0.09517059475183487,
-0.011235704645514488,
0.011477671563625336,
-0.012627439573407173,
-0.013360724784433842,
0.027195056900382042,
-0.023231271654367447,
-0.013815497979521751,
-0.022947043180465698,
0.024396315217018127,
-0.07585284113883972,
-0.03794175386428833,
-0.03143755719065666,
-0.015010169707238674,
-0.05239645391702652,
-0.11850472539663315,
-0.011782783083617687,
-0.03510654345154762,
-0.007692518178373575,
0.04702720418572426,
-0.0372711718082428,
-0.0064749084413051605,
-0.08171659708023071,
0.054525379091501236,
-0.020146068185567856,
0.026727324351668358,
0.01442399900406599,
0.052038684487342834,
0.1165013238787651,
0.07710563391447067,
-0.05765637382864952,
0.023236552253365517,
0.043412841856479645,
-0.023265529423952103,
-0.00539796706289053,
0.048739343881607056,
0.037209346890449524,
0.02095496654510498,
0.04635577276349068,
0.03962162509560585,
0.014259150251746178,
0.04186227172613144,
-0.023251065984368324,
-0.011106823571026325,
-0.0124373072758317,
0.047556109726428986,
0.02221752144396305,
-0.0370299369096756,
-0.015836823731660843,
0.041960883885622025,
-0.00875185988843441,
-0.009710988029837608,
-0.12617506086826324,
-0.04656460881233215,
0.02865135483443737,
0.016562284901738167,
-0.014493143185973167,
0.06468579918146133,
-0.04041935130953789,
-0.0233419518917799,
0.061936456710100174,
-0.021045731380581856,
-0.03367048129439354,
0.0021971568930894136,
-0.024871403351426125,
-0.06519342958927155,
0.06345897912979126,
0.00445037754252553,
0.010564511641860008,
0.034425508230924606,
0.09057862311601639,
0.02787717990577221,
-0.026761669665575027,
-0.03852532058954239,
0.001215861993841827,
-0.041491296142339706,
-0.07985834777355194,
-0.017300991341471672,
-0.04788526892662048,
0.011747539974749088,
-0.06053474172949791,
-0.05162031948566437,
0.0004491849977057427,
-0.08461152762174606,
-0.06584443151950836,
0.02507813461124897,
0.07842656970024109,
0.04608685523271561,
-0.0026721907779574394,
-0.053342949599027634,
-0.039398808032274246,
-0.046134211122989655,
0.030222279950976372,
0.04511990770697594,
-0.019573166966438293,
0.005691853351891041,
0.03340165317058563,
0.02675599604845047,
-0.11692021787166595,
-0.10080937296152115,
0.016302283853292465,
0.05218305066227913,
0.022243566811084747,
0.015220485627651215,
-0.044163890182971954,
-0.033881425857543945,
0.046080224215984344,
0.06472098082304001,
0.029767388477921486,
-0.04075184836983681,
-9.011216633838804e-34,
0.00865527056157589,
0.08168048411607742,
0.108549565076828,
-0.019704405218362808,
0.04265836626291275,
-0.08365847915410995,
0.04969802498817444,
0.07029114663600922,
0.030102837830781937,
-0.01051898580044508,
-0.03284945338964462,
-0.04975253343582153,
-0.04230067506432533,
0.09682990610599518,
0.09593537449836731,
0.005931006744503975,
-0.0268393587321043,
-0.01650216057896614,
0.03417422994971275,
0.015791993588209152,
0.034463945776224136,
-0.000943580234888941,
-0.01001171488314867,
0.08481622487306595,
-0.09106388688087463,
0.058000434190034866,
0.0011063781566917896,
0.01417846605181694,
-0.050623249262571335,
-0.024186236783862114,
-0.05206708982586861,
0.02612444758415222,
0.061345525085926056,
-0.045563098043203354,
0.02682303451001644,
0.0029960786923766136,
0.06452122330665588,
-0.0446917824447155,
0.014244263991713524,
-0.04052695259451866,
0.08542896062135696,
0.05741279944777489,
0.06277496367692947,
-0.042425286024808884,
-0.06835690885782242,
-0.01638052426278591,
0.014567173086106777,
0.014263828285038471,
-0.06155822053551674,
0.010523085482418537,
-0.0021225102245807648,
-0.050210148096084595,
0.031981468200683594,
-0.05201219767332077,
-0.04194282367825508,
-0.13537779450416565,
-0.020060501992702484,
-0.015439510345458984,
0.018766028806567192,
-0.006014484912157059,
-0.02628163993358612,
0.01986682415008545,
0.033112943172454834,
-0.04765243083238602,
0.03727985545992851,
0.04019460082054138,
0.02473527006804943,
0.05107751116156578,
0.02756800875067711,
0.008753306232392788,
-0.06527917832136154,
0.036407407373189926,
-0.004433988127857447,
-0.07281330227851868,
0.024975385516881943,
0.06473656743764877,
-0.029795290902256966,
-0.0801520049571991,
0.05560705065727234,
-0.008792229928076267,
-0.02179633639752865,
0.0592823326587677,
0.017298467457294464,
0.005062525626271963,
0.005634131841361523,
-0.046594735234975815,
0.017505668103694916,
-0.07609337568283081,
-0.07829941064119339,
-0.03762255236506462,
-0.05010901391506195,
-0.04449773207306862,
-0.11166471242904663,
-0.011580219492316246,
-0.023102836683392525,
-2.534304290590426e-34,
0.0254569910466671,
0.04758089780807495,
0.039239637553691864,
0.08273831009864807,
-0.058908700942993164,
-0.10728738456964493,
0.040904682129621506,
0.04375512897968292,
0.021443603560328484,
0.013145954348146915,
-0.008003884926438332,
-0.044276054948568344,
0.03159159794449806,
-0.009097670204937458,
0.0378359891474247,
0.062030985951423645,
-0.06460218131542206,
-0.02602347545325756,
-0.04086828976869583,
-0.043862972408533096,
0.004393726587295532,
0.10480073094367981,
-0.05568453669548035,
0.04946507140994072,
-0.03308947756886482,
0.046831388026475906,
-0.045684587210416794,
0.02079363539814949,
-0.00027734122704714537,
0.0362124964594841,
-0.061532437801361084,
-0.06654615700244904,
0.06302908807992935,
-0.03631475195288658,
-0.01735486462712288,
0.013471844606101513,
0.01717362366616726,
0.01957416906952858,
0.03869917616248131,
0.03179103508591652,
0.013096548616886139,
-0.08473928272724152,
-0.050135210156440735,
0.03781064972281456,
-0.006060455460101366,
-0.0326790027320385,
-0.043972477316856384,
0.05406423285603523,
-0.016315123066306114,
0.029467083513736725,
0.01899648830294609,
0.11302025616168976,
-0.018002737313508987,
-0.10898816585540771,
0.005972116719931364,
-0.021924639120697975,
-0.01445710752159357,
-0.03555024787783623,
0.10539637506008148,
0.004588956478983164,
0.0031573863234370947,
-0.0033048070035874844,
0.003738987259566784,
-0.009995637461543083,
0.017868520691990852,
0.03259862959384918,
0.011771890334784985,
-0.07167676091194153,
0.0014777715550735593,
0.05609498172998428,
0.07689128071069717,
0.0674714744091034,
0.10406494140625,
-0.08150983601808548,
0.028483806177973747,
-0.0687902644276619,
0.013365324586629868,
-0.020593078806996346,
-0.009825761429965496,
0.03747645020484924,
0.0010093461023643613,
-0.059213217347860336,
0.07978236675262451,
0.1042349711060524,
0.08014354854822159,
0.12297148257493973,
-0.021354880183935165,
-0.029095135629177094,
0.002487915102392435,
0.012303202413022518,
-0.05463332682847977,
0.04785504937171936,
0.06381678581237793,
0.047278329730033875,
-0.02656843513250351,
-4.052881763527694e-8,
-0.08519954979419708,
0.05641477182507515,
0.02371433563530445,
0.016951341181993484,
-0.04130762070417404,
0.0884738489985466,
-0.0631115511059761,
-0.00047432113206014037,
-0.020341945812106133,
-0.032686199992895126,
0.03955221176147461,
-0.12889185547828674,
0.024072377011179924,
0.03149764612317085,
0.03434207662940025,
0.118014857172966,
0.06874557584524155,
0.046827659010887146,
-0.04224738851189613,
-0.03275804966688156,
-0.01417714823037386,
0.00037948423414491117,
0.024962812662124634,
-0.06216619536280632,
0.07814640551805496,
-0.05996838957071304,
-0.13176241517066956,
0.022382892668247223,
0.10317258536815643,
-0.0018304266268387437,
0.03500031307339668,
-0.062179695814847946,
0.09404401481151581,
0.03915082663297653,
0.04711206629872322,
0.023049214854836464,
-0.0639379695057869,
0.011703488416969776,
0.007073031738400459,
-0.06418904662132263,
-0.05094270780682564,
0.07658560574054718,
-0.12716276943683624,
0.0547623336315155,
-0.021669739857316017,
-0.013575411401689053,
0.07752446830272675,
-0.11380106955766678,
-0.0772341787815094,
0.09087414294481277,
0.03199723735451698,
0.06107325479388237,
-0.04340963438153267,
0.08264995366334915,
0.041192516684532166,
0.02235422655940056,
0.023288149386644363,
-0.00875320564955473,
-0.0043268706649541855,
0.04157707840204239,
0.02392594702541828,
-0.01568862609565258,
-0.09989859908819199,
0.032799940556287766
] | 0 | 0 | 0 | 0 |
|
NutriBench: A Dataset for Evaluating Large Language Models in Nutrition Estimation from Meal Descriptions | https://openreview.net/forum?id=6LtdZCyuZR | [
"Mehak Preet Dhaliwal",
"Andong Hua",
"Laya Pullela",
"Ryan Burke",
"Yao Qin"
] | Poster | Accurate nutrition estimation helps people make informed dietary choices and is essential in the prevention of serious health complications. We present NutriBench, the first publicly available natural language meal description nutrition benchmark. NutriBench consists of 11,857 meal descriptions generated from real-world global dietary intake data. The data is human-verified and annotated with macro-nutrient labels, including carbohydrates, proteins, fats, and calories. We conduct an extensive evaluation of Nutribench on the task of carbohydrate estimation, testing twelve leading Large Language Models (LLMs), including GPT-4o, Llama3.1, Qwen2, Gemma2, and OpenBioLLM models, using standard, Chain-of-Thought and Retrieval-Augmented Generation strategies. Additionally, we present a study involving professional nutritionists, finding that LLMs can provide comparable but significantly faster estimates. Finally, we perform a real-world risk assessment by simulating the effect of carbohydrate predictions on the blood glucose levels of individuals with type 1 diabetes. Our work highlights the opportunities and challenges of using LLMs for nutrition estimation, demonstrating their potential to aid professionals and laypersons and improve health outcomes. Our benchmark is publicly available at: https://mehak126.github.io/nutribench.html | Large Language Models, Nutrition Estimation, Dataset and Benchmark, AI for healthcare | null | 12,467 | null | [
-0.04233473539352417,
-0.06898479163646698,
0.005452344659715891,
0.08612502366304398,
0.043958209455013275,
0.06775949150323868,
0.02363319881260395,
0.04494965448975563,
0.03364619240164757,
-0.062050145119428635,
-0.062470532953739166,
-0.0910850539803505,
-0.024045411497354507,
-0.04696166142821312,
0.019244316965341568,
-0.07443597167730331,
0.14470991492271423,
-0.003483508713543415,
-0.057816337794065475,
-0.04938758164644241,
0.10792157799005508,
0.05893765017390251,
0.09187246859073639,
0.000901444407645613,
0.10206031054258347,
-0.07474871724843979,
0.008569768629968166,
-0.035430219024419785,
-0.04897841066122055,
0.04979207366704941,
0.05167299136519432,
0.028686339035630226,
0.08382585644721985,
0.014033826068043709,
-0.06209611892700195,
-0.012426529079675674,
-0.042223598808050156,
-0.05616480112075806,
-0.10081059485673904,
-0.060408588498830795,
-0.004776477348059416,
-0.021603550761938095,
0.024415485560894012,
0.018237203359603882,
0.08352120220661163,
-0.09783072024583817,
-0.012970522046089172,
0.030901342630386353,
-0.017464516684412956,
0.0471847727894783,
-0.13842856884002686,
-0.09302765876054764,
0.03427540510892868,
-0.01272846944630146,
0.014073008671402931,
0.006908380892127752,
-0.030169209465384483,
-0.05882252752780914,
-0.07189194858074188,
-0.0007695755921304226,
-0.03173598274588585,
-0.09093805402517319,
-0.03353703394532204,
0.006705339998006821,
0.04260743409395218,
0.016325967386364937,
-0.020820917561650276,
0.06982651352882385,
-0.029346588999032974,
0.011331028304994106,
-0.09142088890075684,
0.009299921803176403,
0.001719564781524241,
0.09069390594959259,
0.021524762734770775,
0.02740546688437462,
0.011058416217565536,
-0.023676134645938873,
0.047472380101680756,
-0.0294174924492836,
-0.025706589221954346,
0.012839023023843765,
0.12820817530155182,
0.03395474702119827,
0.07080239802598953,
0.022318007424473763,
0.04070543870329857,
0.09642691165208817,
-0.03432286158204079,
0.027113672345876694,
-0.023632556200027466,
-0.05354837700724602,
0.10287889838218689,
0.09797576814889908,
-0.013644288294017315,
0.03495018184185028,
-0.02567407675087452,
-0.1502879559993744,
0.05100589990615845,
0.02832644060254097,
0.0017701706383377314,
0.05802299082279205,
0.06088678166270256,
-0.05738440901041031,
-0.07085822522640228,
-0.015954067930579185,
0.0017651909729465842,
0.050867512822151184,
0.04608425125479698,
0.01785886101424694,
0.04594028741121292,
0.0996592715382576,
-0.0024194205179810524,
0.04056459665298462,
-0.040644433349370956,
0.0016895728185772896,
0.051209915429353714,
-0.07624609768390656,
0.028634730726480484,
0.08097068220376968,
-0.07898547500371933,
0.055253371596336365,
0.08793053776025772,
-0.02943497896194458,
0.031768038868904114,
0.04730856418609619,
0.0016959502827376127,
1.223586028766335e-33,
0.016669170930981636,
-0.0329720564186573,
0.09095683693885803,
0.08053945749998093,
0.06898436695337296,
0.0013678891118615866,
-0.04597032815217972,
0.027414264157414436,
0.03566833212971687,
0.0019037056481465697,
-0.011241085827350616,
-0.01518274750560522,
-0.022349515929818153,
0.04612558335065842,
0.03416455537080765,
0.04344862699508667,
-0.07209624350070953,
0.057629287242889404,
0.015406535007059574,
0.06110250577330589,
-0.021712014451622963,
-0.025339625775814056,
0.062879778444767,
-0.037016745656728745,
-0.00817926600575447,
0.07976573705673218,
0.03141153231263161,
-0.0268968865275383,
-0.012678011320531368,
0.017941446974873543,
-0.09827902913093567,
-0.03126916661858559,
-0.019447242841124535,
-0.03297194093465805,
-0.02037644572556019,
-0.05639445036649704,
-0.01034994050860405,
-0.05559415742754936,
-0.009040199220180511,
0.00030980337760411203,
0.020197665318846703,
-0.013325061649084091,
0.05376512557268143,
-0.050873637199401855,
-0.03406898304820061,
-0.006017487030476332,
0.0014576837420463562,
0.06080576777458191,
-0.006843548733741045,
0.015466771088540554,
-0.003761179745197296,
-0.02951658144593239,
0.025772565975785255,
0.0874929130077362,
-0.06621751189231873,
0.026514314115047455,
-0.0256035253405571,
0.01530515868216753,
-0.02437758259475231,
0.012296143919229507,
0.013272713869810104,
0.08778824657201767,
0.04104957357048988,
-0.011872394941747189,
-0.021338922902941704,
0.018899645656347275,
-0.15091367065906525,
-0.04127481207251549,
-0.014614704065024853,
0.05763648822903633,
0.024653002619743347,
-0.014990760013461113,
0.03722899407148361,
0.03439654782414436,
0.04712517932057381,
-0.05013737455010414,
0.028847133740782738,
-0.04653506726026535,
-0.0023078746162354946,
0.04131411388516426,
0.08814990520477295,
-0.035747017711400986,
-0.006550411693751812,
-0.015518435277044773,
-0.09225854277610779,
0.04204322025179863,
-0.02005208283662796,
-0.08860519528388977,
0.04920496046543121,
-0.0377781055867672,
-0.06314168870449066,
0.023659493774175644,
-0.026978248730301857,
0.040336549282073975,
-0.07029982656240463,
-1.4907535786826746e-33,
-0.08772283792495728,
-0.00012214487651363015,
0.04282243922352791,
0.08734088391065598,
0.03030097484588623,
-0.1457451581954956,
0.00707293301820755,
-0.02986126020550728,
0.0819140076637268,
-0.07976879924535751,
0.05436764284968376,
-0.006805735640227795,
0.05984074994921684,
-0.01312226802110672,
0.07433474063873291,
0.09073039144277573,
-0.061766188591718674,
-0.0007022804929874837,
-0.007488216273486614,
0.015527435578405857,
-0.047629088163375854,
0.03306165337562561,
-0.05983660742640495,
0.04009800776839256,
-0.012941712513566017,
0.05284218117594719,
0.05353790149092674,
0.052755068987607956,
-0.031159862875938416,
-0.06358355283737183,
0.017070723697543144,
-0.011827805079519749,
-0.05426502600312233,
-0.06908634305000305,
-0.05440506339073181,
-0.004403762985020876,
-0.03369208052754402,
-0.08608347177505493,
-0.039649877697229385,
0.0779900774359703,
0.06083447113633156,
0.021622642874717712,
-0.09998626261949539,
-0.05362658575177193,
-0.037220727652311325,
-0.013297145254909992,
-0.0283215194940567,
-0.07117407023906708,
0.03613843768835068,
0.013287987560033798,
0.04991328343749046,
-0.005057519767433405,
-0.017529942095279694,
0.01489524357020855,
-0.018061164766550064,
-0.05837033689022064,
-0.029000326991081238,
-0.05316394194960594,
0.00023061622050590813,
-0.018368879333138466,
-0.10795897245407104,
0.06517813354730606,
0.03405732661485672,
-0.0023869136348366737,
0.03796815127134323,
-0.12030826508998871,
0.002272870624437928,
-0.006040933076292276,
0.004637287929654121,
0.014940671622753143,
0.062411852180957794,
-0.04097417742013931,
0.018486760556697845,
0.02985689975321293,
0.009400502778589725,
0.033730026334524155,
-0.020273447036743164,
-0.013022124767303467,
-0.05393868312239647,
-0.08048060536384583,
0.018761083483695984,
-0.03669125214219093,
0.052945952862501144,
0.01736016571521759,
0.033857692033052444,
0.02407032996416092,
0.008032238110899925,
0.0417974516749382,
0.019136963412165642,
0.08351319283246994,
-0.11686749011278152,
0.027508359402418137,
-0.01734563708305359,
0.05909763649106026,
0.014593513682484627,
-3.621265420861164e-8,
0.06579970568418503,
-0.05314764007925987,
-0.05885262414813042,
0.0615704283118248,
-0.026632854714989662,
0.002263388130813837,
-0.06165338307619095,
0.01872663013637066,
-0.012028006836771965,
0.005528280511498451,
0.04908611997961998,
0.06953049451112747,
-0.06785568594932556,
-0.006509588565677404,
-0.038657546043395996,
0.03171848878264427,
0.015858260914683342,
0.06932754069566727,
-0.04156631603837013,
-0.010182719677686691,
-0.0030276936013251543,
0.06573522835969925,
0.008827232755720615,
-0.045707352459430695,
0.05887315049767494,
-0.062041718512773514,
-0.05225878208875656,
0.1095515564084053,
0.05182715505361557,
0.006731256376951933,
0.00861325766891241,
0.02394951321184635,
-0.00850214809179306,
0.0659903883934021,
0.020675215870141983,
0.023525075986981392,
-0.09345650672912598,
-0.02760121412575245,
0.01769140176475048,
0.027869611978530884,
0.016748204827308655,
0.012317942455410957,
-0.04203261807560921,
0.027845121920108795,
-0.004349735099822283,
-0.013579471968114376,
-0.06673632562160492,
-0.04439361020922661,
0.056527409702539444,
-0.013255417346954346,
-0.06841357797384262,
0.01786414347589016,
-0.03996962681412697,
0.04661199077963829,
0.019970888271927834,
0.05072973296046257,
-0.031141093000769615,
-0.02195795066654682,
0.0776779055595398,
-0.0335213877260685,
0.037949539721012115,
-0.038452789187431335,
-0.081387460231781,
-0.041627075523138046
] | 0 | 0 | 0 | 0 |
|
GOttack: Universal Adversarial Attacks on Graph Neural Networks via Graph Orbits Learning | https://openreview.net/forum?id=YbURbViE7l | [
"Zulfikar Alom",
"Tran Gia Bao Ngo",
"Murat Kantarcioglu",
"Cuneyt Gurcan Akcora"
] | Poster | Graph Neural Networks (GNNs) have demonstrated superior performance in node classification tasks across diverse applications. However, their vulnerability to adversarial attacks, where minor perturbations can mislead model predictions, poses significant challenges. This study introduces GOttack, a novel adversarial attack framework that exploits the topological structure of graphs to undermine the integrity of GNN predictions systematically.
By defining a topology-aware method to manipulate graph orbits, our approach generates adversarial modifications that are both subtle and effective, posing a severe test to the robustness of GNNs. We evaluate the efficacy of GOttack across multiple prominent GNN architectures using standard benchmark datasets. Our results show that GOttack outperforms existing state-of-the-art adversarial techniques and completes training in approximately 55% of the time required by the fastest competing model, achieving the highest average misclassification rate in 155 tasks.
This work not only sheds light on the susceptibility of GNNs to structured adversarial attacks
but also shows that certain topological patterns may play a significant role in the underlying robustness of the GNNs. Our Python implementation is shared at https://github.com/cakcora/GOttack. | graphlet, orbit, adversarial machine learning, graph mining, graph convolutional networks, semi-supervised learning | We identify an equivalence group for graph nodes and show that gradient-based attack models predominantly employ the group in their selection. | 12,466 | null | [
-0.12066149711608887,
-0.04052377864718437,
0.0056440806947648525,
0.06156688928604126,
0.05922620743513107,
0.0036880371626466513,
-0.05912770703434944,
-0.12002132087945938,
-0.05019433796405792,
-0.009955192916095257,
-0.03395065292716026,
-0.04321906715631485,
0.04200790077447891,
0.02514999359846115,
-0.06381252408027649,
-0.047649648040533066,
0.07846483588218689,
0.09813140332698822,
0.000596194586250931,
0.023061003535985947,
-0.034425247460603714,
0.017357230186462402,
-0.001779688405804336,
0.018345100805163383,
0.04185910522937775,
-0.02591494657099247,
-0.070745088160038,
-0.020582161843776703,
-0.00846738088876009,
-0.05736308544874191,
0.06248130649328232,
-0.033081308007240295,
-0.09540128707885742,
0.04381481930613518,
-0.04442386329174042,
0.003499352838844061,
-0.033724844455718994,
0.027128690853714943,
0.08322631567716599,
-0.08205616474151611,
0.04251987114548683,
-0.0267716646194458,
0.058560580015182495,
0.024366382509469986,
0.024797165766358376,
0.015252191573381424,
-0.01944645121693611,
-0.017868254333734512,
0.005055786110460758,
-0.008429402485489845,
-0.02081742137670517,
-0.04501085355877876,
-0.022993646562099457,
-0.0002029548486461863,
-0.009090558625757694,
-0.019390594214200974,
0.037084948271512985,
0.04708608612418175,
0.031063081696629524,
0.056364789605140686,
0.023670397698879242,
-0.06811244785785675,
-0.03809948265552521,
0.05138559639453888,
-0.03864821046590805,
-0.009163854643702507,
0.011683241464197636,
0.06516016274690628,
0.005738629028201103,
0.07250180095434189,
0.1268385946750641,
0.06877882033586502,
-0.07050910592079163,
-0.018550995737314224,
0.10556696355342865,
0.12884372472763062,
-0.017115747556090355,
0.005691370461136103,
0.012889792211353779,
-0.09290844202041626,
0.08382055908441544,
0.01665196567773819,
0.026514476165175438,
0.06525618582963943,
0.07536150515079498,
0.00874598603695631,
-0.0823134332895279,
-0.013902345672249794,
0.028650633990764618,
-0.001514056813903153,
-0.04374723136425018,
-0.017702899873256683,
0.03075632080435753,
0.012673386372625828,
0.059985123574733734,
0.028848422691226006,
-0.015620187856256962,
-0.035613808780908585,
-0.08147092908620834,
0.0833749994635582,
-0.05795546621084213,
-0.035164643079042435,
0.026807714253664017,
-0.07774877548217773,
0.0742814764380455,
0.05757246911525726,
0.05148034915328026,
0.0026275767013430595,
0.04660164192318916,
-0.06673168390989304,
0.02797424979507923,
0.12546345591545105,
0.03294629231095314,
-0.020722374320030212,
-0.02108795754611492,
-0.022028235718607903,
-0.04251561313867569,
-0.0010398316662758589,
-0.01025232020765543,
0.1318383812904358,
-0.0369718074798584,
0.05649368837475777,
0.006795571185648441,
-0.030226590111851692,
-0.019681090489029884,
-0.021565156057476997,
-0.09891007095575333,
2.3959491565116197e-33,
0.053770098835229874,
0.020027825608849525,
0.023635677993297577,
-0.053549688309431076,
0.08805257081985474,
0.00803715642541647,
-0.05069689825177193,
0.027752021327614784,
0.02285769209265709,
0.046092621982097626,
-0.08074774593114853,
-0.025021355599164963,
-0.052708301693201065,
0.08872319757938385,
0.07267005741596222,
0.1165759265422821,
0.03837735205888748,
-0.04005338251590729,
-0.045641228556632996,
-0.048940468579530716,
0.05424709990620613,
-0.04526238515973091,
-0.024247201159596443,
-0.00396061223000288,
0.0047116330824792385,
0.08018729835748672,
-0.0002420380769763142,
0.04374213516712189,
-0.007152520585805178,
0.03662535920739174,
-0.08105602860450745,
0.04771439731121063,
0.04387548193335533,
0.11296258121728897,
-0.009463704191148281,
-0.00019662437262013555,
-0.019791865721344948,
-0.029938261955976486,
-0.02288944646716118,
-0.032566316425800323,
-0.027121856808662415,
0.03013988584280014,
-0.04809102043509483,
-0.023185253143310547,
0.013809029012918472,
-0.0025185775011777878,
0.0065221828408539295,
-0.022867048159241676,
-0.024438681080937386,
-0.04665811359882355,
-0.012733790092170238,
-0.0362628772854805,
-0.035417165607213974,
0.05053091049194336,
-0.031696077436208725,
-0.07047904282808304,
0.05751213803887367,
0.05545201152563095,
0.004283546004444361,
0.1148904412984848,
0.030283542349934578,
-0.02229609712958336,
-0.0021681201178580523,
0.0032491059973835945,
-0.04233269765973091,
-0.04706243798136711,
-0.027225393801927567,
0.004344778135418892,
0.00831932108849287,
-0.019000180065631866,
-0.04895959049463272,
0.06830678880214691,
-0.04908150061964989,
-0.04048280790448189,
-0.0035528959706425667,
-0.030156902968883514,
0.021355561912059784,
-0.027952633798122406,
-0.035197798162698746,
0.024401631206274033,
0.019354645162820816,
-0.08412212878465652,
0.04353000223636627,
0.03553759679198265,
-0.08374033868312836,
-0.026209481060504913,
0.02612626552581787,
0.00014047297008801252,
0.04153870418667793,
0.025369275361299515,
-0.02747352048754692,
-0.07004652172327042,
-0.012822347693145275,
-0.010061494074761868,
-0.039365172386169434,
-2.288585930992243e-33,
-0.04250920191407204,
0.1031750813126564,
-0.024638433009386063,
0.048024605959653854,
-0.030830463394522667,
-0.06370347738265991,
0.00002101723657688126,
0.04431886970996857,
-0.08236277103424072,
0.06195038929581642,
-0.061264727264642715,
-0.02470371127128601,
0.09369990229606628,
-0.06900572031736374,
0.054506316781044006,
-0.0735420286655426,
-0.033561281859874725,
0.030283043161034584,
-0.0330730564892292,
-0.013616520911455154,
0.07461713254451752,
0.06845250725746155,
-0.07506255805492401,
0.01725020818412304,
0.09661491215229034,
0.0021561793982982635,
-0.07620120048522949,
0.0420084223151207,
0.03163159638643265,
0.08075247704982758,
-0.06470613926649094,
0.014415930025279522,
-0.036231376230716705,
-0.030345981940627098,
0.053080324083566666,
0.08813810348510742,
0.10986004769802094,
0.04203822463750839,
-0.028361203148961067,
0.013370193541049957,
-0.011446408927440643,
-0.025437306612730026,
-0.11799170076847076,
0.030952922999858856,
0.022513559088110924,
0.04093111678957939,
-0.027925565838813782,
0.06299258768558502,
-0.05621759966015816,
-0.0567827932536602,
-0.0113362492993474,
-0.01068361010402441,
0.019131118431687355,
0.04527058079838753,
-0.007090441416949034,
-0.04796726256608963,
0.05801913142204285,
0.01482168585062027,
0.03753041476011276,
0.02666894532740116,
-0.040458764880895615,
-0.055623799562454224,
-0.05823701620101929,
0.028375942260026932,
-0.04405922815203667,
-0.010346505790948868,
-0.028012385591864586,
0.04762760177254677,
-0.009860632941126823,
0.06536887586116791,
0.0015256392071023583,
-0.024562615901231766,
-0.03991319611668587,
0.021632077172398567,
-0.03356814384460449,
0.0024498761631548405,
0.015409301035106182,
-0.07401376962661743,
0.00035579290124587715,
0.00898771733045578,
0.022048788145184517,
-0.05103876069188118,
0.024693138897418976,
0.03079616278409958,
0.09459797292947769,
0.0753437951207161,
0.06326531618833542,
0.05350938066840172,
0.0030880612321197987,
0.056121714413166046,
-0.017405692487955093,
0.011773327365517616,
-0.11783044785261154,
0.054592568427324295,
0.013619472272694111,
-3.743464560557186e-8,
-0.13200277090072632,
0.04085884243249893,
0.03486139699816704,
0.005279595497995615,
0.012256008572876453,
0.003508138470351696,
0.03735784441232681,
0.105087511241436,
-0.02852071076631546,
0.03946307301521301,
0.06905487179756165,
0.0016842121258378029,
0.0029133285861462355,
-0.09127301722764969,
-0.02806706167757511,
0.024974487721920013,
0.014697003178298473,
0.01280767098069191,
-0.027645127847790718,
0.04204809293150902,
-0.041686296463012695,
-0.021268948912620544,
-0.031800881028175354,
-0.07240206748247147,
0.004457787144929171,
-0.15129370987415314,
-0.05365760624408722,
0.06620737910270691,
0.03770941495895386,
0.04647241532802582,
-0.1044212281703949,
-0.03814869001507759,
0.04624457657337189,
-0.012003391981124878,
-0.07128722965717316,
0.17032520473003387,
-0.02840900793671608,
-0.01883753202855587,
0.02765808254480362,
-0.006661420688033104,
-0.03518320620059967,
0.023533329367637634,
0.019015278667211533,
-0.02510685659945011,
-0.10179430991411209,
0.0096199382096529,
0.07462732493877411,
-0.01215208601206541,
0.09621680527925491,
-0.02232988364994526,
-0.004845773801207542,
-0.0143737206235528,
-0.03926003351807594,
0.02049245312809944,
-0.011466697789728642,
-0.006057547405362129,
-0.031063320115208626,
-0.10980944335460663,
0.011759045533835888,
0.04984427988529205,
0.01683180034160614,
-0.008085424080491066,
-0.058980487287044525,
-0.06213699281215668
] | 0 | 0 | 0 | 0 |
|
NatureLM-audio: an Audio-Language Foundation Model for Bioacoustics | https://openreview.net/forum?id=hJVdwBpWjt | [
"David Robinson",
"Marius Miron",
"Masato Hagiwara",
"Olivier Pietquin"
] | Poster | Large language models (LLMs) prompted with text and audio have achieved state-of-the-art performance across various auditory tasks, including speech, music, and general audio, showing emergent abilities on unseen tasks. However, their potential has yet to be fully demonstrated in bioacoustics tasks, such as detecting animal vocalizations in large recordings, classifying rare and endangered species, and labeling context and behavior—tasks that are crucial for conservation, biodiversity monitoring, and animal behavior studies. In this work, we present NatureLM-audio, the first audio-language foundation model specifically designed for bioacoustics. Our training dataset consists of carefully curated text-audio pairs spanning bioacoustics, speech, and music, designed to address the field's limited availability of annotated data. We demonstrate successful transfer of learned representations from music and speech to bioacoustics, and our model shows promising generalization to unseen taxa and tasks. We evaluate NatureLM-audio on a novel benchmark (BEANS-Zero) and it sets a new state of the art on several bioacoustics tasks, including zero-shot classification of unseen species. To advance bioacoustics research, we release our model weights, benchmark data, and open-source the code for training and benchmark data generation and model training. | audio-language foundation models, multimodal large language models (llms), bioacoustics, animal vocalizations, zero-shot learning, in-context learning | null | 12,454 | null | [
-0.03263198956847191,
-0.08888910710811615,
0.03181549161672592,
0.002879038918763399,
0.08744728565216064,
0.06155732646584511,
-0.026226477697491646,
-0.12869581580162048,
0.04403405636548996,
-0.030048735439777374,
-0.06534828990697861,
-0.16170866787433624,
-0.028260722756385803,
0.0537940189242363,
-0.005567967426031828,
0.009692699648439884,
0.057095035910606384,
0.038912560790777206,
-0.027712002396583557,
0.0061196512542665005,
0.08827988803386688,
0.15492035448551178,
0.09205035120248795,
0.029681440442800522,
0.016331292688846588,
-0.010124661959707737,
-0.07100539654493332,
0.004054375924170017,
-0.007143862545490265,
-0.002282568020746112,
0.03697514533996582,
0.023496124893426895,
0.11629495769739151,
-0.01752510294318199,
-0.05868871137499809,
0.0031217976938933134,
-0.0024049493949860334,
-0.022991430014371872,
0.015881912782788277,
-0.002219769638031721,
0.022879280149936676,
0.046940430998802185,
0.007479664403945208,
-0.04998518526554108,
-0.03710893914103508,
-0.051967330276966095,
-0.07753117382526398,
-0.0940777063369751,
0.03509056195616722,
0.022422874346375465,
-0.06370653212070465,
-0.06856108456850052,
-0.05276225507259369,
0.06972917914390564,
-0.03696248680353165,
0.012150697410106659,
0.0860513374209404,
-0.04596596583724022,
0.010862641036510468,
-0.03211655467748642,
0.033522624522447586,
-0.11590086668729782,
-0.02692168764770031,
0.01076879259198904,
0.0068923053331673145,
-0.03206393122673035,
-0.08436371386051178,
0.025261815637350082,
0.0511745922267437,
-0.007638855837285519,
-0.019524360075592995,
0.035447411239147186,
0.04214981943368912,
0.06668972969055176,
0.02479456178843975,
0.06497634947299957,
0.030216887593269348,
0.02483323961496353,
0.046630579978227615,
-0.06430506706237793,
0.017764946445822716,
-0.05791127309203148,
0.008419019170105457,
-0.04040079563856125,
0.11925714462995529,
0.02581179328262806,
0.03594624996185303,
0.003606329206377268,
-0.0789535716176033,
0.0008887150324881077,
-0.09255920350551605,
-0.056989703327417374,
0.06987107545137405,
0.005933679174631834,
-0.02803283929824829,
0.0423201322555542,
0.0251155998557806,
-0.05658898875117302,
0.03905033692717552,
0.03692881390452385,
0.0035132719203829765,
0.014812787063419819,
-0.0206240713596344,
-0.012297753244638443,
0.024882540106773376,
-0.01799359731376171,
-0.06902075558900833,
0.023703768849372864,
0.04984121397137642,
-0.03309202194213867,
0.046870894730091095,
0.04550132527947426,
-0.02995157428085804,
0.11660531908273697,
0.062239862978458405,
0.022781984880566597,
-0.04525773972272873,
-0.0964430719614029,
0.03191537782549858,
0.01874694973230362,
-0.010212515480816364,
0.007262470666319132,
-0.04527383670210838,
-0.033780958503484726,
0.02297421358525753,
0.001219119061715901,
-0.08191477507352829,
-2.9753010982059846e-34,
0.11371788382530212,
0.04174076393246651,
0.04697718843817711,
-0.008322612382471561,
0.08027856796979904,
-0.10398116707801819,
-0.08240681886672974,
-0.026967203244566917,
-0.044794075191020966,
0.03921213746070862,
-0.0421096608042717,
0.057552505284547806,
-0.0258374884724617,
0.01834757998585701,
0.001921928720548749,
0.003447101917117834,
-0.08070727437734604,
0.02098223567008972,
0.03880989924073219,
-0.024369146674871445,
-0.014649118296802044,
0.05758088082075119,
0.046582743525505066,
0.01641952618956566,
0.11883536726236343,
-0.009150771424174309,
0.03636967018246651,
-0.1896193027496338,
-0.032792843878269196,
0.02891400270164013,
-0.10372080653905869,
-0.07298047095537186,
0.03737756237387657,
0.005480178166180849,
0.034801337867975235,
-0.05156778544187546,
0.03082018531858921,
-0.0353240892291069,
0.022267363965511322,
-0.009866134263575077,
-0.023414144292473793,
0.028106024488806725,
-0.006455263588577509,
-0.07266059517860413,
-0.05428899824619293,
-0.0379021093249321,
0.020057542249560356,
0.13843080401420593,
-0.034668534994125366,
0.05719691887497902,
0.009356371127068996,
-0.00841513741761446,
0.05253731459379196,
-0.022618727758526802,
-0.050771668553352356,
0.06214331462979317,
0.06120661273598671,
-0.0032969273161143064,
0.00617348263040185,
-0.07827410101890564,
0.06904863566160202,
0.040797412395477295,
0.09412838518619537,
-0.01599583774805069,
0.07582119107246399,
0.04476884752511978,
-0.04473388195037842,
0.02486913837492466,
0.07105516642332077,
-0.0390453040599823,
-0.029731107875704765,
0.018919534981250763,
-0.01685628481209278,
-0.07538974285125732,
0.06390459090471268,
0.01378580741584301,
0.09436576813459396,
-0.1177288293838501,
-0.022769702598452568,
0.07411091774702072,
0.027741748839616776,
0.02550005353987217,
-0.0495515838265419,
-0.021007705479860306,
-0.04159724712371826,
0.008631839416921139,
0.04042544588446617,
-0.11745935678482056,
-0.050172142684459686,
-0.02358073741197586,
-0.0885763019323349,
0.03570634126663208,
-0.06305769085884094,
-0.01661701686680317,
-0.0239868126809597,
8.442567453414306e-35,
-0.001590954721905291,
0.021063385531306267,
-0.00202547968365252,
0.05534584820270538,
-0.014166873879730701,
0.011387418024241924,
0.03919033333659172,
0.056275516748428345,
0.005261839833110571,
-0.04231475666165352,
0.014644129201769829,
0.03028440661728382,
0.07932574301958084,
-0.04868777468800545,
-0.03800312057137489,
-0.03941108286380768,
-0.09517263621091843,
0.04667399823665619,
0.019691841676831245,
0.013031032867729664,
-0.04139098897576332,
0.027952680364251137,
-0.05142797529697418,
0.011935655027627945,
-0.04723278060555458,
0.04588194936513901,
-0.046985600143671036,
0.021459486335515976,
-0.00829649530351162,
-0.059104613959789276,
0.021413637325167656,
-0.008686876855790615,
-0.07286019623279572,
-0.11354882270097733,
-0.03868088498711586,
0.047714099287986755,
0.052784085273742676,
-0.04633798077702522,
0.040111344307661057,
0.039984606206417084,
0.03120408020913601,
0.03736191242933273,
-0.058793745934963226,
-0.0247294083237648,
-0.009711451828479767,
-0.056875891983509064,
-0.04368885979056358,
0.07607920467853546,
-0.023623621091246605,
0.011319609358906746,
0.09878117591142654,
0.02255230024456978,
0.006826038938015699,
0.016035014763474464,
-0.008077415637671947,
-0.08186864852905273,
0.020769784227013588,
-0.047316476702690125,
0.025871723890304565,
-0.009172058664262295,
-0.0703202337026596,
0.01977860927581787,
-0.07213793694972992,
0.0018412582576274872,
-0.016953231766819954,
0.014007522724568844,
0.014631190337240696,
0.011844340711832047,
0.02650930918753147,
-0.02727915532886982,
0.06815546751022339,
0.034231577068567276,
-0.020189551636576653,
0.011030877940356731,
-0.008278466761112213,
0.07026033848524094,
-0.0734814703464508,
-0.09847185015678406,
-0.03333890810608864,
0.021005779504776,
0.021269137039780617,
-0.03993216156959534,
0.014977444894611835,
0.014871702529489994,
0.1550533026456833,
0.08610881865024567,
0.022659869864583015,
0.019890721887350082,
0.009030490182340145,
0.05861017853021622,
-0.038255248218774796,
0.0824955627322197,
0.0396743007004261,
0.018003791570663452,
0.0021110426168888807,
-3.529320125039703e-8,
-0.0386456698179245,
0.06293310225009918,
-0.00702283950522542,
-0.012064987793564796,
-0.002127522835507989,
-0.03410269692540169,
0.005957671441137791,
-0.06990483403205872,
-0.0030505263712257147,
-0.04374711215496063,
0.0969095528125763,
-0.027695665135979652,
-0.03266020864248276,
0.05032585188746452,
0.0004314056714065373,
0.009218634106218815,
0.027025535702705383,
0.06890532374382019,
-0.03971090540289879,
-0.006887788884341717,
-0.020058242604136467,
0.07686588913202286,
0.009761306457221508,
-0.07676799595355988,
0.04424354434013367,
-0.05707772821187973,
0.040651798248291016,
-0.02914818562567234,
0.07605520635843277,
0.015789398923516273,
-0.061782319098711014,
0.07531406730413437,
-0.02170608937740326,
-0.025258362293243408,
0.005106095224618912,
0.044535115361213684,
-0.049528419971466064,
-0.03876651078462601,
-0.019207142293453217,
-0.01374345738440752,
0.011278879828751087,
0.08906290680170059,
-0.07774271816015244,
0.004076709505170584,
-0.023333560675382614,
-0.019485676661133766,
0.02023547887802124,
-0.05118861049413681,
0.021602995693683624,
0.048598513007164,
-0.07628446817398071,
0.008189590647816658,
-0.01692379079759121,
0.030539005994796753,
0.04494099318981171,
0.0692453607916832,
-0.016201166436076164,
0.01496249157935381,
0.015381013974547386,
0.03676258400082588,
0.02470589056611061,
-0.02851172909140587,
-0.053505342453718185,
-0.015544559806585312
] | 0 | 0 | 0 | 0 |
|
Privacy-Preserving Personalized Federated Prompt Learning for Multimodal Large Language Models | https://openreview.net/forum?id=Equ277PBN0 | [
"Linh Tran",
"Wei Sun",
"Stacy Patterson",
"Ana Milanova"
] | Poster | Multimodal Large Language Models (LLMs) are pivotal in revolutionizing customer support and operations by integrating multiple modalities such as text, images, and audio. Federated Prompt Learning (FPL) is a recently proposed approach that combines pre-trained multimodal LLMs such as vision-language models with federated learning to create personalized, privacy-preserving AI systems. However, balancing the competing goals of personalization, generalization, and privacy remains a significant challenge. Over-personalization can lead to overfitting, reducing generalizability, while stringent privacy measures, such as differential privacy, can hinder both personalization and generalization. In this paper, we propose a Differentially Private Federated Prompt Learning (DP-FPL) approach to tackle this challenge by leveraging a low-rank factorization scheme to capture generalization while maintaining a residual term that preserves expressiveness for personalization. To ensure privacy, we introduce a novel method where we apply local differential privacy to the two low-rank components of the local prompt, and global differential privacy to the global prompt. Our approach mitigates the impact of privacy noise on the model performance while balancing the tradeoff between personalization and generalization. Extensive experiments demonstrate the effectiveness of our approach over other benchmarks. | Multimodal Large Language Model, Federated Prompt Learning, Personalization, Differential Privacy | We propose a differentially private approach for federated prompt learning using multimodal large language model by leveraging low-rank factorization to balance personalization, generalization and privacy. | 12,452 | 2501.13904 | [
-0.03891139104962349,
-0.10230989009141922,
0.04790545254945755,
0.019611140713095665,
0.13903270661830902,
0.030602533370256424,
0.020418642088770866,
-0.016959205269813538,
0.02621396631002426,
-0.04965856671333313,
0.007213034201413393,
-0.03962276130914688,
0.051191285252571106,
0.014920021407306194,
0.03854912519454956,
-0.02887716330587864,
0.07525546848773956,
0.06701483577489853,
-0.15447869896888733,
-0.027272436767816544,
-0.021911785006523132,
-0.0016450595576316118,
0.05995624512434006,
0.0025950863491743803,
0.04651057347655296,
0.005472594406455755,
0.047627273947000504,
-0.02133581042289734,
0.016363080590963364,
-0.0292609054595232,
0.07524561882019043,
0.031761284917593,
0.019732242450118065,
0.011964930221438408,
0.004260118585079908,
0.035158220678567886,
-0.09059787541627884,
0.02484292723238468,
-0.05912313982844353,
-0.06473641097545624,
-0.0035195685923099518,
-0.056877270340919495,
-0.01765654981136322,
-0.011065431870520115,
0.01582907699048519,
-0.001415671780705452,
-0.0058281924575567245,
-0.016198420897126198,
-0.008339405991137028,
0.042601313441991806,
-0.04563514143228531,
-0.08178240805864334,
-0.054270874708890915,
0.04361112043261528,
-0.0032142912968993187,
-0.04077611491084099,
0.005290421657264233,
-0.04825004190206528,
-0.012432863935828209,
0.012820866890251637,
-0.07826720178127289,
-0.12112855166196823,
-0.03856239840388298,
0.016876472160220146,
0.011605490930378437,
0.02656358666718006,
0.010741110891103745,
0.012320506386458874,
-0.03534210845828056,
0.022073514759540558,
-0.044695738703012466,
0.07764150202274323,
0.02217140980064869,
0.04603087157011032,
0.024893980473279953,
0.009094749577343464,
-0.024093395099043846,
-0.04387650266289711,
0.005386000964790583,
-0.06399767100811005,
0.00435654865577817,
0.04327833652496338,
0.06087132543325424,
-0.08495888113975525,
0.07708336412906647,
0.007895533926784992,
0.005824041552841663,
-0.030608411878347397,
0.005037212278693914,
0.02439490146934986,
-0.05800824984908104,
0.0007102545932866633,
0.017425861209630966,
0.008381722494959831,
-0.013494244776666164,
0.020097240805625916,
-0.014774451963603497,
-0.12892630696296692,
0.01901591382920742,
0.08337858319282532,
-0.03297772258520126,
0.04369615763425827,
-0.021469829604029655,
0.001991653349250555,
0.047933757305145264,
0.02308613620698452,
0.024047117680311203,
0.04877841845154762,
0.05340641736984253,
-0.0338192880153656,
0.026895251125097275,
0.0000719253221177496,
0.023268986493349075,
0.027254225686192513,
0.01782357506453991,
-0.04034394398331642,
0.05238225683569908,
0.02744479291141033,
0.1219550147652626,
-0.0076211607083678246,
0.017215639352798462,
-0.01155468262732029,
0.03309735283255577,
-0.01744205877184868,
-0.039873555302619934,
-0.03625411167740822,
-0.08752710372209549,
-8.27643715565887e-34,
0.07868973910808563,
0.03687485679984093,
-0.0032257805578410625,
0.022849498316645622,
0.05261989310383797,
-0.03706603869795799,
-0.06848632544279099,
0.04246660694479942,
-0.07463947683572769,
0.0058395760133862495,
0.052488718181848526,
0.03431565687060356,
0.0007600077078677714,
0.07962150871753693,
-0.019393058493733406,
0.046508148312568665,
0.0031245483551174402,
0.06884277611970901,
0.07674741744995117,
-0.021459857001900673,
0.09546283632516861,
0.00936082936823368,
0.020818213000893593,
0.028893791139125824,
0.06453745812177658,
-0.02772441692650318,
0.14142443239688873,
-0.050624821335077286,
0.04902906343340874,
0.02253873273730278,
-0.11531028896570206,
-0.07722026854753494,
0.04716665670275688,
0.020728357136249542,
0.09063843637704849,
-0.02427644655108452,
-0.02723879925906658,
-0.05179903283715248,
0.034443557262420654,
-0.018792180344462395,
0.015868039801716805,
-0.06505721807479858,
-0.038726091384887695,
-0.03206418454647064,
-0.1429261565208435,
0.01505910512059927,
-0.03665813431143761,
0.08815072476863861,
-0.05855083465576172,
-0.07350374758243561,
0.013083474710583687,
0.0018473780946806073,
-0.07418862730264664,
-0.057981688529253006,
-0.06292740255594254,
-0.053657419979572296,
0.002034580335021019,
-0.02444681152701378,
0.0071999188512563705,
-0.10140957683324814,
-0.038103751838207245,
0.08897661417722702,
0.037737105041742325,
0.014248985797166824,
-0.035770922899246216,
-0.03358301520347595,
0.04849722236394882,
-0.004517412278801203,
0.09196162223815918,
-0.05493564158678055,
-0.041914720088243484,
0.05116651952266693,
-0.024592511355876923,
-0.01722153276205063,
-0.008473259396851063,
-0.014548884704709053,
0.07501905411481857,
-0.0007733057136647403,
0.07558255642652512,
0.02322922646999359,
-0.003570037428289652,
0.04124899581074715,
-0.027793318033218384,
0.02371666021645069,
-0.03304694592952728,
-0.011465448886156082,
0.015654372051358223,
-0.03098350763320923,
-0.06339752674102783,
-0.03442469611763954,
-0.05298067629337311,
-0.010067430324852467,
-0.027027206495404243,
-0.010549788363277912,
-0.04837958887219429,
1.0365138399735193e-34,
-0.018034568056464195,
0.04687255248427391,
-0.05770085006952286,
0.018930397927761078,
0.040550023317337036,
-0.062407661229372025,
0.012572802603244781,
0.07887077331542969,
0.016570815816521645,
-0.05254608765244484,
-0.0516299307346344,
0.01723492704331875,
0.063697949051857,
-0.03802017867565155,
0.04015270620584488,
0.04524754360318184,
0.004110720939934254,
0.05164967477321625,
-0.07136186212301254,
0.02092432789504528,
-0.03073856793344021,
0.07060925662517548,
-0.08699661493301392,
0.08462325483560562,
0.008204991929233074,
0.02735931985080242,
-0.13504274189472198,
0.07888931035995483,
0.04894586279988289,
-0.010449044406414032,
-0.01938103511929512,
-0.02224113792181015,
-0.10194651782512665,
-0.1171504482626915,
-0.053923942148685455,
0.03887396678328514,
-0.028261177241802216,
-0.010539594106376171,
-0.05457977578043938,
0.11637146770954132,
0.07126683741807938,
-0.07159339636564255,
-0.09542427957057953,
-0.007249975576996803,
0.0361768901348114,
-0.09011482447385788,
-0.032026614993810654,
-0.07676887512207031,
0.06369024515151978,
0.053508397191762924,
-0.028716234490275383,
0.02341422811150551,
0.03286347910761833,
0.025383764877915382,
-0.10888873785734177,
-0.06352593004703522,
0.054150309413671494,
0.0089266961440444,
0.058710142970085144,
0.0058130258694291115,
0.04232073575258255,
-0.09243796765804291,
0.025487134233117104,
-0.03359869122505188,
-0.028303833678364754,
-0.0757296234369278,
-0.004154379479587078,
-0.020732831209897995,
-0.038332194089889526,
-0.02270609326660633,
0.019718460738658905,
-0.07943017035722733,
-0.03450439125299454,
0.050487250089645386,
0.010908004827797413,
0.020827889442443848,
0.0554451085627079,
-0.05681697279214859,
0.0023876961786299944,
0.05027296766638756,
0.116317518055439,
0.009767929092049599,
0.04181141406297684,
0.010380406863987446,
0.09580202400684357,
0.014153581112623215,
0.09039608389139175,
0.06827384978532791,
0.0031804819591343403,
0.009865245781838894,
0.04597118869423866,
0.01042960211634636,
-0.02308320626616478,
0.04700591787695885,
-0.03434258699417114,
-3.5606849024816256e-8,
-0.047097887843847275,
-0.013469289988279343,
0.0007121710805222392,
0.01549238245934248,
0.022862087935209274,
0.0046472251415252686,
-0.02445763535797596,
-0.01837044768035412,
0.016369152814149857,
-0.09099072217941284,
0.013784000650048256,
-0.0041742813773453236,
-0.016371583566069603,
-0.06794288754463196,
-0.04483329877257347,
0.09655746072530746,
-0.0769105777144432,
-0.008042036555707455,
-0.004520091228187084,
-0.03267934173345566,
-0.005252011585980654,
0.02031543478369713,
-0.0172281377017498,
-0.03018621727824211,
0.021387405693531036,
0.012625283561646938,
0.03081876039505005,
0.04145713523030281,
0.03810494765639305,
0.05238576605916023,
-0.019989147782325745,
-0.02326495386660099,
0.03487448766827583,
-0.023016270250082016,
-0.05452708527445793,
0.12811709940433502,
0.005146821029484272,
-0.058013640344142914,
0.017390910536050797,
-0.04515619948506355,
0.0785069391131401,
0.04129938781261444,
-0.04662254825234413,
0.053692951798439026,
-0.02135852910578251,
-0.009981561452150345,
0.03225201740860939,
-0.18663035333156586,
0.0431155301630497,
0.01692359708249569,
-0.010930920019745827,
-0.02923894114792347,
0.044966865330934525,
0.12758769094944,
0.05517128109931946,
-0.01589791476726532,
0.03926989063620567,
0.022778360173106194,
0.08231101185083389,
0.10157036036252975,
0.011136738583445549,
0.05544480308890343,
-0.033116865903139114,
0.03382648527622223
] | 0 | 0 | 0 | 0 |
|
Optimizing 4D Gaussians for Dynamic Scene Video from Single Landscape Images | https://openreview.net/forum?id=IcYDRzcccP | [
"In-Hwan Jin",
"Haesoo Choo",
"Seong-Hun Jeong",
"Park Heemoon",
"Junghwan Kim",
"Oh-joon Kwon",
"Kyeongbo Kong"
] | Poster | To achieve realistic immersion in landscape images, fluids such as water and clouds need to move within the image while revealing new scenes from various camera perspectives. Recently, a field called dynamic scene video has emerged, which combines single image animation with 3D photography. These methods use pseudo 3D space, implicitly represented with Layered Depth Images (LDIs). LDIs separate a single image into depth-based layers, which enables elements like water and clouds to move within the image while revealing new scenes from different camera perspectives. However, as landscapes typically consist of continuous elements, including fluids, the representation of a 3D space separates a landscape image into discrete layers, and it can lead to diminished depth perception and potential distortions depending on camera movement. Furthermore, due to its implicit modeling of 3D space, the output may be limited to videos in the 2D domain, potentially reducing their versatility. In this paper, we propose representing a complete 3D space for dynamic scene video by modeling explicit representations, specifically 4D Gaussians, from a single image. The framework is focused on optimizing 3D Gaussians by generating multi-view images from a single image and creating 3D motion to optimize 4D Gaussians. The most important part of proposed framework is consistent 3D motion estimation, which estimates common motion among multi-view images to bring the motion in 3D space closer to actual motions. As far as we know, this is the first attempt that considers animation while representing a complete 3D space from a single landscape image. Our model demonstrates the ability to provide realistic immersion in various landscape images through diverse experiments and metrics. Extensive experimental results are https://anonymous.4open.science/r/ICLR_3D_MOM-7B9E/README.md. | Dynamic Scene Video, 4D Gaussian | null | 12,445 | null | [
-0.012756985612213612,
-0.0785260796546936,
0.05596870929002762,
-0.0749252438545227,
0.09195193648338318,
-0.016400951892137527,
-0.009489663876593113,
-0.08728621155023575,
0.09752242267131805,
-0.05449429899454117,
-0.05464838445186615,
-0.07033893465995789,
-0.023748479783535004,
0.08300914615392685,
-0.039968110620975494,
-0.02481352910399437,
0.044042933732271194,
0.126338928937912,
-0.09926603734493256,
0.028688160702586174,
0.06978677958250046,
-0.006569319870322943,
-0.003898859955370426,
-0.019496023654937744,
0.026026351377367973,
0.019310005009174347,
0.08453270047903061,
-0.010848559439182281,
-0.00711138267070055,
0.028159186244010925,
0.08773273974657059,
0.019383084028959274,
0.01665342040359974,
-0.025653202086687088,
-0.06546653062105179,
0.021060418337583542,
-0.0040361094288527966,
0.023964734748005867,
-0.09221561998128891,
-0.015616954304277897,
0.0058328574523329735,
0.04578419402241707,
0.028784720227122307,
0.014774405397474766,
0.023497015237808228,
0.09793489426374435,
0.01869281567633152,
-0.07846487313508987,
0.096070297062397,
0.033445995301008224,
-0.04024222865700722,
-0.04397612437605858,
-0.01877482421696186,
0.015032073482871056,
-0.01377120427787304,
0.03534701094031334,
-0.00792953185737133,
-0.0664319097995758,
0.02288956195116043,
0.00046315137296915054,
0.04682088643312454,
-0.005985569674521685,
0.11666195839643478,
0.027600135654211044,
0.04173717275261879,
-0.014481952413916588,
-0.051201608031988144,
-0.022744975984096527,
-0.05172087252140045,
-0.026843257248401642,
-0.12244204431772232,
0.01658606342971325,
0.04273237660527229,
-0.08602149039506912,
-0.050003260374069214,
-0.03649305924773216,
-0.022398483008146286,
-0.06680300086736679,
-0.01301864068955183,
-0.054427530616521835,
0.10985290259122849,
0.012046613730490208,
0.023546524345874786,
0.05768444761633873,
-0.08065546303987503,
0.046527255326509476,
0.01265433244407177,
0.03777877241373062,
-0.01970943808555603,
0.004334081895649433,
-0.015472027473151684,
-0.02724243327975273,
-0.07071541994810104,
-0.05898566544055939,
0.03335593268275261,
-0.06870520114898682,
-0.01898149959743023,
-0.20329202711582184,
0.07192116230726242,
-0.027988720685243607,
0.004033464472740889,
-0.03501511737704277,
0.04911351948976517,
-0.012164289131760597,
0.09560278803110123,
0.027717560529708862,
0.04148473963141441,
0.05536952242255211,
0.03374116122722626,
0.011988223530352116,
-0.009523111395537853,
-0.026750294491648674,
-0.02794489450752735,
0.0215603057295084,
-0.02838275395333767,
0.06392348557710648,
-0.013882189057767391,
-0.034242790192365646,
-0.048176608979701996,
0.009507693350315094,
0.024310292676091194,
-0.03792434558272362,
0.0001277979463338852,
0.022793924435973167,
0.07089799642562866,
-0.04147575423121452,
0.025088200345635414,
4.255204142965254e-34,
0.01652253232896328,
0.04331083595752716,
-0.006756691262125969,
0.11065509915351868,
0.015900757163763046,
0.004335429519414902,
0.022398456931114197,
-0.027272818610072136,
-0.004258825909346342,
-0.009522585198283195,
-0.03175511956214905,
-0.03826369717717171,
0.015316157601773739,
0.06006861478090286,
0.010422544553875923,
0.021388055756688118,
-0.009674572385847569,
0.005803117528557777,
0.0734630674123764,
0.0664834976196289,
0.01994164288043976,
0.0319853238761425,
0.01810932718217373,
-0.000056908003898570314,
-0.04991662874817848,
-0.027747131884098053,
0.04642076790332794,
-0.019306043162941933,
0.05520036071538925,
-0.05164095014333725,
-0.04205489158630371,
-0.01001232024282217,
-0.05414856970310211,
-0.041656412184238434,
0.04510752111673355,
0.042147181928157806,
-0.03535261005163193,
-0.006159281823784113,
0.006001057103276253,
-0.014048460870981216,
-0.08732526749372482,
0.0517643466591835,
-0.019553842023015022,
-0.059696413576602936,
-0.06281615793704987,
0.01188473217189312,
0.05638442188501358,
0.05138234794139862,
-0.09745977818965912,
0.05805734172463417,
0.0925367921590805,
0.009408358484506607,
0.03413749486207962,
-0.03836880251765251,
-0.08551060408353806,
-0.03225036337971687,
0.06305401027202606,
-0.03263643756508827,
0.02908279560506344,
-0.005696640349924564,
-0.039935458451509476,
0.05253244936466217,
-0.08682108670473099,
0.005110863130539656,
-0.025606459006667137,
-0.021169578656554222,
0.10164538770914078,
0.0499027818441391,
0.03822914883494377,
0.045617155730724335,
-0.03213859722018242,
-0.029672643169760704,
0.07242786139249802,
-0.05302770808339119,
0.014226987957954407,
0.03279230743646622,
-0.0157286636531353,
-0.08416686952114105,
0.010117107070982456,
0.08153072744607925,
-0.11625458300113678,
0.029875127598643303,
0.030329477041959763,
0.02704791910946369,
-0.018501421436667442,
-0.003328822087496519,
-0.002765361685305834,
0.033720292150974274,
0.000819538370706141,
-0.08046194911003113,
-0.06616781651973724,
-0.06060844287276268,
0.022999273613095284,
-0.09017081558704376,
0.03500443696975708,
-7.850981669058129e-34,
0.08119261264801025,
0.009209416806697845,
-0.05467532202601433,
0.0357036367058754,
0.060988664627075195,
-0.0554422028362751,
0.020723477005958557,
0.021659178659319878,
0.03557034581899643,
-0.06945332884788513,
-0.023568082600831985,
0.0311420951038599,
-0.01500783208757639,
0.05352896451950073,
-0.02475638873875141,
0.0030344869010150433,
-0.05349958688020706,
-0.05206236243247986,
-0.055286068469285965,
0.027257895097136497,
0.06180790066719055,
0.019664881750941277,
-0.018972214311361313,
-0.0015743991825729609,
-0.039888788014650345,
0.14322924613952637,
0.06101241335272789,
0.04769404977560043,
0.004116520751267672,
-0.009254273027181625,
-0.0455058328807354,
-0.024534381926059723,
0.011640287935733795,
-0.001959022367373109,
-0.07235750555992126,
0.02405373379588127,
0.044953249394893646,
-0.019185276702046394,
-0.07903716713190079,
-0.007544414140284061,
-0.028450455516576767,
-0.0676049143075943,
0.04497269168496132,
0.009351855143904686,
-0.02091093920171261,
0.037343692034482956,
0.011178301647305489,
0.010259488597512245,
-0.044212523847818375,
0.06108450889587402,
-0.10933519154787064,
0.0923369750380516,
-0.10851720720529556,
0.032302141189575195,
0.07740043848752975,
-0.08430083841085434,
-0.10724005103111267,
-0.008630850352346897,
0.04102950915694237,
-0.02932680770754814,
0.016211969777941704,
-0.03614940121769905,
-0.05291765555739403,
0.0020658632274717093,
-0.032085925340652466,
0.0026273936964571476,
-0.050718873739242554,
-0.030402176082134247,
-0.04208246245980263,
0.05508534982800484,
-0.02587021142244339,
-0.020074650645256042,
0.04859713837504387,
0.014357478357851505,
0.057945605367422104,
-0.06144818663597107,
0.01757386140525341,
0.024876339361071587,
0.09339723736047745,
-0.017750997096300125,
0.023669246584177017,
-0.051759760826826096,
0.005007456988096237,
0.09257779270410538,
0.06743427366018295,
0.0409088134765625,
-0.026938995346426964,
-0.0505545511841774,
0.07049717754125595,
0.020343748852610588,
0.010402400977909565,
-0.01695013977587223,
0.0006066570058465004,
0.03251315653324127,
-0.004030607640743256,
-3.6409030457207336e-8,
-0.035524774342775345,
0.06176792085170746,
0.09173493832349777,
-0.06286802887916565,
-0.05891921743750572,
-0.04670516774058342,
0.07285116612911224,
0.08468428999185562,
0.0506264790892601,
-0.11899705976247787,
0.010825295932590961,
-0.04677305370569229,
-0.015008618123829365,
-0.03306413069367409,
0.020617397502064705,
0.07110723108053207,
-0.07335101813077927,
-0.05551915988326073,
-0.015845302492380142,
0.06950001418590546,
-0.05468180030584335,
0.027700187638401985,
-0.09498559683561325,
0.000033895230444613844,
-0.031982194632291794,
0.05202712491154671,
-0.009459318593144417,
-0.07511065155267715,
0.10097069293260574,
-0.08157090842723846,
0.027656342834234238,
0.02011638693511486,
0.046557001769542694,
-0.04411330074071884,
-0.04097660258412361,
-0.000762228446546942,
-0.030047861859202385,
0.015523131005465984,
-0.00294366292655468,
-0.04709145054221153,
-0.08665737509727478,
0.0946589931845665,
0.03473275527358055,
-0.002790017519146204,
0.07269424200057983,
0.056960899382829666,
0.06693784892559052,
-0.042487550526857376,
0.019210489466786385,
0.041770193725824356,
-0.07148434966802597,
-0.0034052515402436256,
0.0007406892837025225,
0.13288506865501404,
0.07911416888237,
0.038859009742736816,
0.0028125273529440165,
-0.03187549114227295,
0.021940503269433975,
0.008327917195856571,
-0.05521950498223305,
-0.023880034685134888,
-0.009006614796817303,
0.00034724155557341874
] | 0 | 0 | 0 | 0 |
|
Explore Theory of Mind: program-guided adversarial data generation for theory of mind reasoning | https://openreview.net/forum?id=246rHKUnnf | [
"Melanie Sclar",
"Jane Dwivedi-Yu",
"Maryam Fazel-Zarandi",
"Yulia Tsvetkov",
"Yonatan Bisk",
"Yejin Choi",
"Asli Celikyilmaz"
] | Poster | Do large language models (LLMs) have theory of mind? A plethora of papers and benchmarks have been introduced to evaluate if current models have been able to develop this key ability of social intelligence. However, all rely on limited datasets with simple patterns that can potentially lead to problematic blind spots in evaluation and an overestimation of model capabilities. We introduce ExploreToM, the first framework to allow large-scale generation of diverse and challenging theory of mind data for robust training and evaluation. Our approach leverages an A* search over a custom domain-specific language to produce complex story structures and novel, diverse, yet plausible scenarios to stress test the limits of LLMs. Our evaluation reveals that state-of-the-art LLMs, such as Llama-3.1-70B and GPT-4o, show accuracies as low as 0% and 9% on ExploreToM-generated data, highlighting the need for more robust theory of mind evaluation. As our generations are a conceptual superset of prior work, fine-tuning on our data yields a 27-point accuracy improvement on the classic ToMi benchmark (Le et al., 2019). ExploreToM also enables uncovering underlying skills and factors missing for models to show theory of mind, such as unreliable state tracking or data imbalances, which may contribute to models' poor performance on benchmarks. | theory of mind reasoning, adversarial data generation, program-guided data generation | We develop an A*-powered algorithm for adversarially generating challenging and diverse theory of mind data, that can be effectively used as to stress-test LLMs capabilities or as fine-tuning data | 12,444 | 2412.12175 | [
-0.06426406651735306,
-0.06316953897476196,
0.0037054407875984907,
0.006818069145083427,
0.037294451147317886,
-0.023676060140132904,
-0.02811848744750023,
-0.05036051571369171,
0.0183979794383049,
-0.009234486147761345,
-0.07806037366390228,
-0.05336359515786171,
0.03875517472624779,
0.021118301898241043,
-0.009312598966062069,
-0.0019351060036569834,
0.06919221580028534,
0.006542421877384186,
-0.0451977513730526,
-0.024778451770544052,
0.06319982558488846,
0.05655571445822716,
0.06824138015508652,
-0.02852090448141098,
0.01041571144014597,
0.0017156064277514815,
0.004647810477763414,
-0.03835132718086243,
-0.021192410960793495,
0.038236841559410095,
0.03253817930817604,
0.0812254548072815,
-0.05036483332514763,
0.03557417169213295,
-0.038711633533239365,
0.06507369875907898,
-0.06757188588380814,
0.04537642002105713,
0.07819707691669464,
-0.07006967067718506,
-0.0695786103606224,
-0.04498898983001709,
0.06066730618476868,
-0.007388708181679249,
0.0662941187620163,
-0.05315079167485237,
-0.036444809287786484,
-0.017319194972515106,
-0.06130904704332352,
-0.006543681025505066,
-0.1398400217294693,
-0.011902451515197754,
0.00610363157466054,
0.013965793885290623,
-0.034696709364652634,
-0.009717200882732868,
0.06891871988773346,
0.02166592888534069,
-0.03919599950313568,
0.009552279487252235,
-0.042020440101623535,
-0.08417762815952301,
-0.04653104022145271,
-0.019896656274795532,
-0.0015139988390728831,
-0.039914846420288086,
-0.06595062464475632,
0.0790017694234848,
-0.06265084445476532,
-0.018367746844887733,
0.035083334892988205,
0.060826994478702545,
0.01267499290406704,
0.05433416739106178,
0.009122316725552082,
0.08992123603820801,
0.011204027570784092,
-0.058627303689718246,
0.06023719161748886,
-0.1003628745675087,
0.02954418957233429,
0.04734466224908829,
0.04690214991569519,
0.07828261703252792,
0.06407720595598221,
0.01458794716745615,
0.03507883474230766,
0.039997000247240067,
0.023459382355213165,
0.12144516408443451,
-0.02578911744058132,
-0.1162358969449997,
-0.025555945932865143,
0.03359859064221382,
0.11033827811479568,
0.08627285063266754,
-0.04029024392366409,
-0.07082508504390717,
-0.05071781948208809,
0.07158847153186798,
0.016212692484259605,
0.12646988034248352,
0.027472391724586487,
-0.06317869573831558,
0.03852235525846481,
-0.054534174501895905,
0.09405194222927094,
0.04182003065943718,
-0.0259349774569273,
-0.05185345187783241,
0.007586539722979069,
0.07695536315441132,
0.036890871822834015,
0.02156049944460392,
0.025437066331505775,
-0.060109272599220276,
0.0006814483786001801,
-0.004641873762011528,
0.025020863860845566,
0.014320294372737408,
0.008500751107931137,
0.061943065375089645,
0.034118566662073135,
0.028482096269726753,
0.0638490691781044,
-0.035320986062288284,
-0.07118484377861023,
6.079148932159117e-34,
0.09670626372098923,
-0.0055167702957987785,
0.01950792409479618,
0.06309286504983902,
0.07925236225128174,
0.0019305248279124498,
-0.027002327144145966,
0.011495602317154408,
-0.10570493340492249,
0.07283686846494675,
-0.10260434448719025,
-0.012385961599647999,
-0.07525216788053513,
0.05724533647298813,
0.01993384398519993,
0.054577041417360306,
-0.06507642567157745,
0.017041878774762154,
-0.0579686276614666,
-0.028546703979372978,
0.07139729708433151,
0.026619333773851395,
-0.008158917538821697,
-0.07341417670249939,
-0.021200072020292282,
0.08897816389799118,
0.060057640075683594,
-0.05824168771505356,
-0.05954297259449959,
0.02619365230202675,
-0.12827423214912415,
0.055832307785749435,
-0.016180602833628654,
0.07130572199821472,
0.03150496631860733,
-0.03179299458861351,
-0.04647386446595192,
-0.058613408356904984,
0.059869587421417236,
0.040496185421943665,
0.043502986431121826,
0.0841713547706604,
0.04865856468677521,
0.02324565127491951,
-0.04374310001730919,
0.009294804185628891,
0.025732919573783875,
0.01080362033098936,
-0.01979876309633255,
0.0065928976982831955,
0.025275468826293945,
0.04473438858985901,
-0.05077106133103371,
0.008738839998841286,
-0.08268959820270538,
-0.053578197956085205,
0.027334582060575485,
0.043433912098407745,
0.006051215808838606,
0.025200657546520233,
0.020995991304516792,
0.03390355035662651,
0.004305160138756037,
0.007916485890746117,
0.013105885125696659,
0.04305221512913704,
-0.08147620409727097,
0.0070170140825212,
0.055857159197330475,
-0.00901836808770895,
-0.00007198850653367117,
-0.0011938161915168166,
0.02167366072535515,
-0.058843277394771576,
0.03219833970069885,
-0.05684255063533783,
0.10344468057155609,
-0.026709189638495445,
-0.013001007959246635,
-0.016512185335159302,
0.08367586880922318,
-0.01593947410583496,
-0.05411151796579361,
-0.044337835162878036,
-0.047246482223272324,
-0.038774240761995316,
0.047277357429265976,
-0.04346025735139847,
-0.02935088612139225,
-0.05154090002179146,
-0.0527479387819767,
-0.03398571163415909,
-0.05238319933414459,
-0.0507151298224926,
-0.0745524913072586,
-7.967727543548975e-34,
-0.06466654688119888,
-0.06432121247053146,
0.01772758737206459,
0.045841775834560394,
-0.0484108105301857,
-0.11109911650419235,
-0.03012048825621605,
0.060429077595472336,
-0.042233120650053024,
-0.015019852668046951,
-0.07784011960029602,
-0.09015441685914993,
0.07311501353979111,
-0.02081286907196045,
0.041307393461465836,
-0.08594244718551636,
-0.030140839517116547,
-0.026177693158388138,
-0.007125441450625658,
0.045390404760837555,
0.02598070353269577,
0.13228026032447815,
-0.09907208383083344,
0.035345498472452164,
0.026316294446587563,
0.005926911253482103,
-0.0046178982593119144,
0.037323907017707825,
-0.010617326945066452,
-0.01948646828532219,
-0.008107036352157593,
-0.02523711696267128,
-0.008813637308776379,
0.01953108049929142,
-0.08509358763694763,
0.03840623423457146,
0.08906383067369461,
-0.025165053084492683,
-0.043403249233961105,
0.07205838710069656,
0.04941214621067047,
-0.015999112278223038,
-0.08607635647058487,
0.03921179473400116,
-0.05736996605992317,
0.030104869976639748,
-0.0709766373038292,
0.0173011664301157,
-0.006803707219660282,
0.043390903621912,
-0.026989620178937912,
-0.03739609569311142,
0.012980150058865547,
-0.020746072754263878,
0.025415195152163506,
-0.12145835161209106,
-0.01091268565505743,
-0.09543868899345398,
-0.003896439913660288,
0.009663349948823452,
-0.03518791124224663,
0.010164679028093815,
0.03165954351425171,
-0.018563367426395416,
-0.009695373475551605,
0.010344711132347584,
-0.02291673980653286,
0.0537710040807724,
0.005431422032415867,
0.014800104312598705,
0.005605453159660101,
-0.025453118607401848,
-0.03154584392905235,
-0.00036738693597726524,
0.059282295405864716,
0.030733631923794746,
-0.03718429431319237,
-0.061330366879701614,
-0.01800883375108242,
-0.06049581244587898,
0.031853582710027695,
-0.0758967399597168,
0.058338992297649384,
0.026379525661468506,
0.1134674921631813,
0.05140905827283859,
0.0856754332780838,
0.11265759915113449,
-0.04194283112883568,
0.039378270506858826,
-0.06062202528119087,
0.03667299076914787,
-0.02213330566883087,
0.055778082460165024,
-0.0495627261698246,
-4.621292504225494e-8,
-0.10262762010097504,
-0.02504001185297966,
-0.03050164133310318,
0.11506539583206177,
0.037205420434474945,
-0.013583880849182606,
-0.0619451068341732,
0.09525103121995926,
-0.008360529318451881,
0.020678114145994186,
0.05508524179458618,
-0.014361896552145481,
-0.08733248710632324,
-0.007400958798825741,
0.03789540380239487,
-0.00871442724019289,
0.016257429495453835,
0.014773214235901833,
-0.017213569954037666,
0.03096301667392254,
0.07957597821950912,
0.0897306501865387,
0.0012277164496481419,
-0.04760416969656944,
0.03158384934067726,
-0.0339084155857563,
-0.0663064569234848,
0.051053617149591446,
-0.01640208624303341,
-0.0318957082927227,
-0.004592409357428551,
-0.008281673304736614,
-0.017046652734279633,
-0.004378108307719231,
0.023135807365179062,
0.047912515699863434,
-0.0359327532351017,
-0.03426460921764374,
0.033832136541604996,
-0.040052223950624466,
0.08791689574718475,
0.049186162650585175,
-0.00352094485424459,
0.040917154401540756,
0.015816878527402878,
-0.01995813474059105,
-0.05311306565999985,
-0.09816575795412064,
0.01959136314690113,
-0.018094507977366447,
0.02514643222093582,
0.029906749725341797,
0.025125697255134583,
0.04881412535905838,
0.09422153234481812,
0.024926472455263138,
-0.04163814336061478,
0.03281356766819954,
-0.0026546488516032696,
0.08341005444526672,
0.04593373462557793,
0.06500116735696793,
-0.08316022157669067,
-0.020744698122143745
] | 0 | 0 | 0 | 0 |
|
Controllable Context Sensitivity and the Knob Behind It | https://openreview.net/forum?id=Igm9bbkzHC | [
"Julian Minder",
"Kevin Du",
"Niklas Stoehr",
"Giovanni Monea",
"Chris Wendler",
"Robert West",
"Ryan Cotterell"
] | Poster | When making predictions, a language model must trade off how much it relies on its context vs. its prior knowledge.
Choosing how sensitive the model is to its context is a fundamental functionality, as it enables the model to excel at tasks like retrieval-augmented generation and question-answering.
In this paper, we search for a knob which controls this sensitivity, determining whether language models answer from the context or their prior knowledge.
To guide this search, we design a task for controllable context sensitivity.
In this task, we first feed the model a context ("Paris is in England") and a question ("Where is Paris?"); we then instruct the model to either use its prior or contextual knowledge and evaluate whether it generates the correct answer for both intents (either "France" or "England").
When fine-tuned on this task, instruct versions of Llama-3.1, Mistral-v0.3, and Gemma-2 can solve it with high accuracy (85-95%).
Analyzing these high-performing models, we narrow down which layers may be important to context sensitivity using a novel linear time algorithm.
Then, in each model, we identify a 1-D subspace in a single layer that encodes whether the model follows context or prior knowledge.
Interestingly, while we identify this subspace in a fine-tuned model, we find that the exact same subspace serves as an effective knob in not only that model but also non-fine-tuned instruct and base models of that model family.
Finally, we show a strong correlation between a model's performance and how distinctly it separates context-agreeing from context-ignoring answers in this subspace.
These results suggest a single fundamental subspace facilitates how the model chooses between context and prior knowledge. | analysis, interpretability, mechanistic interpretability, context vs prior knowledge, large language models | The tension of choosing between in-context information and prior knowledge when prompted is fundamental to LMs; we use mechanistic interpretability techniques to find a knob which controls this. | 12,441 | 2411.07404 | [
0.030223354697227478,
-0.038519684225320816,
-0.02110285498201847,
0.01621180586516857,
0.022466517984867096,
-0.006081909406930208,
0.06170080602169037,
0.014593630097806454,
0.013588129542768002,
-0.0593874417245388,
-0.0626053512096405,
-0.12024043500423431,
0.09152284264564514,
0.030249521136283875,
-0.01227414608001709,
-0.027726788073778152,
0.14111949503421783,
0.032463397830724716,
-0.08335159718990326,
-0.02206786535680294,
0.05720151960849762,
0.019713152199983597,
0.056742917746305466,
-0.0159866102039814,
-0.0020330934785306454,
-0.00005886836879653856,
-0.008752943016588688,
-0.018313895910978317,
0.016683975234627724,
0.013021908700466156,
0.022271690890192986,
0.08745276182889938,
-0.0032930641900748014,
0.04896968975663185,
-0.09346074610948563,
0.038378577679395676,
-0.07544289529323578,
0.030413709580898285,
-0.006487295497208834,
-0.052556753158569336,
-0.07757455110549927,
-0.02140001393854618,
0.05352034047245979,
0.022805865854024887,
0.07449328899383545,
0.008308560587465763,
0.01887342520058155,
0.07878787815570831,
-0.02428305707871914,
0.005887577775865793,
-0.08922198414802551,
-0.038214173167943954,
-0.010392896831035614,
-0.07562265545129776,
-0.06346406787633896,
0.04084901139140129,
0.02090543881058693,
0.08126381784677505,
-0.021928846836090088,
-0.0010420774342492223,
-0.0762263685464859,
-0.09678039699792862,
-0.1458703577518463,
0.035534318536520004,
0.021326730027794838,
0.014014114625751972,
-0.025699736550450325,
-0.01377180777490139,
-0.026775212958455086,
-0.006248857825994492,
-0.019418852403759956,
0.08361835777759552,
0.04427649825811386,
0.042006753385066986,
0.03494454175233841,
0.0028647540602833033,
-0.0069297924637794495,
-0.025605522096157074,
0.03548521548509598,
-0.058421794325113297,
0.10651988536119461,
0.0032566238660365343,
0.0346093475818634,
0.00012103949120501056,
0.11979008466005325,
-0.0326240211725235,
0.06970338523387909,
0.056638818234205246,
0.06122775003314018,
-0.017440715804696083,
-0.05102979391813278,
-0.13316909968852997,
0.05866223946213722,
0.011376168578863144,
0.039987266063690186,
0.07621318846940994,
0.0665699914097786,
-0.014965418726205826,
-0.05119602382183075,
0.08812036365270615,
0.044980213046073914,
0.032793525606393814,
-0.05734390392899513,
-0.04777819663286209,
0.015238660387694836,
0.0015363912098109722,
-0.05254707485437393,
0.024764930829405785,
-0.010386427864432335,
-0.09869474917650223,
-0.0054929545149207115,
0.028723660856485367,
0.008313138037919998,
-0.026786942034959793,
-0.02044566720724106,
-0.04204888269305229,
0.11307722330093384,
-0.053607966750860214,
0.009206495247781277,
0.017510296776890755,
-0.019719241186976433,
0.03715697303414345,
0.04726998880505562,
0.016002999618649483,
-0.036047983914613724,
-0.016318634152412415,
-0.031020665541291237,
-4.2238124744910485e-34,
0.03364234045147896,
0.02144073136150837,
0.024838652461767197,
0.07827890664339066,
0.0100662000477314,
0.026576800271868706,
-0.044541217386722565,
0.0260004885494709,
-0.0021362656261771917,
-0.02394036389887333,
0.05659830942749977,
-0.007410351186990738,
-0.06957690417766571,
0.028955448418855667,
0.0909930020570755,
0.05190913751721382,
0.022917063906788826,
0.043926410377025604,
-0.05845984071493149,
-0.037413597106933594,
0.027982506901025772,
-0.0075459349900484085,
0.017521781846880913,
-0.05533862113952637,
0.022929521277546883,
0.011479456909000874,
0.0737089291214943,
-0.0360734760761261,
-0.05147762596607208,
-0.01095578633248806,
-0.12396315485239029,
-0.043942056596279144,
-0.023499615490436554,
0.04084361344575882,
0.010386491194367409,
-0.012397188693284988,
-0.022032247856259346,
-0.05291373282670975,
0.027131903916597366,
-0.04028085246682167,
-0.011870237998664379,
0.04814602807164192,
0.06314907222986221,
0.044021982699632645,
-0.07042773813009262,
-0.04509807750582695,
-0.052831243723630905,
-0.015465945936739445,
-0.0012295570923015475,
0.022288115695118904,
0.06909015029668808,
0.020110100507736206,
-0.01187338400632143,
-0.05880386382341385,
0.023645464330911636,
0.024955052882432938,
-0.01563153974711895,
0.03316400200128555,
0.030631404370069504,
0.09208060055971146,
0.039661917835474014,
-0.025044750422239304,
0.042056918144226074,
0.09132269024848938,
0.08028365671634674,
0.061401862651109695,
-0.03674778342247009,
0.05652070418000221,
0.07341761142015457,
-0.0519939623773098,
-0.08496348559856415,
-0.05031831935048103,
0.07421048730611801,
-0.028207596391439438,
0.05547504127025604,
0.04293888062238693,
0.013180646114051342,
-0.03361944481730461,
0.009255991317331791,
-0.02742970734834671,
0.03457866981625557,
-0.01909126713871956,
-0.00987138319760561,
0.021109141409397125,
-0.07650554180145264,
-0.04902484267950058,
0.04209262505173683,
-0.04565439373254776,
-0.038863200694322586,
-0.039323706179857254,
-0.061072614043951035,
-0.026040999218821526,
-0.009770959615707397,
-0.004773743916302919,
0.013868476264178753,
4.53399053618451e-34,
-0.00855456292629242,
-0.0876927524805069,
-0.02358279749751091,
0.014463982544839382,
-0.037466131150722504,
-0.1036541536450386,
-0.01756282150745392,
-0.006723463535308838,
-0.0598704032599926,
-0.03889177367091179,
-0.06252913177013397,
-0.05930119380354881,
0.0701640173792839,
0.011207430623471737,
0.02328873798251152,
0.031173313036561012,
0.013009516522288322,
-0.04788986220955849,
-0.015123472549021244,
0.08011401444673538,
-0.04669856280088425,
0.03230840340256691,
-0.18065239489078522,
0.032822296023368835,
-0.06828373670578003,
-0.018740905448794365,
-0.09824206680059433,
0.002977149561047554,
-0.041152168065309525,
0.008046300150454044,
-0.11290254443883896,
0.011668330058455467,
0.030983012169599533,
-0.020238297060132027,
-0.09973926842212677,
0.0623522587120533,
0.015964888036251068,
-0.07108846306800842,
-0.02891712263226509,
0.11146049946546555,
0.020103441551327705,
0.005642957519739866,
-0.07327017933130264,
-0.02339240349829197,
-0.025933563709259033,
0.04878925532102585,
-0.11409120261669159,
-0.031127169728279114,
0.018953220918774605,
-0.00571600953117013,
0.032542917877435684,
-0.02188286930322647,
-0.060314979404211044,
-0.014425629749894142,
-0.012278812006115913,
-0.04991956055164337,
-0.032700926065444946,
-0.07945001125335693,
0.09048212319612503,
-0.017663350328803062,
-0.016039716079831123,
-0.0192315261811018,
0.004395582247525454,
-0.026161300018429756,
0.0591781809926033,
0.002043761545792222,
-0.056740403175354004,
0.007122655864804983,
0.028824834153056145,
-0.013929667882621288,
0.07237821072340012,
-0.010833779349923134,
0.030847935006022453,
-0.03651563450694084,
0.06708125025033951,
0.006166456267237663,
0.03543556481599808,
-0.05286955088376999,
-0.013028321787714958,
-0.05695130676031113,
0.0007722560549154878,
-0.01725853979587555,
0.019126301631331444,
0.05186910554766655,
-0.005728615447878838,
0.09293216466903687,
0.021052569150924683,
0.06290994584560394,
0.05617900565266609,
0.025701504200696945,
0.01434130035340786,
0.04041614010930061,
-0.05375798046588898,
0.04849229007959366,
-0.04036615416407585,
-5.2708550413171906e-8,
-0.07024122774600983,
0.059488214552402496,
-0.042487259954214096,
0.04176986217498779,
0.022061971947550774,
0.014305531047284603,
-0.0072976332157850266,
0.055046986788511276,
0.008455782197415829,
-0.03243235871195793,
0.01778595894575119,
0.015928665176033974,
-0.029712023213505745,
-0.06195421889424324,
-0.013978231698274612,
0.06268582493066788,
0.04846269637346268,
0.022706547752022743,
-0.03945336118340492,
0.0159876961261034,
0.15071694552898407,
0.062382135540246964,
0.009365436621010303,
0.03535924851894379,
0.07554657012224197,
-0.02919476293027401,
-0.04105618596076965,
0.1497477889060974,
0.09190214425325394,
0.03320609778165817,
0.03768719732761383,
-0.04970172047615051,
-0.08653077483177185,
0.01283340249210596,
-0.026099547743797302,
0.07481791824102402,
-0.012433678843080997,
-0.08101356029510498,
0.02649584785103798,
-0.014741245657205582,
0.10425565391778946,
0.02686714008450508,
-0.09875179082155228,
0.06847147643566132,
0.05533616989850998,
0.023523077368736267,
0.05337829887866974,
-0.10111586004495621,
0.02051205188035965,
-0.019279373809695244,
-0.06039039045572281,
-0.017162414267659187,
0.0190932285040617,
0.06802918761968613,
0.05999980494379997,
-0.029743865132331848,
-0.04943859577178955,
-0.004022904671728611,
0.03734312579035759,
0.04047580808401108,
-0.0032695697154849768,
0.06927929073572159,
-0.05056814104318619,
0.03232404589653015
] | https://github.com/kdu4108/context-vs-prior-finetuning | 9 | 0 | 0 | 0 |
The 3D-PC: a benchmark for visual perspective taking in humans and machines | https://openreview.net/forum?id=UIFAJZ22ZF | [
"Drew Linsley",
"Peisen Zhou",
"Alekh Karkada Ashok",
"Akash Nagaraj",
"Gaurav Gaonkar",
"Francis E Lewis",
"Zygmunt Pizlo",
"Thomas Serre"
] | Poster | Visual perspective taking (VPT) is the ability to perceive and reason about the perspectives of others. It is an essential feature of human intelligence, which develops over the first decade of life and requires an ability to process the 3D structure of visual scenes. A growing number of reports have indicated that deep neural networks (DNNs) become capable of analyzing 3D scenes after training on large image datasets. We investigated if this emergent ability for 3D analysis in DNNs is sufficient for VPT with the 3D perception challenge (3D-PC): a novel benchmark for 3D perception in humans and DNNs. The 3D-PC is comprised of three 3D-analysis tasks posed within natural scene images: (i.) a simple test of object depth order, (ii.) a basic VPT task (VPT-basic), and (iii.) a more challenging version of VPT (VPT-perturb) designed to limit the effectiveness of "shortcut" visual strategies. We tested human participants (N=33) and linearly probed or text-prompted over 300 DNNs on the challenge and found that nearly all of the DNNs approached or exceeded human accuracy in analyzing object depth order. Surprisingly, DNN accuracy on this task correlated with their object recognition performance. In contrast, there was an extraordinary gap between DNNs and humans on VPT-basic. Humans were nearly perfect, whereas most DNNs were near chance. Fine-tuning DNNs on VPT-basic brought them close to human performance, but they, unlike humans, dropped back to chance when tested on VPT-perturb. Our challenge demonstrates that the training routines and architectures of today's DNNs are well-suited for learning basic 3D properties of scenes and objects but are ill-suited for reasoning about these properties like humans do. We release our 3D-PC datasets and code to help bridge this gap in 3D perception between humans and machines. | 3D vision, visual cognition, developmental psychology, visual reasonsing | Humans have a unique ability to view the world from another's perspective, which is unmatched by deep neural networks. | 12,436 | null | [
-0.08868162333965302,
-0.06080646067857742,
0.0032206696923822165,
-0.004981941543519497,
0.06546271592378616,
0.014462688006460667,
-0.04928065463900566,
-0.034003838896751404,
0.016734963282942772,
-0.020258378237485886,
-0.10269269347190857,
-0.10366976261138916,
-0.05353621393442154,
0.10567696392536163,
-0.057298582047224045,
-0.02437608316540718,
0.10137565433979034,
0.11110678315162659,
-0.08537366986274719,
0.0890425443649292,
-0.004024075344204903,
-0.02225635200738907,
0.057064928114414215,
-0.008001468144357204,
0.04729713872075081,
-0.007093637250363827,
0.04160330817103386,
-0.06361118704080582,
-0.011958981864154339,
-0.0014040839159861207,
0.01760960929095745,
-0.04236993193626404,
0.0008241915493272245,
0.054571833461523056,
-0.0065963659435510635,
0.02687142603099346,
-0.012413503602147102,
0.012399626895785332,
0.0030512833036482334,
-0.01414522249251604,
-0.01715867593884468,
0.07694835215806961,
0.030516773462295532,
0.02442527376115322,
0.13506877422332764,
0.03626685589551926,
-0.008338134735822678,
-0.10054916888475418,
0.028568964451551437,
-0.03265317901968956,
-0.059299666434526443,
-0.042466141283512115,
0.003983621019870043,
0.010713279247283936,
-0.06443207710981369,
0.10905233770608902,
0.03271876275539398,
0.017974790185689926,
0.019686292856931686,
0.00717671075835824,
0.05302439630031586,
-0.04678523913025856,
0.011130385100841522,
-0.012023461982607841,
0.0038356203585863113,
0.008895760402083397,
-0.018277272582054138,
-0.04711219668388367,
-0.02072271518409252,
0.014350509271025658,
-0.008598599582910538,
0.09343047440052032,
-0.043995585292577744,
-0.029354818165302277,
0.003008487168699503,
0.0178249329328537,
0.022727595642209053,
-0.0991397276520729,
-0.025865986943244934,
-0.0886697918176651,
0.0989253968000412,
0.02051590196788311,
0.028278503566980362,
0.05577535182237625,
0.04027237370610237,
-0.0070507037453353405,
-0.09557319432497025,
0.07572249323129654,
-0.09511157125234604,
0.032451894134283066,
-0.051085904240608215,
-0.1078682392835617,
-0.05259818956255913,
-0.027901390567421913,
0.05291090905666351,
-0.06004638597369194,
0.04822393134236336,
-0.04478982463479042,
0.006104466039687395,
0.0426752045750618,
0.008850953541696072,
0.032149165868759155,
0.018218008801341057,
-0.016197791323065758,
0.04817439243197441,
0.032579876482486725,
0.09068821370601654,
-0.05534878373146057,
0.022727787494659424,
-0.028312988579273224,
0.0012277596397325397,
-0.0005390268051996827,
0.041169289499521255,
0.0121223833411932,
-0.009354489855468273,
-0.0364474393427372,
-0.06633094698190689,
0.014021425507962704,
0.07295216619968414,
0.008359766565263271,
-0.03656107932329178,
-0.0049740285612642765,
0.002261177636682987,
0.01405241433531046,
0.04213807359337807,
0.023890864104032516,
0.004795357119292021,
1.3852535117093513e-33,
-0.06898675858974457,
0.04730949550867081,
-0.002614989411085844,
0.020151296630501747,
0.039294663816690445,
-0.00517739774659276,
-0.004096232820302248,
0.034370459616184235,
-0.0381203293800354,
0.0346413180232048,
-0.04476140812039375,
-0.05271890014410019,
-0.01799996756017208,
0.11645855754613876,
-0.05193960294127464,
0.015577846206724644,
-0.017082715407013893,
0.03883570060133934,
-0.104449562728405,
0.049455951899290085,
0.027021003887057304,
-0.04804566130042076,
0.01360889058560133,
0.002253742655739188,
0.0039353021420538425,
-0.07000136375427246,
0.032136332243680954,
0.011721408925950527,
-0.03453611582517624,
-0.02919064834713936,
-0.03871527686715126,
0.037576574832201004,
-0.02378833293914795,
0.05532826483249664,
0.011896890588104725,
0.03266715630888939,
0.03464163839817047,
-0.02483713999390602,
0.03676270693540573,
0.03338973596692085,
-0.01895087957382202,
0.09416592866182327,
-0.010022396221756935,
-0.040657784789800644,
-0.08060995489358902,
0.019867563620209694,
0.051599059253931046,
0.08731290698051453,
-0.08814731985330582,
0.017100563272833824,
0.04702748730778694,
0.024858178570866585,
-0.029477201402187347,
-0.07709663361310959,
-0.002574353711679578,
-0.038435760885477066,
0.04646628350019455,
-0.0038337726145982742,
0.07707784324884415,
0.027415800839662552,
-0.008099730126559734,
0.035278089344501495,
-0.030531276017427444,
0.03983831778168678,
-0.03162140026688576,
0.024557597935199738,
-0.056270696222782135,
0.1340421438217163,
0.07332997024059296,
0.031055744737386703,
-0.04592219367623329,
-0.017116760835051537,
-0.01012615580111742,
-0.09026829153299332,
0.1145898699760437,
0.0399002805352211,
0.015182795003056526,
-0.033815648406744,
0.033969804644584656,
0.03744661062955856,
-0.020947443321347237,
0.031463902443647385,
0.023039212450385094,
-0.062230270355939865,
-0.05578571557998657,
0.0008332653669640422,
-0.04350128397345543,
-0.1520036906003952,
0.05551118031144142,
-0.042355976998806,
-0.05992551147937775,
-0.09767419099807739,
-0.0333666056394577,
-0.03337899595499039,
0.03407996520400047,
-3.329364818037206e-34,
-0.0014853328466415405,
0.03432348370552063,
-0.08948604017496109,
0.015952816233038902,
0.03286196291446686,
-0.052457813173532486,
-0.044681254774332047,
-0.0035887740086764097,
-0.01752273365855217,
-0.025123542174696922,
0.04222172498703003,
-0.006384511943906546,
0.02134445309638977,
-0.009208178147673607,
0.02260982058942318,
-0.010694058611989021,
-0.058004483580589294,
-0.034158822149038315,
0.005104857496917248,
0.0008407323621213436,
0.04504433274269104,
0.03634045273065567,
-0.0894230529665947,
-0.0030466497410088778,
-0.1318274289369583,
0.09430702775716782,
-0.019518034532666206,
-0.03127114847302437,
0.02736586518585682,
-0.02393527515232563,
0.0005515645607374609,
-0.0027577723376452923,
-0.056870236992836,
-0.018390381708741188,
0.02852904610335827,
0.07740457355976105,
0.030053989961743355,
-0.014864375814795494,
-0.05544765666127205,
0.02726498618721962,
0.05548211559653282,
-0.00470969220623374,
0.01760665513575077,
0.05315149575471878,
-0.041093114763498306,
0.06020103767514229,
0.02339559979736805,
-0.07747326046228409,
0.009559492580592632,
0.1280388981103897,
-0.05617079883813858,
0.05436030402779579,
-0.004351559095084667,
0.02894335240125656,
0.020864199846982956,
-0.06902673095464706,
-0.042496729642152786,
0.019604234024882317,
0.12052231281995773,
0.03657642379403114,
0.002670504618436098,
-0.023928051814436913,
-0.05783366039395332,
0.05969305336475372,
-0.023521307855844498,
0.039819344878196716,
-0.09571947157382965,
0.05589793995022774,
0.08744940161705017,
0.019226280972361565,
-0.029272252693772316,
-0.018262304365634918,
0.02639349177479744,
-0.017674963921308517,
-0.05561276152729988,
-0.068116694688797,
-0.018323950469493866,
0.01951741799712181,
0.042362142354249954,
0.007418482098728418,
0.0819624587893486,
-0.10868504643440247,
0.004335299599915743,
0.09971777349710464,
0.0820525735616684,
0.03571370244026184,
0.01647176221013069,
0.019728034734725952,
-0.025164199993014336,
0.0354144461452961,
-0.03170699626207352,
-0.026312556117773056,
-0.03166647255420685,
0.024086279794573784,
0.014569545164704323,
-4.261710984110323e-8,
-0.05220454931259155,
0.045179352164268494,
0.06748891621828079,
-0.010269819758832455,
-0.013111834414303303,
-0.05872870236635208,
-0.029827937483787537,
0.16910092532634735,
-0.08511855453252792,
0.027402712032198906,
0.0738757848739624,
-0.04442436620593071,
-0.03554520383477211,
-0.037936173379421234,
0.055807244032621384,
0.041080866008996964,
0.013366591185331345,
0.005681685172021389,
0.01609855517745018,
0.02113690786063671,
0.004949395079165697,
0.06717047095298767,
0.01232884731143713,
-0.012940797954797745,
0.019596414640545845,
-0.010458976961672306,
-0.06628139317035675,
0.12119122594594955,
0.05581078678369522,
-0.040465906262397766,
0.04948794096708298,
0.05978626757860184,
-0.0024141953326761723,
-0.05399168282747269,
-0.0037932666018605232,
0.02457522228360176,
-0.04079410433769226,
-0.006679181009531021,
0.061305154114961624,
-0.06340570002794266,
-0.056449197232723236,
0.08194121718406677,
-0.011008425615727901,
0.06368662416934967,
-0.019040275365114212,
-0.03874550387263298,
-0.01189263816922903,
-0.04854075238108635,
-0.0021236364264041185,
-0.03888707607984543,
-0.02726050652563572,
-0.07122176140546799,
-0.010076232254505157,
0.15578536689281464,
0.045223966240882874,
0.0678967610001564,
-0.013011296279728413,
-0.05944488197565079,
-0.031904540956020355,
0.12705160677433014,
-0.02570878528058529,
0.0010636398801580071,
-0.07911857217550278,
-0.03143762797117233
] | 0 | 0 | 0 | 0 |
|
HelpSteer2-Preference: Complementing Ratings with Preferences | https://openreview.net/forum?id=MnfHxPP5gs | [
"Zhilin Wang",
"Alexander Bukharin",
"Olivier Delalleau",
"Daniel Egert",
"Gerald Shen",
"Jiaqi Zeng",
"Oleksii Kuchaiev",
"Yi Dong"
] | Poster | Reward models are critical for aligning models to follow instructions, and are typically trained following one of two popular paradigms: Bradley-Terry style or Regression style. However, there is a lack of evidence that either approach is better than the other, when adequately matched for data. This is primarily because these approaches require data collected in different (but incompatible) formats, meaning that adequately matched data is not available in existing public datasets. To tackle this problem, we release preference annotations (designed for Bradley-Terry training) to complement existing ratings (designed for Regression style training) in the HelpSteer2 dataset. To improve data interpretability, preference annotations are accompanied with human-written justifications. Using this data, we conduct the first head-to-head comparison of Bradley-Terry and Regression models when adequately matched for data. Based on insights derived from such a comparison, we propose a novel approach to combine Bradley-Terry and Regression reward modeling. A Llama-3.1-70B-Instruct model tuned with this approach scores 94.1 on RewardBench, emerging top of more than 140 reward models as of 1 Oct 2024. This reward model can then be used with REINFORCE algorithm (RLHF) to align an Instruct model to reach 85.0 on Arena Hard, which is No. 1 as of 1 Oct 2024.
We open-source this dataset (CC-BY-4.0 license) at https://huggingface.co/datasets/nvidia/HelpSteer2#preferences-new---1-oct-2024 and openly release the trained Reward and Instruct models at https://huggingface.co/nvidia/Llama-3.1-Nemotron-70B-Reward and https://huggingface.co/nvidia/Llama-3.1-Nemotron-70B-Instruct . | reward modelling, rlhf, model alignment | We release data to support a data-matched comparison of Bradley-Terry and Regression style Reward Models, which inspires an approach resulting in a Reward Model scoring 94.1% on RewardBench, emerging the top model out of 140+ models as of 1 Oct 2024. | 12,434 | 2410.01257 | [
-0.12071898579597473,
-0.029539136216044426,
-0.01032694149762392,
0.00388162094168365,
0.021768055856227875,
0.08301714062690735,
-0.012788215652108192,
0.005003790836781263,
0.0024645200464874506,
-0.01296705286949873,
-0.041577886790037155,
-0.07471595704555511,
0.10918106883764267,
0.039045728743076324,
0.06626582890748978,
0.041005153208971024,
0.09788525104522705,
-0.02569306083023548,
-0.04861835762858391,
-0.008942870423197746,
-0.017296195030212402,
0.031224222853779793,
0.024499056860804558,
-0.002239840803667903,
-0.041402578353881836,
-0.00742416363209486,
0.003087269840762019,
0.06675378233194351,
-0.04413600265979767,
-0.012534198351204395,
-0.036397967487573624,
-0.027085809037089348,
0.0007438096799887717,
-0.015620887279510498,
-0.13170763850212097,
-0.001004871679469943,
-0.05747196078300476,
0.03404989838600159,
-0.0798981785774231,
-0.04683275893330574,
-0.07543100416660309,
-0.034502796828746796,
0.014107157476246357,
0.01920764334499836,
0.0666489452123642,
-0.02742352895438671,
-0.006325811147689819,
-0.057635121047496796,
-0.025335535407066345,
0.029681609943509102,
-0.06021106243133545,
-0.0763721764087677,
0.029942769557237625,
-0.001751260831952095,
0.0090940585359931,
0.04932643845677376,
-0.032889168709516525,
0.07170841097831726,
0.03464644402265549,
-0.05314697325229645,
-0.034930869936943054,
-0.07930973172187805,
-0.07980132102966309,
0.03339170292019844,
0.019775619730353355,
-0.08002214878797531,
-0.05003074184060097,
0.10075779259204865,
-0.08925087004899979,
-0.04123667627573013,
-0.009853384457528591,
0.05696730315685272,
0.03467785194516182,
0.028365114703774452,
0.04645409435033798,
0.0696525126695633,
0.03834451362490654,
0.008622931316494942,
0.06341376155614853,
-0.042099278420209885,
-0.1115008071064949,
-0.016763979569077492,
0.044541217386722565,
0.034249674528837204,
0.03620113059878349,
-0.03679898753762245,
0.05572601035237312,
0.013014416210353374,
-0.004343664273619652,
0.07428869605064392,
-0.06178949400782585,
-0.020337142050266266,
0.03944868594408035,
-0.013728714548051357,
-0.0407540500164032,
0.09408841282129288,
-0.06680610030889511,
-0.052092161029577255,
-0.04460444301366806,
0.07646135985851288,
0.028085727244615555,
0.11193570494651794,
-0.022449782118201256,
-0.05157216638326645,
0.09600552916526794,
-0.05768635496497154,
0.05188987776637077,
0.03744488209486008,
0.0664091482758522,
-0.09987682104110718,
0.021130355075001717,
0.06733856350183487,
-0.047869723290205,
-0.014364244416356087,
0.018060199916362762,
-0.027082646265625954,
-0.016751503571867943,
0.016930365934967995,
0.04487936571240425,
0.01890575885772705,
0.02155124768614769,
-0.018233967944979668,
0.036054015159606934,
0.008752919733524323,
0.011495140381157398,
0.033051129430532455,
-0.0710122138261795,
4.375533438516185e-33,
0.03509564325213432,
0.03416464850306511,
0.03877492621541023,
0.02006860077381134,
0.07262230664491653,
-0.04515840858221054,
-0.09893225133419037,
0.07237870991230011,
-0.08161577582359314,
-0.008210488595068455,
-0.023065906018018723,
0.04486607387661934,
-0.04642977565526962,
0.13723166286945343,
0.08261281251907349,
0.027900340035557747,
-0.040026530623435974,
0.015544390305876732,
-0.06188325583934784,
-0.004360855557024479,
0.06252223998308182,
0.03420647606253624,
0.0378548838198185,
-0.04224123805761337,
0.025678670033812523,
-0.04565401002764702,
0.08645869791507721,
-0.035137757658958435,
-0.06815998256206512,
-0.0009607016108930111,
-0.02493196353316307,
-0.06247681751847267,
0.08512051403522491,
-0.041401226073503494,
-0.016355372965335846,
-0.01850143074989319,
-0.007534862961620092,
-0.0530812032520771,
0.034642819315195084,
-0.02017175406217575,
0.04167317971587181,
0.028067318722605705,
-0.01079777255654335,
0.040956467390060425,
-0.03031015209853649,
0.005706639029085636,
0.06955792009830475,
-0.0017248047515749931,
-0.012131909839808941,
0.02438708022236824,
0.04161371663212776,
0.01987091265618801,
0.0065570794977247715,
-0.010965700261294842,
-0.16739019751548767,
0.04110301285982132,
0.0029582511633634567,
0.0031465364154428244,
0.007661996874958277,
-0.08842156082391739,
0.014499669894576073,
0.03924423083662987,
0.06287555396556854,
-0.07241256535053253,
-0.04114223271608353,
0.010711063630878925,
0.023072533309459686,
0.010963700711727142,
-0.009380396455526352,
-0.027462812140583992,
-0.0635279193520546,
0.06069984659552574,
-0.020048806443810463,
-0.028918204829096794,
0.050856854766607285,
-0.012664076872169971,
0.04942713677883148,
-0.09792426973581314,
0.1039021760225296,
-0.0846983790397644,
0.07047760486602783,
0.00890396162867546,
0.013364708051085472,
-0.08023142069578171,
-0.03150707855820656,
-0.05760699883103371,
0.03813041001558304,
-0.09733697026968002,
0.021894115954637527,
0.0106422770768404,
-0.007182248402386904,
0.015051811933517456,
-0.048382461071014404,
0.012057199142873287,
0.028800535947084427,
-4.8571959221854636e-33,
-0.044813372194767,
-0.04113393276929855,
0.004429022315889597,
0.04698828607797623,
-0.0028633710462599993,
-0.04384306073188782,
-0.03319883346557617,
-0.021139226853847504,
0.047756414860486984,
-0.05550840124487877,
0.024229466915130615,
-0.006378769874572754,
0.03412104770541191,
-0.03683821111917496,
0.016868798062205315,
-0.020178748294711113,
-0.09514705836772919,
0.037485212087631226,
-0.018245374783873558,
-0.03295276314020157,
0.013152962550520897,
0.10123054683208466,
-0.057020001113414764,
0.025376377627253532,
-0.058355897665023804,
0.03332405909895897,
0.005320913624018431,
0.02736116386950016,
-0.006834443658590317,
0.023728346452116966,
-0.026160944253206253,
-0.06516571342945099,
-0.03319907933473587,
-0.06737979501485825,
-0.04734575003385544,
0.03543698415160179,
-0.009656036272644997,
-0.09275555610656738,
-0.04772505536675453,
0.13795028626918793,
0.0034928626846522093,
0.004209463484585285,
-0.09008584171533585,
0.03140096738934517,
-0.03592870011925697,
-0.02158055454492569,
-0.020068790763616562,
-0.04270944744348526,
0.023628195747733116,
-0.018021106719970703,
-0.026784364134073257,
0.017979640513658524,
-0.06938385218381882,
0.08012706786394119,
-0.048966649919748306,
-0.12213075906038284,
-0.02690809778869152,
-0.038849640637636185,
0.025120146572589874,
0.01448425929993391,
-0.045473288744688034,
0.03427066653966904,
0.014779884368181229,
0.010872483253479004,
-0.006783970165997744,
-0.06743764132261276,
0.053785040974617004,
0.04224548488855362,
-0.04725213348865509,
0.007625130005180836,
0.025704871863126755,
-0.035068392753601074,
0.03447506204247475,
-0.03230482339859009,
-0.01438729465007782,
0.015226291492581367,
0.029099203646183014,
-0.020069915801286697,
-0.05393566936254501,
-0.009843027219176292,
-0.038005080074071884,
-0.015396754257380962,
0.04788952320814133,
0.016079848632216454,
0.053390022367239,
0.06374891102313995,
0.11670918762683868,
0.11003626883029938,
-0.06631425768136978,
-0.014353876933455467,
-0.03016345016658306,
0.045191578567028046,
0.037754543125629425,
0.08847641199827194,
-0.03787437081336975,
-4.955789378868758e-8,
-0.029022008180618286,
0.01731407828629017,
-0.001684995018877089,
0.13568207621574402,
-0.060661494731903076,
0.0030321090016514063,
-0.06799882650375366,
0.06130869686603546,
-0.04201953858137131,
0.007604977115988731,
0.08466809242963791,
0.09267818182706833,
-0.09101264923810959,
-0.06037682294845581,
-0.021187789738178253,
0.03889741376042366,
0.04391336441040039,
0.04303458705544472,
-0.03507983684539795,
-0.004948415327817202,
-0.002196640707552433,
0.06342534720897675,
0.0373404286801815,
-0.09073087573051453,
0.0795595571398735,
-0.02731654979288578,
-0.10038405656814575,
0.12355467677116394,
0.042537350207567215,
0.057030241936445236,
-0.006109909620136023,
0.0046981703490018845,
-0.006918823812156916,
-0.08251456916332245,
0.06461954861879349,
0.028513966128230095,
0.04236520454287529,
-0.005156399682164192,
0.06993116438388824,
0.032171767204999924,
0.09595146775245667,
0.009597689844667912,
-0.04799901321530342,
0.05435480922460556,
0.031069861724972725,
0.009884214960038662,
0.051076579838991165,
-0.0725174993276596,
-0.02815478853881359,
-0.05500417575240135,
0.036743130534887314,
-0.07209667563438416,
-0.028554577380418777,
0.04443566873669624,
0.07054420560598373,
0.0422317199409008,
-0.0544670931994915,
-0.001341673661954701,
0.01739467680454254,
0.04136281833052635,
0.04690193384885788,
-0.02916855178773403,
-0.04130161181092262,
-0.0010399866150692105
] | 0 | 35 | 4 | 90 |
|
Shared-AE: Automatic Identification of Shared Subspaces in High-dimensional Neural and Behavioral Activity | https://openreview.net/forum?id=zXCnIyX9MG | [
"Daiyao Yi",
"Hao Dong",
"Michael James Higley",
"Anne Churchland",
"Shreya Saxena"
] | Poster | Understanding the relationship between behavior and neural activity is crucial for understanding brain function. An effective method is to learn embeddings for interconnected modalities. For simple behavioral tasks, neural features can be learned based on labels. However, complex behaviors, such as social interactions, require the joint extraction of behavioral and neural characteristics. In this paper, we present an autoencoder (AE) framework, called Shared-AE, which includes a novel regularization term that automatically identifies features shared between neural activity and behavior, while simultaneously capturing the unique private features specific to each modality. We apply Shared-AE to large-scale neural activity recorded across the entire dorsal cortex of the mouse, during two very different behaviors: (i) head-fixed mice performing a self-initiated decision-making task, and (ii) freely-moving social behavior amongst two mice. Our model successfully captures both `shared features', shared across neural and behavioral activity, and `private features', unique to each modality, significantly enhancing our understanding of the alignment between neural activity and complex behaviors. The original code for the entire Shared-AE framework on Pytorch has been made publicly available at: \url{https://github.com/saxenalab-neuro/Shared-AE}. | Computational neuroscience, Multimodal, Social behavior | null | 12,417 | null | [
-0.09272489696741104,
-0.11452662199735641,
0.05077146738767624,
0.03651705011725426,
0.06122903525829315,
0.029233835637569427,
0.049125831574201584,
-0.10352065414190292,
-0.06476107984781265,
-0.06578617542982101,
-0.0232261810451746,
-0.061805326491594315,
-0.04327542334794998,
0.07891319692134857,
-0.011306744068861008,
-0.04328778758645058,
-0.04891594126820564,
0.03507237881422043,
-0.1017918586730957,
0.00464274175465107,
-0.009553815238177776,
-0.010944021865725517,
0.0014962258283048868,
0.04196709766983986,
0.051701392978429794,
0.056135185062885284,
0.06764883548021317,
-0.03260408714413643,
0.08775924146175385,
-0.01145052257925272,
0.16512425243854523,
-0.022095024585723877,
0.033752575516700745,
0.010798062197864056,
-0.055809102952480316,
0.0384734682738781,
-0.12670265138149261,
0.031401656568050385,
-0.016869135200977325,
-0.12169601768255234,
-0.00814126804471016,
-0.04529070109128952,
0.012561910785734653,
0.03787056356668472,
-0.05352093651890755,
0.07861904799938202,
0.006590169854462147,
-0.1063830703496933,
0.01910657063126564,
-0.0317002534866333,
0.023786555975675583,
0.0388588085770607,
-0.0031734120566397905,
0.07187207043170929,
0.040027543902397156,
0.0833209902048111,
0.00836638081818819,
-0.04865649342536926,
-0.039163149893283844,
0.02810688503086567,
-0.0335838720202446,
-0.010648069903254509,
0.05534854158759117,
0.002600626088678837,
-0.009794864803552628,
0.055126626044511795,
-0.004578712396323681,
0.08775637298822403,
-0.03707226365804672,
-0.03932790830731392,
-0.03047318011522293,
0.0212753526866436,
0.022573933005332947,
0.027940765023231506,
0.04727272316813469,
-0.007405699696391821,
-0.060105014592409134,
-0.029471009969711304,
0.04834659770131111,
-0.0747622400522232,
-0.06195235252380371,
0.02575124427676201,
0.009988338686525822,
0.05248824134469032,
0.1541871875524521,
0.03647708520293236,
-0.0018947017379105091,
0.02680346556007862,
-0.02711675688624382,
0.0637742429971695,
-0.0610373392701149,
-0.03758832812309265,
-0.019827933982014656,
-0.057479340583086014,
0.029689710587263107,
-0.006396686192601919,
-0.016117269173264503,
-0.03926777094602585,
-0.008373424410820007,
0.05496472120285034,
-0.05780940130352974,
0.0767473354935646,
-0.008190970867872238,
0.024344580247998238,
0.09827826917171478,
0.0495586171746254,
0.04892604425549507,
0.004059252329170704,
0.06561538577079773,
-0.0056215329095721245,
-0.06353489309549332,
0.014712936244904995,
-0.06149837374687195,
0.0017978082178160548,
0.0855017676949501,
-0.012505419552326202,
0.0018379556713625789,
0.021628480404615402,
0.16370514035224915,
0.016896987333893776,
-0.01163931842893362,
-0.030260255560278893,
-0.01089046336710453,
0.007072216831147671,
0.04173291102051735,
-0.023286452516913414,
-0.12243077903985977,
9.653000733544272e-34,
0.026307398453354836,
-0.029038364067673683,
-0.051490362733602524,
-0.005609497427940369,
-0.024271635338664055,
0.012679069302976131,
-0.04384699836373329,
0.038658257573843,
-0.004030915442854166,
0.02256430871784687,
-0.0178678035736084,
0.03546880558133125,
0.023007329553365707,
0.14385713636875153,
0.03241397812962532,
0.008690480142831802,
0.008935721591114998,
0.006631826050579548,
0.06881053745746613,
-0.08231784403324127,
0.03887397423386574,
0.01743723452091217,
-0.04369499161839485,
0.01820055954158306,
-0.034696854650974274,
-0.08088712394237518,
0.052155692130327225,
-0.0641164630651474,
0.005042750854045153,
0.040798284113407135,
-0.05870893597602844,
0.03592779487371445,
0.014171568676829338,
-0.01571999117732048,
-0.006350221112370491,
-0.010378883220255375,
0.07230009883642197,
-0.0268801748752594,
-0.00846031028777361,
0.020968051627278328,
0.05252893269062042,
-0.02277647890150547,
-0.017364749684929848,
-0.06926098465919495,
-0.08634769916534424,
-0.04494515061378479,
-0.03479870408773422,
0.0637010931968689,
-0.05564994737505913,
-0.006402095314115286,
0.03329983726143837,
-0.0016345081385225058,
0.008122156374156475,
-0.036153487861156464,
-0.08037179708480835,
0.027920899912714958,
-0.012985744513571262,
-0.010316573083400726,
0.016285207122564316,
0.02450646460056305,
-0.025432638823986053,
0.05627283453941345,
0.01731189340353012,
-0.012564756907522678,
0.06133243441581726,
0.03919093683362007,
-0.01195782981812954,
0.018703443929553032,
-0.0016075443709269166,
0.004825323820114136,
-0.05059031397104263,
0.08752340823411942,
-0.008653836324810982,
-0.13111858069896698,
0.07202285528182983,
-0.017836693674325943,
-0.01718534342944622,
-0.12064436823129654,
-0.031247872859239578,
0.026798591017723083,
0.0017072566552087665,
0.04778235778212547,
-0.0029670281801372766,
0.0545598603785038,
-0.050066988915205,
-0.027046339586377144,
0.021881183609366417,
-0.06256859749555588,
-0.10387442260980606,
0.05720115453004837,
0.047357603907585144,
-0.07300865650177002,
0.05535799264907837,
0.012364518828690052,
-0.03856053203344345,
-1.6750043284859025e-33,
-0.00625323411077261,
0.03475888445973396,
-0.021667612716555595,
0.026749147102236748,
0.0009834151715040207,
-0.021325722336769104,
-0.03536709398031235,
0.02549642324447632,
-0.049450114369392395,
0.03420257940888405,
-0.044936493039131165,
-0.04718362167477608,
0.06684820353984833,
-0.022086182609200478,
0.10444461554288864,
0.0181649848818779,
-0.033951740711927414,
0.08819322288036346,
0.05203839763998985,
0.049840740859508514,
0.001788371242582798,
0.07127963751554489,
-0.00397998932749033,
0.0691642165184021,
0.0241096094250679,
0.008209223859012127,
0.0027415677905082703,
0.03771480545401573,
0.03527919575572014,
-0.010573270730674267,
-0.10871383547782898,
0.013837788254022598,
-0.05291774496436119,
-0.08678918331861496,
-0.03665214404463768,
0.02576505020260811,
-0.03659195452928543,
-0.04982135444879532,
-0.054052986204624176,
-0.045514751225709915,
0.07589197903871536,
-0.040292058140039444,
-0.09458622336387634,
0.04291688650846481,
0.018236389383673668,
-0.026928747072815895,
-0.11168745905160904,
0.043473973870277405,
-0.03517356887459755,
0.028513016179203987,
-0.006971101742237806,
0.020099131390452385,
-0.09308138489723206,
-0.031829070299863815,
-0.024986008182168007,
-0.022080887109041214,
0.08977498859167099,
0.008033253252506256,
0.05472542718052864,
0.0029615378007292747,
-0.0034680671524256468,
-0.01339578814804554,
-0.012392327189445496,
0.0352972112596035,
0.011081964708864689,
-0.08488152176141739,
-0.015655482187867165,
0.02992815151810646,
0.02840987592935562,
-0.021129591390490532,
0.027789274230599403,
-0.019108956679701805,
0.018853016197681427,
-0.006307260598987341,
0.011628607288002968,
-0.051618531346321106,
0.05850723013281822,
-0.09282118827104568,
0.008560197427868843,
0.042995892465114594,
0.05523442104458809,
-0.02245239168405533,
0.06169786676764488,
-0.002261986257508397,
-0.04813935607671738,
0.054769016802310944,
0.0036580655723810196,
0.06621706485748291,
-0.012191284447908401,
-0.025832897052168846,
0.008258464746177197,
0.0737219899892807,
-0.07836227118968964,
0.027785649523139,
-0.04439036548137665,
-4.3960280748933656e-8,
-0.10684455186128616,
0.016450060531497,
-0.008073070086538792,
0.03940434008836746,
0.019634567201137543,
-0.03764219954609871,
0.05490732192993164,
0.02895927242934704,
-0.055290337651968,
-0.006688034161925316,
0.10873749852180481,
-0.0385727733373642,
-0.012692030519247055,
0.014516344293951988,
-0.033875539898872375,
0.07619117200374603,
0.008163394406437874,
-0.02333100140094757,
-0.015474526211619377,
0.04911405220627785,
-0.019614912569522858,
-0.02095540426671505,
0.02301478385925293,
-0.09517001360654831,
-0.03956042602658272,
-0.08576899021863937,
-0.029115693643689156,
0.08458472788333893,
0.04213294759392738,
-0.011951718479394913,
-0.02202211692929268,
0.021157514303922653,
-0.002107231877744198,
-0.0008363878005184233,
0.006156057119369507,
0.11725933849811554,
0.03309512510895729,
-0.08458859473466873,
-0.005769530311226845,
0.0095868781208992,
0.02433089166879654,
0.04656281694769859,
0.01715417020022869,
0.010771281085908413,
0.03955848142504692,
0.03527464345097542,
0.04723699763417244,
-0.10047302395105362,
0.07586099207401276,
-0.0659274160861969,
0.04092077538371086,
0.0034614959731698036,
0.02669970691204071,
0.07848111540079117,
0.022983243688941002,
0.026952233165502548,
-0.09643535315990448,
-0.056296609342098236,
0.09125436097383499,
-0.007785580586642027,
0.009238924831151962,
0.09861216694116592,
-0.13662581145763397,
-0.01982834003865719
] | 0 | 0 | 0 | 0 |
|
Bounds on Lp Errors in Density Ratio Estimation via f-Divergence Loss Functions | https://openreview.net/forum?id=ttq44QjKda | [
"Yoshiaki Kitazawa"
] | Poster | Density ratio estimation (DRE) is a core technique in machine learning used to capture relationships between two probability distributions.
$f$-divergence loss functions, which are derived from variational representations of $f$-divergence, have become a standard choice in DRE for achieving cutting-edge performance.
This study provides novel theoretical insights into DRE by deriving upper and lower bounds on the $L_p$ errors through $f$-divergence loss functions.
These bounds apply to any estimator belonging to a class of Lipschitz continuous estimators, irrespective of the specific $f$-divergence loss function employed.
The derived bounds are expressed as a product involving the data dimensionality and the expected value of the density ratio raised to the $p$-th power.
Notably, the lower bound includes an exponential term that depends on the Kullback--Leibler (KL) divergence, revealing that the $L_p$ error increases significantly as the KL divergence grows when $p > 1$.
This increase becomes even more pronounced as the value of $p$ grows.
The theoretical insights are validated through numerical experiments. | density ratio estimation, variational divergence optimization, Kullback–Leibler divergence, $f$-divergence, $L_p$ error, the curse of dimensionality, GAN | We provide both lower and upper bounds on $L_p$ errors in DRE, which hold for any member of a group of Lipschitz continuous estimators regardless of the specific $f$-divergence loss functions used. | 12,415 | null | [
-0.03377993777394295,
-0.05820171907544136,
0.033833205699920654,
0.03050893172621727,
0.06566458195447922,
0.011796901002526283,
-0.02178144082427025,
0.029436951503157616,
0.04525996744632721,
0.013762119226157665,
-0.10376915335655212,
0.04056403040885925,
0.03958220034837723,
-0.010700284503400326,
-0.02324889600276947,
-0.04388101398944855,
-0.008295423351228237,
0.12467604875564575,
-0.11402440071105957,
0.03995434194803238,
0.02847418002784252,
0.07082279771566391,
0.015853429213166237,
0.005805197637528181,
0.03827499598264694,
-0.05455005541443825,
-0.009284990839660168,
-0.010121640749275684,
0.0032724200282245874,
-0.04642008617520332,
0.040844693779945374,
0.0018942990573123097,
0.010885401628911495,
-0.03209211677312851,
0.013322296552360058,
-0.00996505655348301,
-0.015166210941970348,
0.010336246341466904,
0.06498045474290848,
0.03051656112074852,
-0.02700592577457428,
0.016410237178206444,
0.0284411683678627,
0.021929649636149406,
0.01585967279970646,
-0.014032389968633652,
-0.03976346552371979,
-0.0474700890481472,
0.0002996229741256684,
0.029407775029540062,
0.057778432965278625,
0.05162932351231575,
-0.016791975125670433,
0.02510649710893631,
-0.04454956203699112,
-0.04180550202727318,
0.011796323582530022,
0.011892384849488735,
-0.0031510095577687025,
0.11000426113605499,
-0.050320353358983994,
-0.08980601280927658,
-0.02049979567527771,
0.004410822410136461,
-0.06650856137275696,
-0.07254267483949661,
0.030475448817014694,
0.03154490515589714,
-0.04920186474919319,
0.0522502176463604,
-0.10032409429550171,
0.07647093385457993,
-0.1485763043165207,
0.040999963879585266,
0.04064319282770157,
-0.0012870361097157001,
0.06492762267589569,
0.031156424432992935,
0.011212687008082867,
-0.09668394178152084,
0.01953618787229061,
0.047224633395671844,
-0.049890805035829544,
-0.05797707661986351,
0.06882546842098236,
-0.027794990688562393,
-0.06119590252637863,
0.05749167874455452,
0.08865488320589066,
-0.051776692271232605,
-0.0308774895966053,
-0.00014811633445788175,
-0.08174241334199905,
0.0720660462975502,
0.028621844947338104,
-0.030191130936145782,
0.027975866571068764,
-0.049694787710905075,
0.04222952201962471,
0.053758736699819565,
-0.04190268740057945,
0.04238656163215637,
-0.019594207406044006,
-0.06357051432132721,
0.037779033184051514,
-0.1477842479944229,
0.042280785739421844,
-0.045101817697286606,
0.06658238172531128,
-0.019959833472967148,
0.040359001606702805,
0.018590465188026428,
0.026347488164901733,
0.04136605188250542,
0.007491038180887699,
-0.06323519349098206,
0.008028487674891949,
0.02296232059597969,
0.04545699059963226,
0.015310900285840034,
-0.12562188506126404,
-0.03686892241239548,
0.010440511628985405,
0.11311075836420059,
-0.00142851157579571,
-0.019762033596634865,
-0.07154814153909683,
5.7863106978372306e-33,
-0.04502741992473602,
-0.029323125258088112,
0.05885053426027298,
0.019596092402935028,
0.027448704466223717,
0.05533578619360924,
-0.0037673464976251125,
-0.03490319103002548,
0.017757665365934372,
0.04709133878350258,
-0.044889021664857864,
0.033800289034843445,
-0.024559682235121727,
-0.058063775300979614,
0.023343661800026894,
-0.0030901185236871243,
-0.06868379563093185,
0.12455356121063232,
-0.014240183867514133,
-0.06664252281188965,
0.013588794507086277,
-0.0991978868842125,
0.056680191308259964,
0.05746360868215561,
-0.01639445312321186,
0.0033536201808601618,
0.014072866179049015,
0.03479804843664169,
-0.010395502671599388,
0.010822337120771408,
-0.07066706568002701,
0.016854846850037575,
0.07103212177753448,
-0.08781053125858307,
-0.034958504140377045,
-0.02340499684214592,
0.007592206820845604,
0.0070838527753949165,
-0.09800170361995697,
0.017148522660136223,
-0.08688007295131683,
0.021977601572871208,
-0.01511551532894373,
-0.07371898740530014,
-0.10270287096500397,
-0.12430406361818314,
0.07319990545511246,
-0.04339561611413956,
0.0054721469059586525,
-0.07121290266513824,
-0.04397884011268616,
0.011579031124711037,
0.003931042272597551,
0.021082470193505287,
0.05567304044961929,
0.033724986016750336,
-0.028623564168810844,
0.03473973646759987,
0.0348685085773468,
-0.004876389633864164,
-0.01835891231894493,
0.10470464080572128,
-0.0003809196350630373,
-0.0059397402219474316,
0.04845022037625313,
0.06458836048841476,
0.05099836364388466,
0.08286416530609131,
-0.016581164672970772,
0.028320923447608948,
0.06714674830436707,
-0.013163828290998936,
-0.019867228344082832,
-0.040961865335702896,
0.0501219779253006,
0.04301214963197708,
0.013600964099168777,
-0.02345135807991028,
0.06374523788690567,
0.008756473660469055,
0.003355755005031824,
0.013524199835956097,
0.000799844041466713,
-0.038923799991607666,
-0.08439133316278458,
-0.09670143574476242,
0.009548481553792953,
0.01122113224118948,
-0.025905439630150795,
-0.01664244942367077,
-0.09364095330238342,
-0.07874015718698502,
-0.02092798613011837,
0.01058034785091877,
-0.0034834188409149647,
-6.807206235691279e-33,
-0.0891181007027626,
0.029867790639400482,
-0.01855957694351673,
0.038891591131687164,
0.007971557788550854,
-0.07101570814847946,
0.01034505944699049,
0.053966570645570755,
0.025228086858987808,
-0.0973457470536232,
-0.029090501368045807,
0.017044346779584885,
-0.024423256516456604,
0.00627874443307519,
-0.016657723113894463,
0.035988010466098785,
-0.06251361966133118,
0.0159784946590662,
0.027470501139760017,
-0.005821465514600277,
0.062308814376592636,
-0.02052195742726326,
0.0200313962996006,
0.09514480829238892,
-0.05059751123189926,
0.005724220536649227,
-0.056983582675457,
0.12535874545574188,
-0.06715905666351318,
-0.10298685729503632,
0.025905901566147804,
-0.031091637909412384,
-0.08808255940675735,
-0.006815213244408369,
-0.031145019456744194,
-0.08978337794542313,
-0.015948232263326645,
0.062226831912994385,
-0.03554791957139969,
0.10376383364200592,
0.043459538370370865,
0.040337853133678436,
0.059027012437582016,
-0.09403454512357712,
0.048226747661828995,
-0.037892796099185944,
0.06472745537757874,
-0.04673638567328453,
0.13212406635284424,
-0.003948333207517862,
0.007655663415789604,
0.014683298766613007,
-0.036891888827085495,
0.10568428784608841,
-0.06222231313586235,
0.02713143639266491,
-0.00024329214647877961,
-0.002727649873122573,
0.029038146138191223,
0.06186511367559433,
-0.08600091189146042,
-0.031152023002505302,
-0.04688168689608574,
0.10406912863254547,
-0.08343096077442169,
-0.030910082161426544,
-0.03368854522705078,
0.01277951616793871,
0.05210421606898308,
0.08580503612756729,
0.02209068089723587,
-0.031840067356824875,
0.0631990134716034,
0.08076485246419907,
-0.03663206845521927,
0.04973412677645683,
0.020478343591094017,
-0.04083042964339256,
-0.031955499202013016,
0.020217617973685265,
0.020054660737514496,
0.0003278304648119956,
0.005170776974409819,
0.034732941538095474,
0.06526602059602737,
-0.004568936303257942,
-0.010885694995522499,
-0.086641326546669,
0.014582918956875801,
-0.0670725554227829,
-0.03906913846731186,
0.004107233136892319,
-0.020966438576579094,
-0.029151979833841324,
-0.03849407657980919,
-5.3413053535678046e-8,
-0.042115919291973114,
0.0421283133327961,
0.020492099225521088,
-0.0308912992477417,
0.11031628400087357,
-0.018527517095208168,
-0.027981704100966454,
0.012733998708426952,
-0.009648894891142845,
-0.04760102927684784,
0.1028909832239151,
-0.03597467765212059,
0.007518645375967026,
0.02448621578514576,
0.02255295217037201,
0.03055446781218052,
0.023245342075824738,
-0.014300910755991936,
0.07219096273183823,
-0.006598563399165869,
0.024840205907821655,
0.06103998050093651,
0.08590161800384521,
-0.10586338490247726,
0.05777851864695549,
-0.11032548546791077,
-0.031049121171236038,
0.07473461329936981,
-0.025122512131929398,
-0.0424775704741478,
0.005179297178983688,
0.07002294063568115,
0.0070124296471476555,
0.012506240978837013,
0.018611427396535873,
0.07759373635053635,
0.0062737539410591125,
-0.037953201681375504,
-0.04915968328714371,
0.08598431944847107,
0.03860415518283844,
0.010938580147922039,
-0.002192818559706211,
-0.036740805953741074,
-0.009519390761852264,
0.059785451740026474,
0.08093568682670593,
-0.012123491615056992,
0.002221823437139392,
-0.09842053055763245,
0.05555914714932442,
0.04661109670996666,
0.011221583932638168,
0.023634662851691246,
0.08838491141796112,
-0.04517005756497383,
-0.05212000757455826,
-0.021398378536105156,
-0.01785798743367195,
0.019612884148955345,
-0.02831103652715683,
-0.005550430156290531,
-0.021829087287187576,
0.013042120262980461
] | 0 | 0 | 0 | 0 |
|
Palu: KV-Cache Compression with Low-Rank Projection | https://openreview.net/forum?id=LWMS4pk2vK | [
"Chi-Chih Chang",
"Wei-Cheng Lin",
"Chien-Yu Lin",
"Chong-Yan Chen",
"Yu-Fang Hu",
"Pei-Shuo Wang",
"Ning-Chi Huang",
"Luis Ceze",
"Mohamed S. Abdelfattah",
"Kai-Chiang Wu"
] | Poster | Post-training KV-Cache compression methods typically either sample a subset of effectual tokens or quantize the data into lower numerical bit width. However, these methods cannot exploit redundancy in the hidden dimension of the KV tenors. This paper presents a hidden dimension compression approach called Palu, a KV-Cache compression framework that utilizes low-rank projection to reduce inference-time LLM memory usage. Palu decomposes the linear layers into low-rank matrices, caches compressed intermediate states, and reconstructs the full keys and values on the fly. To improve accuracy, compression rate, and efficiency, Palu further encompasses (1) a medium-grained low-rank decomposition scheme, (2) an efficient rank search algorithm, (3) low-rank-aware quantization compatibility enhancements, and (4) an optimized GPU kernel with matrix fusion. Extensive experiments with popular LLMs show that Palu compresses KV-Cache by 50% while maintaining strong accuracy and delivering up to 1.89× speedup on the RoPE-based attention module. When combined with quantization, Palu’s
inherent quantization-friendly design yields small to negligible extra accuracy degradation while saving additional memory than quantization-only methods and achieving up to 2.91× speedup for the RoPE-based attention. Moreover, it maintains comparable or even better accuracy (up to 1.19 lower perplexity) compared to quantization-only methods. These results demonstrate Palu’s superior capability to effectively address the efficiency and memory challenges of LLM inference posed by KV-Cache. Our code is publicly available at: https://github.com/shadowpa0327/Palu. | KV-Cache, Low-Rank Compression, Large Language Model | We propose a novel low-rank KV-Cache compression method to reduce the memory footprint and accelerate decoding efficiency | 12,412 | 2407.21118 | [
-0.010840684175491333,
0.05122053995728493,
-0.03577963262796402,
0.02561609074473381,
0.0032473215833306313,
0.03426337614655495,
-0.06263148784637451,
-0.008161322213709354,
0.0020155294332653284,
-0.02301400527358055,
0.0003487439244054258,
-0.004924263339489698,
-0.015001687221229076,
0.011900033801794052,
-0.033357199281454086,
-0.03420519083738327,
0.0439477376639843,
0.15708394348621368,
-0.01787077635526657,
-0.06303592026233673,
0.013958074152469635,
-0.028145870193839073,
-0.001002261065877974,
0.012346332892775536,
0.1302151381969452,
0.013635835610330105,
-0.0045976173132658005,
-0.004916192498058081,
0.058791011571884155,
-0.01675551012158394,
0.046354323625564575,
0.038087960332632065,
0.034879814833402634,
-0.005612237378954887,
-0.07166757434606552,
0.05220871418714523,
-0.020921051502227783,
-0.03257140889763832,
0.005359156057238579,
-0.03171619027853012,
-0.019398072734475136,
0.04526404291391373,
-0.0355919785797596,
0.033517517149448395,
0.07970331609249115,
0.08510252088308334,
0.06750788539648056,
-0.03997046500444412,
-0.030442342162132263,
-0.031087491661310196,
0.009750369936227798,
0.03511926159262657,
-0.03220422938466072,
0.10062810778617859,
-0.12557223439216614,
-0.04010361060500145,
0.013555706478655338,
-0.011323905549943447,
-0.00919833779335022,
0.0157993845641613,
-0.036412909626960754,
-0.04435475543141365,
0.03231309726834297,
-0.03203772380948067,
-0.010034778155386448,
-0.05035829171538353,
0.09662727266550064,
-0.00702928751707077,
-0.008474403992295265,
-0.009931503795087337,
-0.029402554035186768,
-0.019413353875279427,
-0.006615978665649891,
0.014399705454707146,
-0.05143425241112709,
0.059693578630685806,
0.016002854332327843,
-0.05401386320590973,
0.08666035532951355,
0.0207234974950552,
0.03132520988583565,
-0.013856285251677036,
-0.014049661345779896,
-0.01730283349752426,
0.02240767329931259,
-0.019648561254143715,
0.08012295514345169,
-0.03031047433614731,
-0.011341837234795094,
-0.07844870537519455,
0.002863469999283552,
0.013214699923992157,
0.07523979246616364,
0.0029804043006151915,
0.07530874758958817,
-0.02557450160384178,
0.019637567922472954,
0.07385143637657166,
0.002449299907311797,
0.07532849907875061,
0.04792933538556099,
0.03583228588104248,
-0.06749065965414047,
0.03647661954164505,
0.03800171613693237,
0.033019911497831345,
0.012437708675861359,
0.07191731780767441,
-0.000604226894211024,
-0.060976698994636536,
0.04625367000699043,
0.02954547479748726,
-0.04542754217982292,
-0.04546300694346428,
0.03693762049078941,
0.029732229188084602,
-0.007504930254071951,
-0.017656205222010612,
0.059906281530857086,
0.01171464566141367,
-0.01831119880080223,
-0.06846258044242859,
-0.02365317940711975,
0.020676646381616592,
-0.02258988656103611,
-0.011243109591305256,
-0.05916370078921318,
3.9029637709925014e-33,
0.037833068519830704,
0.04510022699832916,
-0.011181415989995003,
-0.019629016518592834,
0.03984016180038452,
0.01059903297573328,
0.04503410682082176,
0.005705299321562052,
0.00013830004900228232,
0.07532232999801636,
-0.06344466656446457,
-0.00052351487101987,
-0.03945554420351982,
0.07623446732759476,
0.020471913740038872,
0.03968823701143265,
-0.023946335539221764,
0.033207591623067856,
-0.05521329119801521,
-0.08767937123775482,
0.08107936382293701,
-0.0001542562822578475,
0.05212119594216347,
-0.05307942256331444,
0.05062070116400719,
0.02410464733839035,
0.03937339782714844,
-0.12281915545463562,
-0.07670820504426956,
-0.013989715836942196,
-0.044909726828336716,
-0.09188570827245712,
-0.022655954584479332,
-0.000072532806370873,
0.022748906165361404,
0.0028300562407821417,
0.03400213643908501,
-0.007001867517828941,
0.042752690613269806,
-0.05618475377559662,
0.0027054487727582455,
0.01948116160929203,
0.018444739282131195,
-0.11426154524087906,
-0.16435091197490692,
-0.013508050702512264,
0.023545902222394943,
-0.005397280212491751,
-0.037445005029439926,
-0.001803667633794248,
0.06531578302383423,
-0.026492774486541748,
-0.042427629232406616,
0.04360952973365784,
0.04302724078297615,
-0.12465395033359528,
0.11562036722898483,
0.040715668350458145,
0.06629213690757751,
0.10861002653837204,
-0.10713877528905869,
-0.03674441576004028,
-0.06317011266946793,
0.08537130057811737,
-0.04990072920918465,
-0.019404783844947815,
-0.0006736965733580291,
0.02454116754233837,
0.015580018050968647,
0.03642589598894119,
-0.0714876651763916,
0.047206275165081024,
0.023653723299503326,
-0.08375725895166397,
0.006371700670570135,
0.0007512680604122579,
0.030925268307328224,
-0.04282389581203461,
-0.04924260452389717,
-0.021025346592068672,
0.012655913829803467,
0.04560655355453491,
0.009990064427256584,
-0.005887081380933523,
-0.037461187690496445,
-0.07803785800933838,
0.008223346434533596,
-0.02046777680516243,
0.02422775700688362,
-0.045465391129255295,
-0.06937339901924133,
-0.01864379458129406,
0.06124025583267212,
-0.02287054806947708,
-0.06727886199951172,
-1.4719594443939417e-33,
-0.003360279370099306,
0.05878368392586708,
-0.007958375848829746,
0.06147022917866707,
-0.013421319425106049,
-0.045450322329998016,
0.03865094855427742,
-0.011755887418985367,
-0.06614722311496735,
-0.09663648158311844,
-0.06859805434942245,
-0.07619109004735947,
0.046040721237659454,
0.04947134852409363,
0.08777809888124466,
0.07945504784584045,
-0.07890323549509048,
-0.04963256046175957,
-0.10264324396848679,
-0.006075617391616106,
-0.06596220284700394,
0.0763743445277214,
-0.026153845712542534,
0.024995630607008934,
-0.005287970881909132,
0.03235333412885666,
0.037849172949790955,
0.005943238735198975,
0.024924222379922867,
-0.060519371181726456,
-0.006794255692511797,
-0.029056543484330177,
-0.039919208735227585,
0.01745804399251938,
-0.0074977390468120575,
0.06399805843830109,
0.06093273311853409,
0.03230210393667221,
-0.040521036833524704,
0.08713168650865555,
0.017319153994321823,
-0.0051509179174900055,
-0.08143429458141327,
-0.009482099674642086,
0.009074501693248749,
0.054432861506938934,
-0.06604637950658798,
-0.004671306349337101,
0.04590034484863281,
0.02781773917376995,
0.03215966001152992,
0.04527091234922409,
-0.021998055279254913,
0.07288751006126404,
0.014912408776581287,
-0.07335597276687622,
-0.12659364938735962,
0.10005570948123932,
0.04024152085185051,
-0.04161571338772774,
0.1180429607629776,
-0.09876763820648193,
0.02523275464773178,
-0.09959512203931808,
-0.04773182049393654,
-0.06264391541481018,
0.020401638001203537,
-0.012125594541430473,
-0.05819215252995491,
0.055420029908418655,
-0.017476756125688553,
-0.047306351363658905,
0.026463640853762627,
-0.027147119864821434,
-0.03600505739450455,
-0.015864817425608635,
0.01689709722995758,
-0.038302384316921234,
0.017299193888902664,
0.0342782586812973,
0.05319890007376671,
0.009864605963230133,
0.02740676887333393,
0.028695058077573776,
0.09069111943244934,
0.11902427673339844,
0.03653258830308914,
0.037688903510570526,
0.031125959008932114,
-0.043100506067276,
-0.03754856809973717,
0.044292621314525604,
0.0015553102130070329,
0.031013481318950653,
0.004293718375265598,
-4.113531204552601e-8,
-0.01484763901680708,
-0.0046875737607479095,
-0.06006305664777756,
0.06350934505462646,
-0.028973160311579704,
-0.05910082533955574,
0.03512752056121826,
0.07825417816638947,
0.05742400139570236,
-0.032914016395807266,
0.1297159641981125,
-0.11860685795545578,
-0.06762178987264633,
-0.0493658185005188,
0.04660015180706978,
0.010419096797704697,
-0.04976329952478409,
0.015501570887863636,
0.03432716056704521,
-0.04290458932518959,
-0.07200197875499725,
0.09284511208534241,
0.027859827503561974,
-0.008086981251835823,
-0.0066667357459664345,
-0.053332533687353134,
0.06918163597583771,
0.11400893330574036,
0.1394803524017334,
0.010748685337603092,
0.023582838475704193,
0.03135534003376961,
0.08871623873710632,
-0.08565615862607956,
0.06557992845773697,
0.0414738692343235,
0.023733587935566902,
0.08301617205142975,
0.04267145320773125,
0.005903137382119894,
-0.03162076324224472,
-0.05166042223572731,
-0.07702779769897461,
0.04183338209986687,
-0.008290287107229233,
0.025638852268457413,
-0.002743989462032914,
-0.054071493446826935,
0.022797537967562675,
0.03710411489009857,
0.03894668072462082,
-0.0026083732955157757,
-0.018092261627316475,
0.035010602325201035,
0.012924401089549065,
0.01649457775056362,
-0.00558881089091301,
0.03484038636088371,
0.06600019335746765,
-0.0414535216987133,
0.024883724749088287,
-0.09825915098190308,
-0.08555273711681366,
-0.005669207312166691
] | https://github.com/shadowpa0327/Palu | 101 | 0 | 0 | 0 |
Samba: Simple Hybrid State Space Models for Efficient Unlimited Context Language Modeling | https://openreview.net/forum?id=bIlnpVM4bc | [
"Liliang Ren",
"Yang Liu",
"Yadong Lu",
"yelong shen",
"Chen Liang",
"Weizhu Chen"
] | Poster | Efficiently modeling sequences with infinite context length has long been a challenging problem. Previous approaches have either suffered from quadratic computational complexity or limited extrapolation ability in length generalization. In this
work, we present Samba, a simple hybrid architecture that layer-wise combines
Mamba, a selective State Space Model (SSM), with Sliding Window Attention
(SWA). Samba selectively compresses a given sequence into recurrent hidden
states while still maintaining the ability to precisely recall recent memories with the
attention mechanism. We scale Samba up to 3.8B parameters with 3.2T training
tokens and demonstrate that it significantly outperforms state-of-the-art models
across a variety of benchmarks. Pretrained on sequences of 4K length, Samba
shows improved perplexity in context lengths of up to 1M in zero-shot. When
finetuned on 4K-length sequences, Samba efficiently extrapolates to a 256K context length with perfect memory recall on the Passkey Retrieval task, and exhibits
superior retrieval extrapolation on the challenging Phonebook task compared to
full-attention models. As a linear-time sequence model, Samba achieves a 3.73×
higher throughput compared to Transformers with grouped-query attention for user
prompts of 128K length, and a 3.64× speedup when generating 64K tokens with
unlimited streaming. | Large Language Models;Length Extrapolation;Efficiency;Hybrid State Space Models | Simple hybrid state space models outperform SOTA Transformers and SSMs. | 12,408 | 2406.07522 | [
-0.03519893437623978,
-0.06807400286197662,
-0.014219937846064568,
0.06892021745443344,
0.0024100621230900288,
0.06342937052249908,
-0.03745841607451439,
-0.0050075422041118145,
0.04630516469478607,
-0.07057874649763107,
0.01115509681403637,
-0.04592546448111534,
0.05096893012523651,
0.008710737340152264,
0.027756283059716225,
-0.011321346275508404,
0.08639319986104965,
0.01779044046998024,
-0.060573041439056396,
-0.05640518292784691,
-0.002992978785187006,
-0.0189176257699728,
0.05274254456162453,
-0.02526339702308178,
0.010405385866761208,
0.038343530148267746,
-0.019300663843750954,
-0.04995390772819519,
0.01710641197860241,
0.013016395270824432,
0.08093693852424622,
0.04953722655773163,
0.0003262509999331087,
0.025515612214803696,
-0.03532238304615021,
0.021070446819067,
-0.09565624594688416,
-0.015768636018037796,
-0.04006897658109665,
-0.12526074051856995,
0.052381839603185654,
0.0198508407920599,
0.01044715940952301,
0.10056690871715546,
0.05950012430548668,
-0.007132640108466148,
0.07236698269844055,
-0.011680470779538155,
0.014731843955814838,
-0.023220716044306755,
-0.10208878666162491,
0.030614737421274185,
-0.016368942335247993,
0.09130655974149704,
-0.06394556909799576,
-0.04140122979879379,
-0.023464152589440346,
0.10619865357875824,
-0.023152664303779602,
-0.008104551583528519,
-0.0785033255815506,
-0.11372904479503632,
-0.0334126353263855,
-0.059948962181806564,
-0.010278062894940376,
0.010374115779995918,
-0.018527086824178696,
0.04921681433916092,
-0.025989916175603867,
0.029815133661031723,
-0.031908340752124786,
0.0422661192715168,
-0.03146906569600105,
0.05377133563160896,
-0.013570990413427353,
0.03506385535001755,
0.024285469204187393,
-0.005144143011420965,
0.06422549486160278,
-0.09256414324045181,
0.017405299469828606,
-0.04168586805462837,
-0.014189057983458042,
-0.051801860332489014,
0.06129852682352066,
0.018581947311758995,
-0.04480297863483429,
0.043980106711387634,
-0.007814018987119198,
0.051108211278915405,
0.015863653272390366,
-0.10150384902954102,
0.04502703621983528,
-0.01831125095486641,
0.01597079634666443,
0.049003638327121735,
-0.023043781518936157,
-0.009526272304356098,
-0.07151526212692261,
0.09428947418928146,
0.0399177260696888,
0.1061621680855751,
0.0326007641851902,
-0.03627200052142143,
-0.0035486165434122086,
0.0092973867431283,
0.016645807772874832,
0.06993235647678375,
-0.009694010019302368,
-0.023400135338306427,
0.01730448007583618,
0.06260967999696732,
0.035738810896873474,
-0.005391273647546768,
0.03823511675000191,
-0.04428961127996445,
0.020265569910407066,
-0.07538405805826187,
0.023395374417304993,
0.0950692892074585,
0.016013873741030693,
-0.016587818041443825,
-0.0461074598133564,
-0.007929879240691662,
-0.058052193373441696,
-0.02147219330072403,
-0.03441166505217552,
2.3863174496745696e-33,
0.011275386437773705,
0.008972649462521076,
-0.004108574241399765,
0.05041966214776039,
0.053632766008377075,
0.012396949343383312,
0.0568632073700428,
0.016497526317834854,
-0.07605186104774475,
0.022162742912769318,
0.004982318263500929,
0.008325465954840183,
-0.02221124805510044,
0.030595632269978523,
0.032986804842948914,
-0.03404652327299118,
-0.05370860546827316,
0.02524714730679989,
0.010046087205410004,
-0.009902648627758026,
-0.010436657816171646,
0.03416833281517029,
0.01206861063838005,
-0.03888922184705734,
0.0801858976483345,
0.005533584393560886,
0.08661741018295288,
-0.05820224806666374,
-0.015712494030594826,
0.02348952554166317,
-0.10665380209684372,
0.05425252765417099,
-0.06224178150296211,
-0.01661856099963188,
0.04654235020279884,
-0.040418658405542374,
-0.018586255609989166,
-0.0779903307557106,
0.02181035466492176,
-0.09249086678028107,
-0.03751586750149727,
-0.0007345114136114717,
0.04458235949277878,
0.08061905205249786,
-0.11141505092382431,
-0.1031254455447197,
0.005012787878513336,
0.027775723487138748,
-0.03219352662563324,
0.03396541252732277,
0.10438328236341476,
0.03587184101343155,
-0.04807883873581886,
-0.01812499202787876,
-0.010914820246398449,
-0.02552207186818123,
0.050402820110321045,
0.078181691467762,
0.010023646056652069,
0.1150006577372551,
0.06460651755332947,
0.04116995260119438,
0.007825720123946667,
0.15133172273635864,
-0.004632623866200447,
0.019365163519978523,
-0.03469211235642433,
0.13951022922992706,
0.027095718309283257,
-0.014538588002324104,
-0.0007096912595443428,
-0.029153939336538315,
0.00569223752245307,
0.018327776342630386,
0.04686737805604935,
-0.0333576425909996,
0.11497824639081955,
-0.10583943128585815,
-0.03103059157729149,
0.00021370779722929,
0.0309443399310112,
-0.018509963527321815,
0.004911676049232483,
0.017991608008742332,
-0.10327159613370895,
-0.06127389892935753,
0.03986091911792755,
-0.07232152670621872,
-0.07339637726545334,
-0.12547852098941803,
0.04157789796590805,
0.017077460885047913,
0.07112856954336166,
0.04175962507724762,
-0.04152147099375725,
-1.2778472153422765e-33,
0.06936479359865189,
-0.031278062611818314,
-0.019185159355401993,
0.00715757766738534,
-0.04604477807879448,
-0.06232369691133499,
0.014101057313382626,
0.055474162101745605,
-0.08296684175729752,
-0.021543145179748535,
-0.045457229018211365,
-0.048597101122140884,
0.12626615166664124,
-0.011031516827642918,
0.08106520771980286,
0.03315212205052376,
0.0682208240032196,
-0.04766266420483589,
-0.016373010352253914,
0.028957193717360497,
-0.008295002393424511,
0.07543522864580154,
-0.05368427559733391,
0.007269138470292091,
-0.07882088422775269,
0.003603052580729127,
-0.051065102219581604,
0.06465785950422287,
0.017714280635118484,
-0.04208173230290413,
-0.02903170883655548,
-0.0005412722821347415,
-0.01577015593647957,
-0.057333048433065414,
-0.1095045879483223,
-0.00046258518705144525,
0.03845993056893349,
0.006212148815393448,
-0.03299270197749138,
0.04147091135382652,
0.1253339648246765,
-0.04036892205476761,
-0.06105949357151985,
0.07701605558395386,
-0.03748512268066406,
0.06336469203233719,
-0.07152090221643448,
-0.007191551383584738,
-0.01749091036617756,
-0.0005553287919610739,
-0.03951864689588547,
0.010341204702854156,
-0.05228434503078461,
-0.022329792380332947,
-0.03724633902311325,
-0.0606938861310482,
-0.07977533340454102,
-0.027980538085103035,
0.038976121693849564,
-0.05949900671839714,
0.031087758019566536,
-0.02321884036064148,
0.022480962797999382,
-0.052176643162965775,
0.04624118655920029,
0.001540740835480392,
0.027456318959593773,
-0.005226537119597197,
-0.03282010555267334,
-0.027043579146265984,
0.027078678831458092,
-0.03592349588871002,
-0.08225010335445404,
0.061738960444927216,
0.002518939785659313,
0.06801827996969223,
0.08023401349782944,
-0.04121981933712959,
-0.0048090615309774876,
-0.01507820188999176,
-0.06455682218074799,
0.06137826293706894,
-0.02301674336194992,
0.0307762548327446,
0.043349552899599075,
0.1074318215250969,
0.053491078317165375,
0.057732392102479935,
-0.003268361324444413,
0.01748502068221569,
-0.05157505348324776,
0.00996820442378521,
0.030401425436139107,
0.061054036021232605,
-0.014296136796474457,
-4.003121745199678e-8,
-0.10257571935653687,
0.03184173256158829,
-0.016732288524508476,
0.028529059141874313,
-0.0067189414985477924,
-0.03300277516245842,
0.013742579147219658,
0.08073040843009949,
0.03438840061426163,
-0.01605279929935932,
0.01706804893910885,
-0.02636890299618244,
-0.06041134148836136,
-0.029372580349445343,
-0.05009372904896736,
0.08771009743213654,
0.03561770170927048,
0.003984478302299976,
-0.005124709103256464,
0.00625733332708478,
0.03787161782383919,
0.05298226326704025,
0.05255195125937462,
0.05026942491531372,
-0.014836619608104229,
-0.05323014408349991,
-0.04900267720222473,
0.10075058788061142,
0.0998385101556778,
-0.03771241009235382,
0.014724229462444782,
-0.04022881016135216,
-0.023965051397681236,
-0.10811448097229004,
-0.14245273172855377,
0.052344631403684616,
0.02873329259455204,
-0.006492478772997856,
0.03321436047554016,
0.008110560476779938,
0.019989565014839172,
-0.020269183441996574,
-0.013466035947203636,
0.022374235093593597,
-0.08693424612283707,
0.023398276418447495,
-0.015417053364217281,
-0.118213951587677,
0.025353437289595604,
-0.04125345125794411,
0.03910042345523834,
-0.028845280408859253,
0.0009894486283883452,
0.024597294628620148,
0.034946613013744354,
0.03133421763777733,
0.024944934993982315,
0.007757777813822031,
0.05094694346189499,
0.03527790307998657,
-0.0012093262048438191,
0.021715233102440834,
-0.12964028120040894,
0.07002883404493332
] | https://github.com/microsoft/Samba | 866 | 0 | 0 | 0 |
Nova: Generative Language Models for Assembly Code with Hierarchical Attention and Contrastive Learning | https://openreview.net/forum?id=4ytRL3HJrq | [
"Nan Jiang",
"Chengxiao Wang",
"Kevin Liu",
"Xiangzhe Xu",
"Lin Tan",
"Xiangyu Zhang",
"Petr Babkin"
] | Poster | Binary code analysis is the foundation of crucial tasks in the security domain; thus building effective binary analysis techniques is more important than ever. Large language models (LLMs) although have brought impressive improvement to source code tasks, do not directly generalize to assembly code due to the unique challenges of assembly: (1) the low information density of assembly and (2) the diverse optimizations in assembly code. To overcome these challenges, this work proposes a hierarchical attention mechanism that builds attention summaries to capture the semantics more effectively and designs contrastive learning objectives to train LLMs to learn assembly optimization. Equipped with these techniques, this work develops Nova, a generative LLM for assembly code. Nova outperforms existing techniques on binary code decompilation by up to 14.84 -- 21.58% higher Pass@1 and Pass@10, and outperforms the latest binary code similarity detection techniques by up to 6.17% Recall@1, showing promising abilities on both assembly generation and understanding tasks. | large language model, hierarchical attention, contrastive learning, assembly code | null | 12,395 | 2311.13721 | [
-0.00616413401439786,
-0.0383281335234642,
-0.02243560366332531,
-0.005947340279817581,
-0.03010747767984867,
-0.04050963744521141,
-0.008703607134521008,
-0.01697300374507904,
-0.010162277147173882,
0.004159079398959875,
-0.059541255235672,
-0.045570049434900284,
0.04226730018854141,
-0.018072059378027916,
-0.06580549478530884,
0.03709119185805321,
0.03656718134880066,
0.07062166929244995,
-0.04544122889637947,
-0.02877470664680004,
0.0545097179710865,
0.011572892777621746,
0.025129156187176704,
0.029105527326464653,
0.020286913961172104,
0.08933183550834656,
-0.010004542768001556,
0.027562089264392853,
0.05014876276254654,
-0.06302684545516968,
0.07006391882896423,
0.07526718080043793,
0.06986831873655319,
0.11701866239309311,
0.01033447589725256,
0.12054148316383362,
-0.04366106912493706,
-0.03892665356397629,
-0.0325310155749321,
-0.1089310348033905,
-0.07712489366531372,
0.09106787294149399,
0.011245042085647583,
0.038132376968860626,
0.058923929929733276,
-0.036449212580919266,
0.0004118139622732997,
-0.056279320269823074,
-0.021534038707613945,
-0.08200487494468689,
-0.0528251975774765,
-0.08178967982530594,
-0.028305748477578163,
-0.057316556572914124,
-0.019477345049381256,
0.028964929282665253,
0.04873228073120117,
-0.01614329405128956,
-0.0010303982999175787,
-0.02291826531291008,
-0.06633609533309937,
-0.041601985692977905,
-0.051771774888038635,
-0.04959217831492424,
0.029306964948773384,
-0.01766033098101616,
0.06966271996498108,
0.0431072898209095,
-0.042210377752780914,
-0.05910106748342514,
-0.04796755686402321,
0.048305314034223557,
-0.0017389910062775016,
0.06638546288013458,
0.0011188848875463009,
0.09756626933813095,
0.06769058108329773,
-0.0189710333943367,
0.037055324763059616,
-0.13061405718326569,
-0.04280674457550049,
0.010780512355268002,
0.0486191026866436,
0.07825807482004166,
0.12267033755779266,
0.08109736442565918,
0.038252025842666626,
0.07978973537683487,
0.01905948854982853,
0.11166026443243027,
-0.05134247988462448,
-0.0758320614695549,
0.07398389279842377,
-0.009818252176046371,
0.04935261979699135,
-0.0021577500738203526,
0.01623678393661976,
-0.07410524785518646,
-0.011744123883545399,
0.04601529985666275,
-0.01179675105959177,
0.0771334320306778,
0.04500391706824303,
-0.06925655901432037,
-0.01723019964993,
-0.033447183668613434,
0.08165533095598221,
0.003736939514055848,
0.00399375706911087,
-0.05629816651344299,
0.04294103384017944,
0.07969990372657776,
-0.025530768558382988,
-0.021034814417362213,
0.035428259521722794,
-0.011258736252784729,
0.015160144306719303,
-0.04751966521143913,
-0.054079364985227585,
0.00567104434594512,
0.020710421726107597,
0.0034413887187838554,
-0.05307698994874954,
0.05124131217598915,
0.022307587787508965,
-0.11106862127780914,
-0.036277711391448975,
3.197621288668798e-34,
0.056636206805706024,
-0.015153191983699799,
-0.07724656909704208,
0.07347861677408218,
0.07564865052700043,
-0.0395781584084034,
0.04286890849471092,
0.056893207132816315,
-0.08644356578588486,
0.032583922147750854,
0.013732977211475372,
-0.0101695004850626,
-0.029567165300250053,
0.09592200070619583,
0.010075482539832592,
-0.02188338339328766,
-0.020281890407204628,
0.007090090773999691,
-0.13712957501411438,
-0.02342892624437809,
-0.05821404233574867,
0.03210853040218353,
0.005693493876606226,
-0.04840681329369545,
0.03871665522456169,
0.024986788630485535,
0.01437756884843111,
-0.03724129498004913,
0.01594165340065956,
0.035268548876047134,
-0.12281718105077744,
-0.011734671890735626,
-0.04296163469552994,
0.050299666821956635,
0.07266923040151596,
-0.0400606133043766,
-0.03286636248230934,
-0.047680214047431946,
0.054879944771528244,
0.011519265361130238,
0.02071583829820156,
0.04041891545057297,
-0.022854812443256378,
-0.015436826273798943,
-0.009167620912194252,
-0.0256164763122797,
0.002047819783911109,
-0.003395683830603957,
0.055717289447784424,
-0.012631416320800781,
0.031610000878572464,
0.04205108433961868,
0.007374370936304331,
-0.03995424881577492,
-0.07488590478897095,
0.015054712072014809,
0.004476575180888176,
0.02650454454123974,
0.026445133611559868,
0.11205039173364639,
-0.036339156329631805,
0.0591425746679306,
-0.013962678611278534,
0.11120937764644623,
0.04193119332194328,
0.014055710285902023,
-0.004775544162839651,
0.07114310562610626,
0.049820732325315475,
0.029039256274700165,
-0.01819201558828354,
0.009235654957592487,
0.022394126281142235,
0.02068815380334854,
0.01250425260514021,
0.0225845780223608,
0.0753859430551529,
-0.03818941488862038,
0.020573168992996216,
-0.06651705503463745,
0.009576261043548584,
-0.013102306984364986,
0.0038758995942771435,
-0.06737055629491806,
-0.0843433290719986,
-0.014559848234057426,
0.07620330899953842,
-0.06706999242305756,
-0.05325350910425186,
0.012432113289833069,
-0.012662994675338268,
-0.15794074535369873,
0.035758744925260544,
-0.07117442786693573,
-0.03131821006536484,
-2.1694201911776336e-33,
-0.005303866229951382,
-0.0008900486282072961,
-0.09507698565721512,
-0.02313131093978882,
-0.15467354655265808,
-0.046009112149477005,
-0.005703836679458618,
-0.027722200378775597,
-0.08836062997579575,
0.009906277991831303,
-0.011727997101843357,
0.02130736969411373,
0.043085336685180664,
-0.04638542979955673,
0.05475926026701927,
-0.010822701267898083,
-0.006193024106323719,
0.01979151740670204,
0.05526166036725044,
0.06034724786877632,
-0.006077907048165798,
0.13763926923274994,
-0.003430778393521905,
-0.06628113240003586,
-0.037844859063625336,
-0.0087368069216609,
-0.0680723786354065,
0.0770789235830307,
0.05458221212029457,
-0.040818095207214355,
-0.01237805001437664,
0.05919594690203667,
0.01693732663989067,
-0.03375326097011566,
-0.06369218230247498,
-0.003837084863334894,
0.05339108407497406,
-0.033250823616981506,
0.008401114493608475,
0.05710489675402641,
0.08800194412469864,
-0.048246223479509354,
-0.04105145111680031,
0.010676887817680836,
-0.04345381259918213,
0.03633352369070053,
-0.05244914069771767,
0.016551217064261436,
0.038010500371456146,
-0.0066240932792425156,
0.012067370116710663,
0.02074514329433441,
0.019546054303646088,
-0.05079483613371849,
-0.049326904118061066,
-0.07752445340156555,
0.014477853663265705,
-0.06264603137969971,
0.04696047306060791,
0.003862101351842284,
0.010271057486534119,
-0.03383457660675049,
0.037384577095508575,
-0.10019191354513168,
0.0173705592751503,
-0.05063548684120178,
-0.03818535432219505,
-0.03234342113137245,
-0.043143048882484436,
-0.04817277565598488,
0.061079200357198715,
-0.06343699991703033,
-0.007271860726177692,
0.018668340519070625,
0.00904588121920824,
-0.03191246837377548,
0.026742447167634964,
-0.06619065254926682,
-0.0419638529419899,
-0.032518982887268066,
0.047977812588214874,
0.02998463064432144,
-0.010867123492062092,
0.07777433842420578,
-0.0004172670596744865,
0.02220672369003296,
0.05274299532175064,
0.030425958335399628,
-0.020071065053343773,
0.007981505244970322,
-0.0589454248547554,
0.028826024383306503,
0.10210977494716644,
-0.015379139222204685,
-0.07004775106906891,
-3.860087005591595e-8,
-0.007436839397996664,
-0.04522423446178436,
0.00014502392150461674,
0.013602515682578087,
0.05109879747033119,
-0.05197368934750557,
-0.12576410174369812,
0.07646359503269196,
-0.03805744647979736,
-0.020118849352002144,
0.0894913375377655,
-0.00854745227843523,
-0.06049695611000061,
-0.06711699068546295,
0.02998178079724312,
0.06697037070989609,
0.0028636299539357424,
-0.0688721239566803,
-0.03929954394698143,
-0.0572529174387455,
-0.017812227830290794,
0.06764111667871475,
0.03563255816698074,
0.019528429955244064,
-0.04108406975865364,
-0.026547642424702644,
-0.007821375504136086,
0.06932424753904343,
0.05569637939333916,
0.013364551588892937,
0.06236723065376282,
0.03878991678357124,
0.007435679901391268,
-0.04359487071633339,
-0.010597413405776024,
0.05658721178770065,
0.05508006736636162,
0.006638512481004,
0.06384126096963882,
-0.012882265262305737,
0.06737960875034332,
-0.019627002999186516,
-0.0348735935986042,
0.0845407247543335,
-0.0180178452283144,
0.0039049778133630753,
-0.03959399834275246,
-0.11079717427492142,
0.03198162466287613,
-0.06233736872673035,
0.038213420659303665,
0.04343150556087494,
-0.05558868125081062,
0.054626163095235825,
-0.05127963051199913,
0.052872661501169205,
-0.042522817850112915,
-0.057445403188467026,
0.045088719576597214,
0.09491129219532013,
0.05705912411212921,
0.07017254084348679,
-0.0030108641367405653,
-0.030760595574975014
] | 0 | 0 | 0 | 0 |
|
Efficient Model-Based Reinforcement Learning Through Optimistic Thompson Sampling | https://openreview.net/forum?id=Ian00SaFHg | [
"Jasmine Bayrooti",
"Carl Henrik Ek",
"Amanda Prorok"
] | Poster | Learning complex robot behavior through interactions with the environment necessitates principled exploration. Effective strategies should prioritize exploring regions of the state-action space that maximize rewards, with optimistic exploration emerging as a promising direction aligned with this idea and enabling sample-efficient reinforcement learning. However, existing methods overlook a crucial aspect: the need for optimism to be informed by a belief connecting the reward and state. To address this, we propose a practical, theoretically grounded approach to optimistic exploration based on Thompson sampling. Our approach is the first that allows for reasoning about _joint_ uncertainty over transitions and rewards for optimistic exploration. We apply our method on a set of MuJoCo and VMAS continuous control tasks. Our experiments demonstrate that optimistic exploration significantly accelerates learning in environments with sparse rewards, action penalties, and difficult-to-explore regions. Furthermore, we provide insights into when optimism is beneficial and emphasize the critical role of model uncertainty in guiding exploration. | reinforcement learning, model-based reinforcement learning, optimistic exploration | We introduce a theoretically-grounded approach to optimistic exploration that leverages joint uncertainty over states and rewards for improved sample efficiency. | 12,371 | 2410.04988 | [
-0.02087421528995037,
-0.03452190011739731,
0.052642107009887695,
0.03730915114283562,
0.037059180438518524,
-0.028377991169691086,
0.018478192389011383,
0.002319095190614462,
0.05571010708808899,
-0.0066703008487820625,
-0.05130293220281601,
-0.02978360280394554,
0.04901828244328499,
0.0373087115585804,
0.03321915119886398,
0.03967664763331413,
0.03225191310048103,
0.04801362752914429,
-0.009146985597908497,
0.020480861887335777,
-0.0015779100358486176,
-0.03973548486828804,
0.0779491513967514,
-0.005318349692970514,
-0.036660339683294296,
-0.019097711890935898,
0.03930291160941124,
-0.05448607727885246,
0.0037660172674804926,
0.002542890142649412,
-0.0029746361542493105,
-0.05368056148290634,
0.036781731992959976,
-0.12207196652889252,
-0.03679259493947029,
0.07034661620855331,
-0.13274045288562775,
-0.057240117341279984,
-0.006833345163613558,
-0.042579833418130875,
0.008488386869430542,
0.01792566291987896,
0.04348605126142502,
0.03853332996368408,
0.02817985601723194,
0.0363396592438221,
0.034628503024578094,
-0.0803929790854454,
0.025805827230215073,
0.009849129244685173,
-0.029368095099925995,
-0.053299423307180405,
0.035021498799324036,
-0.07579314708709717,
-0.023033326491713524,
0.029278643429279327,
-0.0027908976189792156,
0.031602051109075546,
0.05469927191734314,
-0.0450889989733696,
-0.00622508954256773,
-0.08066379278898239,
0.0164646003395319,
-0.028034554794430733,
-0.05713847652077675,
-0.023481041193008423,
-0.01636820286512375,
0.05940360575914383,
-0.006463414523750544,
0.015601346269249916,
0.023844948038458824,
0.07764517515897751,
0.03880380466580391,
-0.01666709966957569,
0.010985967703163624,
0.006113458424806595,
0.01501508615911007,
-0.031008239835500717,
0.004185644444078207,
0.0455191470682621,
0.017046844586730003,
-0.0012066835770383477,
-0.02014756388962269,
-0.0005560417775996029,
0.03875748813152313,
0.013527252711355686,
-0.007003157399594784,
0.029926784336566925,
0.0958625003695488,
-0.03336317092180252,
-0.004779481329023838,
-0.04093589633703232,
-0.053045909851789474,
-0.02396738901734352,
-0.019990095868706703,
0.03568653017282486,
-0.06436306983232498,
-0.08240144699811935,
-0.06809160858392715,
0.05702320486307144,
0.03774835541844368,
0.09215781092643738,
0.0030906086321920156,
-0.004878129344433546,
-0.04771778732538223,
-0.08495806157588959,
-0.019568389281630516,
-0.004060452338308096,
0.01959136314690113,
-0.036443714052438736,
0.041932862251996994,
-0.0825616717338562,
0.10786784440279007,
0.07200486212968826,
0.008014491759240627,
0.05048469454050064,
0.0037775677628815174,
0.04068838059902191,
0.00020550390763673931,
0.09085118770599365,
0.057129815220832825,
-0.04865645244717598,
0.0812726765871048,
0.018046297132968903,
0.048276156187057495,
-0.05037330090999603,
0.04622392728924751,
-7.419554579319714e-34,
0.10288114100694656,
-0.043327104300260544,
0.07514968514442444,
0.0402785986661911,
0.020613539963960648,
0.014943636953830719,
0.06062129884958267,
-0.02400672622025013,
-0.11023122072219849,
0.016700221225619316,
0.03953962400555611,
0.017837507650256157,
-0.04894522204995155,
0.06469327956438065,
0.01903979480266571,
-0.02988053858280182,
-0.03125495836138725,
0.04107894003391266,
0.004583607893437147,
0.002506251446902752,
0.04124991223216057,
-0.011000992730259895,
-0.04303966835141182,
-0.07110758870840073,
0.06069726496934891,
0.02229832485318184,
-0.02366797998547554,
0.056156598031520844,
-0.03969099745154381,
-0.0019052467541769147,
-0.10008645802736282,
0.05346628278493881,
-0.07749618589878082,
0.06383132189512253,
0.006593906786292791,
0.08842366933822632,
-0.06876440346240997,
-0.007755827624350786,
-0.029911518096923828,
-0.06421153247356415,
0.029941655695438385,
0.015755798667669296,
0.051245205104351044,
0.015073949471116066,
-0.0780530720949173,
-0.10135708749294281,
0.019373638555407524,
-0.03713386505842209,
-0.08329270035028458,
0.002550331875681877,
-0.0064699845388531685,
0.02985967881977558,
0.014769569039344788,
-0.05225065350532532,
-0.04401590675115585,
-0.042138513177633286,
0.0249025858938694,
0.054431408643722534,
-0.08741399645805359,
0.0029265934135764837,
0.039642054587602615,
-0.05535265803337097,
-0.04875325411558151,
0.00015828624600544572,
-0.015898192301392555,
0.13912850618362427,
-0.03583171218633652,
0.062257811427116394,
0.07594169676303864,
0.048287712037563324,
0.004270634613931179,
0.009377222508192062,
-0.036339666694402695,
-0.11947042495012283,
0.061459023505449295,
-0.0018144265050068498,
0.02880171500146389,
-0.014542336575686932,
0.023504694923758507,
-0.02404586412012577,
0.011077051050961018,
-0.09017923474311829,
-0.13097800314426422,
0.026656033471226692,
0.011634950526058674,
-0.03307698667049408,
0.01582976058125496,
-0.05493508279323578,
-0.08838368952274323,
0.0191512368619442,
-0.11178923398256302,
-0.025769317522644997,
-0.020567504689097404,
0.08530272543430328,
0.006096405442804098,
-5.471265446682892e-34,
0.07121574133634567,
-0.08486858755350113,
0.06942562013864517,
-0.0152015071362257,
-0.055623140186071396,
-0.02649299055337906,
0.006556304637342691,
-0.05584857985377312,
-0.032308466732501984,
-0.09446550160646439,
-0.1257878541946411,
0.03352198004722595,
0.0343487411737442,
-0.006558244116604328,
-0.021749090403318405,
0.014901977963745594,
0.023152843117713928,
-0.010954839177429676,
0.026690714061260223,
-0.009203103370964527,
0.015370199456810951,
0.0704583004117012,
-0.085816390812397,
0.013805577531456947,
-0.029087383300065994,
0.060988787561655045,
0.016486413776874542,
0.02272222936153412,
0.0022153554018586874,
0.041725583374500275,
-0.10825808346271515,
-0.1109757125377655,
0.001421123743057251,
0.0022047872189432383,
-0.08410363644361496,
0.050543490797281265,
0.10423565655946732,
0.020008066669106483,
0.00940291490405798,
0.0682184174656868,
0.020203126594424248,
-0.1017337292432785,
-0.006487653125077486,
0.010751006193459034,
0.03359237685799599,
-0.019649246707558632,
0.010888233780860901,
0.017720578238368034,
-0.06565187126398087,
0.027197573333978653,
-0.021468201652169228,
0.05988135188817978,
-0.10911446809768677,
-0.016450637951493263,
-0.05738699808716774,
-0.07291725277900696,
-0.07189697027206421,
0.010687926784157753,
0.06501522660255432,
0.01126786321401596,
-0.07954613119363785,
0.05765090510249138,
-0.042087484151124954,
0.06888055801391602,
-0.0659404918551445,
0.05320320278406143,
-0.01869884319603443,
0.030855918303132057,
0.06837963312864304,
-0.03194713592529297,
0.01782447099685669,
0.06170118600130081,
0.05952541157603264,
-0.018137505277991295,
0.07083123922348022,
-0.017412222921848297,
0.03716128319501877,
-0.03205016627907753,
0.05424480140209198,
-0.0056699300184845924,
0.020990943536162376,
0.03993526101112366,
0.07914460450410843,
0.0846790000796318,
0.03297284618020058,
0.10380823910236359,
-0.033196821808815,
0.08630795031785965,
0.021580683067440987,
-0.005711146164685488,
-0.0229426771402359,
-0.01262502558529377,
0.05644325911998749,
0.028822502121329308,
-0.06718194484710693,
-3.711236118419947e-8,
-0.10508348792791367,
0.06063898652791977,
0.05543678626418114,
0.024783791974186897,
0.07542712986469269,
0.04695184901356697,
0.007328698877245188,
-0.045297738164663315,
-0.06937863677740097,
0.0021750363521277905,
0.08347941190004349,
0.057765960693359375,
0.0076894196681678295,
-0.08176758140325546,
-0.0017658524448052049,
0.042637698352336884,
0.0622587613761425,
-0.05429985746741295,
-0.002060395898297429,
0.0020791024435311556,
0.04906687140464783,
-0.043607521802186966,
0.026754170656204224,
-0.03285594657063484,
0.01444519404321909,
-0.06235847249627113,
-0.06308460980653763,
0.06382740288972855,
0.05611851438879967,
0.04057469591498375,
-0.020978964865207672,
-0.04823393374681473,
0.005254992749541998,
0.06021948531270027,
0.03877371549606323,
0.04095758870244026,
0.01858554035425186,
-0.08036217093467712,
0.024488314986228943,
-0.07935502380132675,
0.006986232940107584,
0.11089745163917542,
-0.022669987753033638,
-0.012788170017302036,
-0.09063772857189178,
0.05565464124083519,
-0.07939651608467102,
-0.08053997904062271,
0.028926445171236992,
-0.014994329772889614,
-0.0030421719420701265,
-0.013254842720925808,
-0.07393467426300049,
0.11121903359889984,
0.03986343368887901,
-0.003960497677326202,
0.012564009055495262,
-0.05767723545432091,
-0.012176834046840668,
0.0904143825173378,
0.06686855852603912,
0.0013505020178854465,
-0.08139809221029282,
-0.005345084238797426
] | 0 | 0 | 0 | 0 |
|
Latent Safety-Constrained Policy Approach for Safe Offline Reinforcement Learning | https://openreview.net/forum?id=bDt5qc7TfO | [
"Prajwal Koirala",
"Zhanhong Jiang",
"Soumik Sarkar",
"Cody Fleming"
] | Poster | In safe offline reinforcement learning, the objective is to develop a policy that maximizes cumulative rewards while strictly adhering to safety constraints, utilizing only offline data. Traditional methods often face difficulties in balancing these constraints, leading to either diminished performance or increased safety risks. We address these issues with a novel approach that begins by learning a conservatively safe policy through the use of Conditional Variational Autoencoders, which model the latent safety constraints. Subsequently, we frame this as a Constrained Reward-Return Maximization problem, wherein the policy aims to optimize rewards while complying with the inferred latent safety constraints. This is achieved by training an encoder with a reward-Advantage Weighted Regression objective within the latent constraint space. Our methodology is supported by theoretical analysis, including bounds on policy performance and sample complexity. Extensive empirical evaluation on benchmark datasets, including challenging autonomous driving scenarios, demonstrates that our approach not only maintains safety compliance but also excels in cumulative reward optimization, surpassing existing methods. Additional visualizations provide further insights into the effectiveness and underlying mechanisms of our approach. | Safe RL, Offline RL, Variational Autoencoders, Latent Safety Constraints | This study proposes learning safety constraints in the latent space for safe offline RL, simplifying constraint imposition within the Constrained MDP framework. | 12,366 | 2412.08794 | [
-0.05650906637310982,
-0.08760521560907364,
-0.020883221179246902,
-0.016850905492901802,
0.10770038515329361,
0.10032051801681519,
-0.0051726908423006535,
-0.010989150032401085,
-0.049396395683288574,
0.04960966482758522,
-0.009173095226287842,
-0.059601254761219025,
0.05469755828380585,
0.0623541884124279,
-0.05585067346692085,
0.03741820156574249,
0.09643455594778061,
0.041528478264808655,
-0.14519478380680084,
-0.02336503006517887,
0.02389627695083618,
0.0671197697520256,
0.05016582831740379,
0.044464390724897385,
-0.010944444686174393,
-0.04294819012284279,
0.057803183794021606,
0.02700495533645153,
-0.002741545671597123,
-0.019085342064499855,
0.03168215602636337,
-0.10840900242328644,
-0.010636660270392895,
-0.041337694972753525,
-0.040939606726169586,
0.008500796742737293,
-0.15478380024433136,
-0.06807226687669754,
-0.0647771805524826,
-0.006200016010552645,
-0.09651177376508713,
-0.005810013972222805,
-0.030954528599977493,
0.04549325257539749,
0.04975952208042145,
-0.03510294482111931,
0.009973060339689255,
-0.05206013470888138,
0.031037945300340652,
-0.059152889996767044,
0.03436161205172539,
0.05056394264101982,
0.028412247076630592,
-0.006001744419336319,
-0.061460427939891815,
-0.03780306875705719,
0.07728522270917892,
0.006386605557054281,
-0.018153803423047066,
0.025959542021155357,
0.009634836576879025,
-0.16154751181602478,
-0.02496846206486225,
-0.08346879482269287,
-0.04719207435846329,
-0.05099547281861305,
-0.06121066212654114,
0.01658038981258869,
0.019636521115899086,
0.06408534198999405,
-0.04324435442686081,
0.049289897084236145,
-0.02821504697203636,
-0.019511286169290543,
0.04292808473110199,
0.04367786645889282,
0.057696156203746796,
-0.04682046175003052,
0.017917443066835403,
-0.00380897824652493,
0.006873972713947296,
0.0019220128888264298,
0.03247220814228058,
0.10961786657571793,
0.10896506160497665,
-0.012002131901681423,
-0.02416602335870266,
0.09512334316968918,
0.10763558745384216,
0.02896757982671261,
-0.027262071147561073,
0.01390792615711689,
0.05595765262842178,
0.055669717490673065,
0.025292351841926575,
-0.047326765954494476,
-0.005137710832059383,
-0.1067601665854454,
-0.012018677778542042,
0.06517981737852097,
0.011840823106467724,
0.05151456966996193,
-0.008223336189985275,
0.04437921568751335,
0.06038982793688774,
-0.054001159965991974,
0.07917752861976624,
0.043586745858192444,
0.04891600087285042,
-0.06786342710256577,
0.10616995394229889,
0.026080431416630745,
0.07577157765626907,
0.0005837759817950428,
-0.024281272664666176,
0.001923246425576508,
-0.025507144629955292,
0.01491202600300312,
0.022926408797502518,
0.0760408565402031,
-0.06281067430973053,
-0.043140728026628494,
0.04394039884209633,
0.002178110647946596,
0.048581596463918686,
-0.0684535801410675,
0.0052619818598032,
1.779240188019115e-33,
0.014399333856999874,
0.006648343522101641,
0.061422888189554214,
-0.0247778482735157,
0.025273514911532402,
-0.030237141996622086,
-0.06543941050767899,
0.006250650621950626,
-0.06132325157523155,
-0.005226365756243467,
0.013446875847876072,
-0.003972181119024754,
-0.09197571128606796,
0.06401664763689041,
0.10022135823965073,
-0.03765226528048515,
-0.05534392595291138,
0.04147014021873474,
0.02857036329805851,
-0.022155389189720154,
0.03314448148012161,
-0.1047729030251503,
-0.04443062096834183,
0.0052892891690135,
0.01416734792292118,
0.0023882556706666946,
0.03003040887415409,
-0.039360951632261276,
0.008904651738703251,
-0.014192699454724789,
-0.12666350603103638,
0.044507429003715515,
0.008053005672991276,
0.01297350600361824,
0.0029490229208022356,
0.0515364408493042,
-0.10677177459001541,
-0.019441397860646248,
-0.023381976410746574,
0.025549830868840218,
0.017322003841400146,
-0.01511119119822979,
0.06978970021009445,
-0.015721186995506287,
-0.02139287069439888,
-0.12540343403816223,
0.03819958120584488,
0.0015514991246163845,
-0.05456067994236946,
-0.05275687202811241,
-0.0278305821120739,
0.01285913772881031,
0.034314289689064026,
-0.09253328293561935,
-0.09627863019704819,
0.05348601192235947,
0.01986686885356903,
0.05483066663146019,
0.0016480577178299427,
0.02876584418118,
-0.0002163492317777127,
-0.014704885892570019,
0.055211082100868225,
0.007874267175793648,
-0.03555718809366226,
0.043514955788850784,
-0.013845563866198063,
-0.011498527601361275,
0.03651529923081398,
0.044958025217056274,
-0.04356754198670387,
0.08277929574251175,
0.01005855668336153,
-0.10070157796144485,
0.04971768707036972,
0.08012259006500244,
0.006282128859311342,
-0.07127774506807327,
0.044274263083934784,
-0.03685522824525833,
-0.0024354313500225544,
0.04529862850904465,
-0.023678872734308243,
0.036258820444345474,
-0.014899409376084805,
-0.06095500290393829,
-0.04723990708589554,
-0.0325174555182457,
-0.04294639825820923,
0.06673026829957962,
-0.09943998605012894,
-0.09186884760856628,
-0.0008404164691455662,
0.06339471787214279,
0.024215832352638245,
-2.1075695355282035e-33,
-0.0035512347240000963,
-0.047162145376205444,
0.007741347420960665,
-0.019475581124424934,
0.029563382267951965,
-0.021516086533665657,
-0.01961255818605423,
0.020451856777071953,
0.042482536286115646,
0.01224357821047306,
-0.04215673729777336,
-0.0008928435272537172,
0.06717804819345474,
-0.0024665785022079945,
0.02011863887310028,
-0.03946301341056824,
-0.0363994762301445,
0.1126575618982315,
0.011340866796672344,
-0.011537608690559864,
0.0950959175825119,
0.018113678321242332,
-0.061137694865465164,
0.091854989528656,
0.001485421322286129,
0.00016498086915817112,
-0.028507743030786514,
0.02400357276201248,
-0.02014615386724472,
0.08332948386669159,
0.004221860319375992,
-0.04846997559070587,
-0.020531130954623222,
0.0022644337732344866,
-0.04725552722811699,
0.041925638914108276,
0.07981505244970322,
-0.03845275565981865,
-0.0699695274233818,
0.05194195359945297,
0.047370728105306625,
-0.014038275927305222,
-0.0010420463513582945,
-0.018139159306883812,
0.010237461887300014,
-0.031630825251340866,
0.01616784743964672,
-0.006677455268800259,
0.027781402692198753,
0.03957784175872803,
-0.02682972326874733,
-0.004688600543886423,
-0.06451627612113953,
0.02760365791618824,
-0.03851230815052986,
-0.07519069314002991,
0.010117709636688232,
-0.0452917255461216,
0.09056025743484497,
0.025182368233799934,
-0.0507974848151207,
0.024958735331892967,
-0.08218351006507874,
0.0384163036942482,
-0.05327944457530975,
-0.09979894012212753,
-0.05516272410750389,
-0.013391472399234772,
0.03895123675465584,
0.014043096452951431,
0.0659434124827385,
0.03428935259580612,
0.029019972309470177,
0.0023588864132761955,
0.038989998400211334,
-0.03781064227223396,
0.05120097100734711,
0.00897490605711937,
-0.024773944169282913,
-0.04680120572447777,
0.054369665682315826,
-0.004091551993042231,
0.09736524522304535,
0.10190410912036896,
0.01988869532942772,
0.06247714161872864,
0.01584579050540924,
-0.00043371564242988825,
0.029208963736891747,
0.02023942396044731,
-0.05616739019751549,
0.023990284651517868,
-0.042283207178115845,
0.02880929782986641,
-0.11067453771829605,
-3.80039573144586e-8,
-0.03531423956155777,
0.06856219470500946,
0.020336104556918144,
-0.055532798171043396,
0.010089325718581676,
-0.018770525231957436,
0.047270067036151886,
-0.011253132484853268,
-0.057543378323316574,
0.0006683756364509463,
0.13394993543624878,
0.03347339481115341,
0.010506615042686462,
-0.09254685789346695,
-0.0981496199965477,
0.0785880759358406,
0.03529742360115051,
-0.02828799933195114,
0.04337570071220398,
0.04874517396092415,
0.014940488152205944,
0.038598377257585526,
-0.013435502536594868,
-0.07589977234601974,
0.0496521070599556,
-0.051615528762340546,
-0.026790643110871315,
0.11304547637701035,
0.05159216746687889,
0.009944899939000607,
-0.003965084906667471,
-0.02127878926694393,
0.020863791927695274,
0.021372783929109573,
0.011939718388020992,
0.028706157580018044,
0.023193728178739548,
-0.03759315237402916,
-0.020307833328843117,
-0.023267891258001328,
0.0520796999335289,
0.04757020249962807,
0.014819972217082977,
-0.00008089119364740327,
0.003700931090861559,
0.05879679694771767,
-0.029008259996771812,
-0.06440836936235428,
0.04735823720693588,
-0.0567118376493454,
0.003544250503182411,
-0.057665612548589706,
0.009402367286384106,
0.11941023916006088,
0.09251350909471512,
-0.007036433555185795,
0.0013871730770915747,
-0.09871815890073776,
0.03701777383685112,
0.07753404974937439,
-0.026758573949337006,
0.026489347219467163,
-0.02363721840083599,
-0.023002298548817635
] | 0 | 0 | 0 | 0 |
|
Large (Vision) Language Models are Unsupervised In-Context Learners | https://openreview.net/forum?id=ohJxgRLlLt | [
"Artyom Gadetsky",
"Andrei Atanov",
"Yulun Jiang",
"Zhitong Gao",
"Ghazal Hosseini Mighan",
"Amir Zamir",
"Maria Brbic"
] | Poster | Recent advances in large language and vision-language models have enabled zero-shot inference, allowing models to solve new tasks without task-specific training. Various adaptation techniques such as prompt engineering, In-Context Learning (ICL), and supervised fine-tuning can further enhance the model’s performance on a downstream task, but they require substantial manual effort to construct effective prompts or labeled examples. In this work, we introduce a joint inference framework for fully unsupervised adaptation, eliminating the need for manual prompt engineering and labeled examples. Unlike zero-shot inference, which makes independent predictions, the joint inference makes predictions simultaneously for all inputs in a given task. Since direct joint inference involves computationally expensive optimization, we develop efficient approximation techniques, leading to two unsupervised adaptation methods: unsupervised fine-tuning and unsupervised ICL. We demonstrate the effectiveness of our methods across diverse tasks and models, including language-only Llama-3.1 on natural language processing tasks, reasoning-oriented Qwen2.5-Math on grade school math problems, vision-language OpenFlamingo on vision tasks, and the API-only access GPT-4o model on massive multi-discipline tasks. Our experiments demonstrate substantial improvements over the standard zero-shot approach, including 39% absolute improvement on the challenging GSM8K math reasoning dataset. Remarkably, despite being fully unsupervised, our framework often performs on par with supervised approaches that rely on ground truth labels. | llm, unsupervised, in-context learning, few-shot learning | We introduce a joint inference framework for large (vision) language models to perform unsupervised adaptation on a given task, resulting in the improved performance upon independent zero-shot predictions. | 12,362 | null | [
-0.04977644979953766,
-0.0310652032494545,
0.0232499148696661,
0.02846706099808216,
0.1412629336118698,
-0.029579808935523033,
0.03136465325951576,
-0.017776357010006905,
0.005866254214197397,
-0.0018163742497563362,
-0.042073823511600494,
-0.11991042643785477,
0.03848780691623688,
0.08771637827157974,
0.02165665291249752,
0.01332910731434822,
0.10102574527263641,
0.05074959993362427,
-0.1319318413734436,
-0.07265282422304153,
0.03020363859832287,
0.009474839083850384,
0.08706481009721756,
0.01192521583288908,
0.06510960310697556,
0.018647048622369766,
0.016271192580461502,
-0.052734144032001495,
0.004070692230015993,
-0.0288207046687603,
-0.012828145176172256,
-0.0053631216287612915,
-0.0763847753405571,
0.03733302652835846,
-0.051664937287569046,
0.07449733465909958,
-0.03798173740506172,
0.06604362279176712,
0.0018344313139095902,
0.011365896090865135,
-0.02108764462172985,
0.033293984830379486,
0.01573677361011505,
-0.0167426448315382,
0.06278633326292038,
0.010483039543032646,
-0.016499944031238556,
-0.02343841828405857,
-0.0012776292860507965,
-0.0034319176338613033,
-0.09060432016849518,
-0.05775771290063858,
-0.03222489356994629,
0.03269050270318985,
-0.03942863643169403,
0.0270170196890831,
0.0130160478875041,
-0.01666276715695858,
0.01656203158199787,
-0.022372834384441376,
-0.07609740644693375,
-0.11597110331058502,
-0.02052975632250309,
0.019609646871685982,
-0.06174306198954582,
-0.033099863678216934,
0.0034773482475429773,
-0.0006850334466435015,
-0.026378236711025238,
0.038210272789001465,
0.002333408687263727,
0.07361661642789841,
0.014939791522920132,
0.05439390987157822,
0.04954616725444794,
0.03286674618721008,
0.04593086987733841,
-0.030648859217762947,
0.07087089121341705,
-0.060571327805519104,
0.00628221221268177,
-0.03675447404384613,
-0.006301097571849823,
-0.060831617563962936,
0.1856386363506317,
0.06857921183109283,
-0.021597974002361298,
0.06872070580720901,
0.011953214183449745,
0.018928509205579758,
-0.07745134830474854,
-0.09684711694717407,
0.03155072405934334,
0.05038921907544136,
0.03179149702191353,
0.016460983082652092,
-0.005335578694939613,
-0.10635365545749664,
-0.08012022078037262,
0.11293406784534454,
0.012491261586546898,
0.06619460135698318,
0.02503920905292034,
-0.07906802743673325,
-0.0007068222621455789,
-0.00301903928630054,
0.0441230908036232,
0.049489010125398636,
0.01673736795783043,
-0.12808459997177124,
0.08058087527751923,
0.028129642829298973,
0.035633672028779984,
0.0014590987702831626,
-0.02786353975534439,
-0.025449547916650772,
0.0429261177778244,
-0.06125682219862938,
0.03565199673175812,
0.0004079819191247225,
-0.021329978480935097,
0.027205834165215492,
-0.01726263388991356,
-0.008441093377768993,
0.002264553913846612,
-0.0850878432393074,
-0.024185607209801674,
1.1204397973950372e-33,
0.05959421768784523,
0.04244522377848625,
0.10244123637676239,
0.005902871023863554,
0.05380546674132347,
-0.01676095463335514,
0.013363066129386425,
0.07166463136672974,
-0.040731705725193024,
-0.00665636220946908,
0.04844755306839943,
-0.012767503969371319,
-0.06422019004821777,
0.058640219271183014,
0.050749488174915314,
0.006694959010928869,
-0.02082965523004532,
0.04298819974064827,
-0.03308423236012459,
0.007530174218118191,
-0.00854283757507801,
0.01658732257783413,
-0.0704333633184433,
-0.0195954367518425,
0.028089744970202446,
0.03624198958277702,
0.0696149542927742,
-0.07272124290466309,
-0.008857477456331253,
0.014530245214700699,
-0.11897630244493484,
0.009309966117143631,
-0.024951880797743797,
0.04772680252790451,
-0.0067369285970926285,
-0.08114228397607803,
-0.02571873366832733,
-0.03056434355676174,
0.00011900420940946788,
-0.004816614091396332,
0.014155355282127857,
0.03323615342378616,
0.052275896072387695,
-0.0024100386071950197,
-0.06555625796318054,
-0.1053713709115982,
-0.014016376808285713,
0.08034714311361313,
-0.06055701524019241,
-0.005147214513272047,
0.07082904875278473,
0.0234267245978117,
-0.043012432754039764,
-0.065766341984272,
-0.0664554312825203,
0.016287632286548615,
0.04136386141180992,
0.050063375383615494,
0.01872590184211731,
-0.011811262927949429,
0.016153916716575623,
-0.013548154383897781,
-0.051180124282836914,
0.08332693576812744,
-0.011960391886532307,
0.04261607676744461,
-0.02117680199444294,
0.039721693843603134,
0.0581478625535965,
-0.0010220458498224616,
-0.03935055807232857,
0.012303424067795277,
-0.031732454895973206,
-0.02108481340110302,
0.09933896362781525,
0.06361746042966843,
0.036511339247226715,
-0.09615131467580795,
0.029845893383026123,
0.09158273041248322,
0.07387445867061615,
-0.015931997448205948,
0.004057606216520071,
0.008198714815080166,
-0.04592593386769295,
-0.02522176317870617,
0.06479757279157639,
-0.127842515707016,
-0.009737231768667698,
-0.05610065534710884,
-0.02202041633427143,
-0.018127845600247383,
-0.058217115700244904,
-0.03592000901699066,
0.02175740711390972,
-1.580063001765596e-35,
0.031683243811130524,
0.011327295564115047,
-0.03772971034049988,
-0.003589766565710306,
-0.03901968151330948,
-0.030155271291732788,
0.05489759147167206,
0.02407628856599331,
0.025128837674856186,
-0.12266465276479721,
-0.05482969060540199,
-0.03207450732588768,
0.026852190494537354,
0.007968622259795666,
-0.011378416791558266,
-0.012856673449277878,
-0.03970134258270264,
0.06673897802829742,
-0.05206576734781265,
0.041776206344366074,
-0.0036355373449623585,
0.1143791452050209,
-0.03785994276404381,
0.006969605106860399,
-0.0518006756901741,
0.04981258884072304,
-0.046061452478170395,
0.04697543382644653,
-0.03559102118015289,
0.06248075142502785,
-0.045794494450092316,
-0.02704940177500248,
-0.03777396306395531,
-0.061745136976242065,
-0.09196135401725769,
0.04806998744606972,
0.014374909922480583,
-0.034187983721494675,
-0.06918410956859589,
0.10084723681211472,
0.09075479209423065,
-0.049786701798439026,
-0.03153593838214874,
0.032714612782001495,
-0.04472591355443001,
0.01365472748875618,
-0.04104353487491608,
-0.046831998974084854,
-0.02242017351090908,
0.08301250636577606,
-0.09974643588066101,
0.04587847366929054,
-0.07437053322792053,
0.02932676300406456,
-0.027461545541882515,
-0.11712770164012909,
-0.01054202951490879,
-0.05886198207736015,
0.09689049422740936,
-0.030315488576889038,
-0.0110903549939394,
0.028322875499725342,
0.016303764656186104,
0.02576962113380432,
0.04190575331449509,
-0.029724592342972755,
-0.06972653418779373,
-0.002753831446170807,
0.048660870641469955,
-0.008326048962771893,
0.052821122109889984,
-0.009725773707032204,
0.004080958664417267,
-0.027386760339140892,
-0.015817344188690186,
0.028161879628896713,
0.08015614748001099,
-0.018349269405007362,
0.003674162086099386,
-0.0254196934401989,
0.08736340701580048,
-0.05393745005130768,
0.004154805559664965,
0.08352826535701752,
0.04677020013332367,
0.060464560985565186,
0.12044204771518707,
0.053243525326251984,
0.05112785845994949,
0.036625660955905914,
-0.06775207817554474,
0.03866958245635033,
0.0420784056186676,
0.06132916361093521,
-0.06302420794963837,
-3.6728525998341865e-8,
-0.10428208112716675,
0.05783748999238014,
-0.02533229999244213,
0.05406980961561203,
-0.01903463713824749,
0.008881318382918835,
-0.08916457742452621,
0.07543694972991943,
0.0275114756077528,
-0.01206126157194376,
0.01700575090944767,
0.02594858966767788,
0.004054608289152384,
-0.06089169904589653,
-0.04611068591475487,
0.10086426883935928,
0.04265002906322479,
0.0300239659845829,
0.014370969496667385,
-0.03399379923939705,
0.015247049741446972,
0.03330144286155701,
0.02878616377711296,
0.02100173942744732,
0.033423516899347305,
-0.09526524692773819,
-0.01830385997891426,
0.10119719803333282,
0.09675759077072144,
-0.02739807218313217,
-0.07795961201190948,
0.00940600037574768,
0.010890915989875793,
-0.04756847023963928,
0.015652049332857132,
0.05867936462163925,
0.013309893198311329,
-0.009285986423492432,
0.00988770741969347,
-0.07398894429206848,
-0.017228512093424797,
0.07178974896669388,
-0.06229972839355469,
0.038299571722745895,
0.03668438643217087,
0.00011426945275161415,
0.020565645769238472,
-0.11827420443296432,
0.012203994207084179,
0.000697673240210861,
-0.012938886880874634,
-0.005386112257838249,
-0.0004585432179737836,
0.020060034468770027,
-0.013173247687518597,
0.02645907737314701,
-0.023517847061157227,
-0.022180601954460144,
-0.008420794270932674,
0.06080270931124687,
0.005069575272500515,
0.08215626329183578,
-0.04075959324836731,
0.02428647130727768
] | 0 | 0 | 0 | 0 |
|
LLMs Can Plan Only If We Tell Them | https://openreview.net/forum?id=K3KrOsR6y9 | [
"Bilgehan Sel",
"Ruoxi Jia",
"Ming Jin"
] | Poster | Large language models (LLMs) have demonstrated significant capabilities in natural language processing and reasoning, yet their effectiveness in autonomous planning has been under debate. While existing studies have utilized LLMs with external feedback mechanisms or in controlled environments for planning, these approaches often involve substantial computational and development resources due to the requirement for careful design and iterative backprompting. Moreover, even the most advanced LLMs like GPT-4 struggle to match human performance on standard planning benchmarks, such as the Blocksworld, without additional support. This paper investigates whether LLMs can independently generate long-horizon plans that rival human baselines. Our novel enhancements to Algorithm-of-Thoughts (AoT), which we dub AoT+, help achieve state-of-the-art results in planning benchmarks out-competing prior methods and human baselines all autonomously. | large language models, decision-making, planning | Investigating what enables autonomous planning in large language models | 12,359 | 2501.13545 | [
0.019114378839731216,
-0.04177418723702431,
0.06409993022680283,
-0.022773196920752525,
0.050077248364686966,
-0.05842002481222153,
-0.0577605776488781,
0.04804656282067299,
0.0507042333483696,
0.010213333182036877,
-0.10064668208360672,
-0.009543176740407944,
0.01900104060769081,
0.03861143812537193,
0.052303165197372437,
0.02951059117913246,
0.07913157343864441,
0.00483661238104105,
-0.005711347330361605,
-0.056569650769233704,
-0.019702736288309097,
0.03580556809902191,
0.05231412500143051,
-0.029103538021445274,
0.017242878675460815,
0.00692548556253314,
0.0011203299509361386,
-0.02755000814795494,
0.0507199764251709,
0.06782928854227066,
0.054323259741067886,
0.0058119394816458225,
0.0009215656318701804,
-0.018337685614824295,
-0.036771442741155624,
0.05002867802977562,
-0.11263791471719742,
0.019355114549398422,
-0.04338156431913376,
-0.06662186235189438,
-0.08303380012512207,
-0.028416501358151436,
0.012905856594443321,
0.023472795262932777,
0.06937345117330551,
0.041066329926252365,
-0.00035365988151170313,
0.009211031720042229,
-0.05293498560786247,
0.018724022433161736,
-0.11246442794799805,
-0.021448923274874687,
0.024044601246714592,
-0.0640389546751976,
-0.019274910911917686,
0.03251110389828682,
0.049540579319000244,
0.04740751162171364,
0.04660659655928612,
-0.07636108994483948,
-0.05218875780701637,
-0.13469719886779785,
-0.026131194084882736,
-0.05172569677233696,
-0.02743041142821312,
0.01237565465271473,
0.002600350184366107,
0.06634168326854706,
-0.04554228484630585,
0.024386728182435036,
0.013521552085876465,
0.062338776886463165,
0.031124848872423172,
0.03918398916721344,
0.030947839841246605,
0.09341083467006683,
-0.01439073495566845,
-0.034598398953676224,
0.05230502039194107,
-0.04317827522754669,
-0.002052459167316556,
0.014957677572965622,
-0.006412202026695013,
0.067559614777565,
0.10917863994836807,
0.0034505699295550585,
0.0196140818297863,
0.09806375205516815,
0.05199359729886055,
0.02503141574561596,
-0.012918474152684212,
-0.16257420182228088,
0.01450288575142622,
-0.029482241719961166,
-0.02659459039568901,
0.0885363519191742,
-0.10049626976251602,
-0.11219251900911331,
-0.045870549976825714,
0.057167865335941315,
0.011090961284935474,
0.07875996083021164,
-0.04220473766326904,
-0.04760414734482765,
-0.013660263270139694,
-0.010054015554487705,
0.05880233272910118,
0.030356071889400482,
0.0038270195946097374,
-0.09968365728855133,
0.027445560321211815,
-0.016170676797628403,
0.08546140044927597,
0.024778004735708237,
0.01840350776910782,
-0.037662480026483536,
0.057760197669267654,
-0.023615093901753426,
0.038982123136520386,
0.025295646861195564,
-0.010956007987260818,
0.038671284914016724,
0.08208130300045013,
-0.0038718313444405794,
0.014707650989294052,
-0.0405094176530838,
0.024437522515654564,
1.6781081845850502e-33,
0.07711777836084366,
0.004882404580712318,
0.0925961509346962,
0.05225237458944321,
0.08500248193740845,
-0.009747766889631748,
-0.011541180312633514,
0.017547497525811195,
-0.06142213195562363,
0.05075066536664963,
0.015077712945640087,
-0.10222090780735016,
-0.021453306078910828,
0.06840135157108307,
0.05339835211634636,
-0.0485360324382782,
-0.004367280285805464,
0.04115446284413338,
-0.09523986279964447,
-0.01516296062618494,
0.056416381150484085,
0.04104063659906387,
0.0623960942029953,
-0.08548370003700256,
0.0424402691423893,
0.0645991787314415,
0.01960070990025997,
-0.0586424320936203,
0.017349055036902428,
-0.013078746385872364,
-0.1393090784549713,
0.005261055193841457,
-0.050576336681842804,
0.09609241783618927,
0.007614769972860813,
0.03233065456151962,
-0.04045039415359497,
-0.05267589911818504,
0.047087412327528,
-0.007057960610836744,
-0.004468746483325958,
0.08521072566509247,
0.0492430105805397,
0.02203195169568062,
-0.04937827214598656,
-0.02157985046505928,
-0.011677748523652554,
0.029484201222658157,
-0.032403185963630676,
0.01280900090932846,
0.05223468691110611,
0.04398917779326439,
-0.07231779396533966,
-0.05223622918128967,
-0.0027699864003807306,
-0.08174225687980652,
-0.02664707601070404,
0.03820542246103287,
0.0360536091029644,
0.06714679300785065,
0.06338061392307281,
0.013708454556763172,
-0.042271386831998825,
0.04410632699728012,
0.033048201352357864,
0.070430226624012,
-0.03648475557565689,
0.060981981456279755,
0.10986805707216263,
-0.0419364757835865,
0.022143825888633728,
0.036056987941265106,
0.022327659651637077,
-0.025178352370858192,
0.001742557273246348,
-0.01351746916770935,
0.042889367789030075,
-0.05550125613808632,
-0.00020172560471110046,
0.0019264634465798736,
0.03943875432014465,
-0.00289970263838768,
-0.04934454709291458,
-0.07881046831607819,
0.01818331889808178,
-0.08155225962400436,
0.03629186749458313,
-0.03653078153729439,
-0.025722132995724678,
-0.01705089397728443,
-0.12125103175640106,
-0.031259290874004364,
-0.027846135199069977,
0.010038347914814949,
-0.06021831929683685,
-2.0311157502927066e-33,
-0.056858863681554794,
-0.08184128999710083,
0.041330382227897644,
0.015741797164082527,
-0.021394740790128708,
-0.12790557742118835,
-0.02151462808251381,
-0.05079866573214531,
-0.014084112830460072,
-0.05861696973443031,
-0.0749208927154541,
-0.04625094681978226,
0.07784334570169449,
0.03939288482069969,
0.04199012741446495,
-0.046398770064115524,
-0.00956856831908226,
-0.07103494554758072,
0.014670250937342644,
0.04582732915878296,
-0.04344848543405533,
0.04272228851914406,
-0.14733265340328217,
0.041010841727256775,
-0.01099261175841093,
0.044211383908987045,
-0.0013362823519855738,
0.02337527461349964,
-0.021682659164071083,
-0.036635540425777435,
-0.10010804235935211,
-0.025834782049059868,
-0.08835232257843018,
-0.018555855378508568,
-0.041157349944114685,
0.048358429223299026,
-0.017181580886244774,
0.019298115745186806,
-0.08922994881868362,
0.04763138294219971,
0.04895404726266861,
-0.03855060786008835,
-0.02835896983742714,
-0.0247235968708992,
-0.07279624789953232,
0.05979747325181961,
-0.13205674290657043,
0.027441103011369705,
-0.031591784209012985,
0.023208145052194595,
-0.032516054809093475,
0.02715190313756466,
-0.07499446719884872,
-0.019259175285696983,
-0.0029966835863888264,
-0.09803588688373566,
-0.03775005415081978,
-0.06078089028596878,
0.052076324820518494,
0.02510368824005127,
-0.08722852915525436,
0.04070739075541496,
0.0578894205391407,
-0.03513973206281662,
0.014414037577807903,
0.008136709220707417,
-0.020738638937473297,
-0.01428287848830223,
-0.015835408121347427,
-0.03890224173665047,
0.029282987117767334,
0.005934711080044508,
-0.05555165559053421,
0.01857861876487732,
0.0710272565484047,
0.0438765287399292,
0.047571826726198196,
-0.048634592443704605,
0.0031834300607442856,
-0.0963515043258667,
0.055587656795978546,
-0.006567682605236769,
-0.011348145082592964,
0.06395898014307022,
0.029340891167521477,
0.0330657921731472,
0.02014024741947651,
0.11040589213371277,
0.009325996041297913,
0.02167862467467785,
-0.054775018244981766,
0.027534963563084602,
0.016884662210941315,
0.09459801763296127,
-0.08688800036907196,
-4.005594433920123e-8,
-0.074089415371418,
0.011940822936594486,
0.0005786567344330251,
0.0833108052611351,
0.033948298543691635,
-0.06368906795978546,
-0.0025449960958212614,
0.04294401779770851,
-0.0065192291513085365,
0.0077652339823544025,
0.045014638453722,
0.013907705433666706,
-0.025335008278489113,
-0.03305070474743843,
-0.005921471863985062,
0.01785898208618164,
0.046978019177913666,
-0.023322634398937225,
-0.010817368514835835,
-0.05092042684555054,
0.03187486156821251,
0.04391588270664215,
-0.04701010137796402,
0.012586292810738087,
0.051027338951826096,
-0.036302350461483,
-0.0705040842294693,
0.051544252783060074,
0.05412561818957329,
0.027568485587835312,
0.054731130599975586,
0.024390682578086853,
0.047441985458135605,
0.006392497103661299,
0.10870923101902008,
-0.007516828365623951,
0.012683105655014515,
0.04125913977622986,
0.02678728476166725,
-0.09076552838087082,
0.0450766459107399,
0.06134245917201042,
-0.0434228889644146,
0.0018779394449666142,
0.006300734356045723,
-0.005255404394119978,
-0.07868087291717529,
-0.09992455691099167,
-0.027175042778253555,
-0.08181745558977127,
-0.02420618198812008,
0.026173092424869537,
0.021132005378603935,
0.05543386936187744,
0.10168229043483734,
0.04976309463381767,
0.020286833867430687,
-0.05069654807448387,
0.022391073405742645,
0.014796932227909565,
0.0414072722196579,
0.02757301554083824,
-0.039340145885944366,
0.04176848381757736
] | 0 | 0 | 0 | 0 |
|
Feature Responsiveness Scores: Model-Agnostic Explanations for Recourse | https://openreview.net/forum?id=wsWCVrH9dv | [
"Seung Hyun Cheon",
"Anneke Wernerfelt",
"Sorelle Friedler",
"Berk Ustun"
] | Poster | Machine learning models routinely automate decisions in applications like lending and hiring. In such settings, consumer protection rules require companies that deploy models to explain predictions to decision subjects. These rules are motivated, in part, by the belief that explanations can promote *recourse* by revealing information that individuals can use to contest or improve their outcomes. In practice, many companies comply with these rules by providing individuals with a list of the most important features for their prediction, which they identify based on feature importance scores from feature attribution methods such as SHAP or LIME. In this work, we show how these practices can undermine consumers by highlighting features that would not lead to an improved outcome and by explaining predictions that cannot be changed. We propose to address these issues by highlighting features based on their *responsiveness score*—i.e., the probability that an individual can attain a target prediction by changing a specific feature. We develop efficient methods to compute responsiveness scores for any model and any dataset. We conduct an extensive empirical study on the responsiveness of explanations in lending. Our results show that standard practices in consumer finance can backfire by presenting consumers with *reasons without recourse*, and demonstrate how our approach improves consumer protection by highlighting responsive features and identifying fixed predictions. | explainability, feature attribution, algorithmic recourse, regulation | We introduce *feature responsiveness scores*, the probability that an individual can change their model prediction by altering a feature. | 12,358 | 2410.22598 | [
-0.06769586354494095,
-0.003677776549011469,
0.015984809026122093,
0.03547464311122894,
0.1446203887462616,
0.029582850635051727,
0.033915650099515915,
0.0035735941492021084,
-0.03912443667650223,
-0.007269275840371847,
-0.040950071066617966,
-0.03388125076889992,
0.011132903397083282,
-0.026604875922203064,
-0.007196136750280857,
-0.04742242023348808,
0.1332368552684784,
0.039794664829969406,
-0.08469084650278091,
0.010871702805161476,
0.011409039609134197,
-0.04103830084204674,
0.03668200969696045,
0.0347885861992836,
0.021238261833786964,
-0.0559142529964447,
0.019316164776682854,
0.04225710406899452,
-0.03331850841641426,
-0.024199971929192543,
-0.0016599675873294473,
0.013401364907622337,
-0.010973840951919556,
-0.0032654888927936554,
-0.06774143874645233,
0.025906868278980255,
-0.0463591106235981,
0.010759495198726654,
-0.06624916195869446,
-0.08418701589107513,
0.03518193960189819,
-0.06927074491977692,
-0.04665939137339592,
0.015099204145371914,
0.09435008466243744,
0.017147831618785858,
0.0724109560251236,
0.00038188003236427903,
-0.04308311268687248,
0.0033639331813901663,
-0.10803946852684021,
-0.013560308143496513,
-0.012821880169212818,
-0.020406967028975487,
-0.037656817585229874,
-0.03736884146928787,
0.06257537752389908,
0.03848498687148094,
0.005723231006413698,
-0.07204418629407883,
0.03843946382403374,
-0.11675350368022919,
-0.052647463977336884,
0.045965369790792465,
-0.016826076433062553,
0.03843805566430092,
-0.02602352201938629,
0.033994343131780624,
-0.09617984294891357,
0.030463946983218193,
0.04372667148709297,
0.023690713569521904,
-0.022402849048376083,
0.008311254903674126,
0.05781365558505058,
0.07222263514995575,
0.008507163263857365,
0.07971788197755814,
0.04000913351774216,
-0.0006461612647399306,
-0.08589935302734375,
-0.06862013787031174,
-0.028485504910349846,
-0.0008725041407160461,
0.005507422611117363,
-0.04196897894144058,
0.014257904142141342,
-0.021796245127916336,
0.022221984341740608,
0.03592298552393913,
0.005104179494082928,
-0.006576396059244871,
0.027508758008480072,
-0.07648104429244995,
-0.04221334680914879,
0.017742788419127464,
-0.06959313154220581,
-0.1044590100646019,
0.0010211652843281627,
0.053267136216163635,
-0.059825263917446136,
0.11387588083744049,
0.010621004737913609,
-0.10331539064645767,
0.07286577671766281,
-0.04437461495399475,
-0.013033184222877026,
0.000009092025720747188,
0.07487627118825912,
-0.01738058775663376,
0.005512450356036425,
0.08126594126224518,
0.016670089215040207,
-0.03709013760089874,
0.08238903433084488,
0.0035475606564432383,
-0.06195381283760071,
0.05977801978588104,
-0.007450337987393141,
0.04880496859550476,
0.016416272148489952,
0.08113422244787216,
0.024891337379813194,
-0.010309980250895023,
-0.00471114506945014,
0.041202280670404434,
-0.13548387587070465,
-2.1635735761502313e-34,
0.03632114455103874,
0.02095719613134861,
0.013238403014838696,
0.024266481399536133,
0.06232379376888275,
-0.03838281333446503,
-0.0942312628030777,
0.05938328057527542,
0.0228376816958189,
0.07731522619724274,
-0.005058166105300188,
0.021210944280028343,
0.016872497275471687,
0.01606111042201519,
0.029879311099648476,
0.05533958971500397,
-0.06710202991962433,
0.04790128394961357,
-0.07616175711154938,
-0.027914758771657944,
0.0753450021147728,
-0.06694455444812775,
-0.011194741353392601,
-0.03747239336371422,
-0.017572781071066856,
-0.11463157832622528,
-0.03390371426939964,
0.023855822160840034,
0.007731311488896608,
0.021679773926734924,
-0.017177792266011238,
-0.04287886619567871,
0.07263785600662231,
0.013006323017179966,
-0.034248001873493195,
-0.05914408341050148,
-0.12392614036798477,
-0.0895879939198494,
0.045614637434482574,
0.031201541423797607,
-0.04829375818371773,
-0.029664596542716026,
-0.03212742134928703,
0.02246508002281189,
-0.0010731378570199013,
-0.009375455789268017,
-0.03846042603254318,
-0.0919593796133995,
-0.0899198055267334,
0.03273487091064453,
0.044013861566782,
0.07137705385684967,
0.02941758558154106,
-0.0516631044447422,
-0.11238314211368561,
-0.00014735397417098284,
-0.013147816061973572,
-0.07980509102344513,
-0.0826270654797554,
-0.040797024965286255,
0.01299959234893322,
-0.0037889222148805857,
0.0013703511795029044,
-0.013467776589095592,
-0.10964632034301758,
0.005360759329050779,
0.03247769549489021,
0.05258111283183098,
-0.0312080979347229,
0.02003186196088791,
-0.017320971935987473,
0.03927513584494591,
-0.01969830133020878,
-0.06764859706163406,
0.004714167211204767,
-0.027061613276600838,
0.032282933592796326,
0.0012798791285604239,
0.06339074671268463,
0.005845359992235899,
0.05114324390888214,
-0.004876552615314722,
0.030627919360995293,
0.008296751417219639,
-0.017157714813947678,
0.021051684394478798,
0.050758350640535355,
-0.05363991856575012,
-0.0018446637550368905,
-0.04614429920911789,
-0.08008085191249847,
0.005179054569453001,
-0.04394221678376198,
-0.02014324441552162,
0.0384393148124218,
-2.6631933890672512e-33,
-0.10551047325134277,
-0.03162728622555733,
-0.013045886531472206,
0.0027497783303260803,
-0.05157969519495964,
0.015858476981520653,
-0.10198704153299332,
-0.03009626269340515,
0.040095556527376175,
-0.06784381717443466,
-0.0478668138384819,
-0.069370798766613,
0.0701959952712059,
-0.038447801023721695,
-0.04036879166960716,
0.021745236590504646,
-0.021739697083830833,
0.004906055051833391,
0.038150981068611145,
0.0224322360008955,
0.027145052328705788,
0.07221037894487381,
-0.0005015455535613,
0.03497828543186188,
-0.03293043375015259,
-0.039212942123413086,
-0.049860913306474686,
0.022875837981700897,
0.07232919335365295,
-0.04910273104906082,
-0.060262009501457214,
0.012487415224313736,
-0.08007863163948059,
-0.027141176164150238,
-0.031380388885736465,
-0.013850806280970573,
0.003431508783251047,
-0.035518404096364975,
-0.04830851033329964,
0.17447210848331451,
0.12030119448900223,
-0.022353630512952805,
-0.0766519159078598,
-0.027668887749314308,
-0.03214820846915245,
-0.04531429335474968,
0.06291429698467255,
-0.08262347429990768,
0.06491132825613022,
0.0015064846957102418,
-0.003961306530982256,
0.006981157697737217,
0.007116013206541538,
0.12580178678035736,
-0.08371733874082565,
0.020273109897971153,
0.0383322574198246,
0.009830201044678688,
0.09713441878557205,
0.08405215293169022,
-0.09648092091083527,
0.03324158489704132,
0.049844685941934586,
-0.010915658436715603,
0.04532647132873535,
0.0022239263635128736,
0.0429389588534832,
0.0171382836997509,
0.0393977016210556,
-0.07967410981655121,
-0.005008350592106581,
-0.07675553113222122,
-0.04354707896709442,
0.026852859184145927,
-0.025262242183089256,
0.025677267462015152,
-0.015293043106794357,
-0.10397843271493912,
-0.07826684415340424,
-0.03341546654701233,
0.06972873955965042,
0.01717868447303772,
0.057900913059711456,
0.04291125014424324,
0.08327893167734146,
-0.01725011132657528,
0.08199647814035416,
0.04942607879638672,
0.013703729957342148,
0.012031312100589275,
-0.026142993941903114,
0.06272617727518082,
-0.07164553552865982,
0.1029701828956604,
-0.021130112931132317,
-4.3821430040225096e-8,
-0.00835750624537468,
0.02459106035530567,
0.010119115933775902,
0.0828566625714302,
0.11104998737573624,
0.012744331732392311,
-0.03742777928709984,
0.03176293149590492,
-0.07282646745443344,
-0.02292569912970066,
0.05656407028436661,
0.04842286556959152,
-0.07858402281999588,
-0.009294179268181324,
0.03284301236271858,
0.05408385023474693,
-0.03635350987315178,
-0.007524917833507061,
-0.04595551639795303,
0.02920721098780632,
0.02868712693452835,
0.01702190935611725,
0.037275444716215134,
-0.05519366264343262,
0.023283373564481735,
-0.07733294367790222,
-0.0112570496276021,
0.09054426103830338,
0.054142121225595474,
0.07469479739665985,
-0.05862452834844589,
0.0070448326878249645,
-0.015906022861599922,
-0.04385459050536156,
0.00793988723307848,
0.039550282061100006,
0.03123757615685463,
-0.06912144273519516,
0.06675338745117188,
0.023873480036854744,
0.019684793427586555,
0.08124600350856781,
-0.018788393586874008,
0.06679564714431763,
0.0458800233900547,
0.003532432485371828,
0.0025054982397705317,
-0.0647725984454155,
0.04491511359810829,
0.04296494647860527,
0.018368950113654137,
-0.027252573519945145,
0.08537963777780533,
0.05602729320526123,
0.05033627152442932,
-0.011635999195277691,
-0.019738854840397835,
-0.02174106240272522,
0.03774002194404602,
0.02074919454753399,
0.058672886341810226,
0.012783491052687168,
0.017540743574500084,
0.0030494583770632744
] | 0 | 0 | 0 | 0 |
|
InterMask: 3D Human Interaction Generation via Collaborative Masked Modeling | https://openreview.net/forum?id=ZAyuwJYN8N | [
"Muhammad Gohar Javed",
"chuan guo",
"Li cheng",
"Xingyu Li"
] | Poster | Generating realistic 3D human-human interactions from textual descriptions remains a challenging task. Existing approaches, typically based on diffusion models, often produce results lacking realism and fidelity. In this work, we introduce *InterMask*, a novel framework for generating human interactions using collaborative masked modeling in discrete space. InterMask first employs a VQ-VAE to transform each motion sequence into a 2D discrete motion token map. Unlike traditional 1D VQ token maps, it better preserves fine-grained spatio-temporal details and promotes *spatial awareness* within each token. Building on this representation, InterMask utilizes a generative masked modeling framework to collaboratively model the tokens of two interacting individuals. This is achieved by employing a transformer architecture specifically designed to capture complex spatio-temporal inter-dependencies. During training, it randomly masks the motion tokens of both individuals and learns to predict them. For inference, starting from fully masked sequences, it progressively fills in the tokens for both individuals. With its enhanced motion representation, dedicated architecture, and effective learning strategy, InterMask achieves state-of-the-art results, producing high-fidelity and diverse human interactions. It outperforms previous methods, achieving an FID of $5.154$ (vs $5.535$ of in2IN) on the InterHuman dataset and $0.399$ (vs $5.207$ of InterGen) on the InterX dataset. Additionally, InterMask seamlessly supports reaction generation without the need for model redesign or fine-tuning. | Motion Synthesis, Human Interaction Generation, Masked Generative Transformer, Text-driven Generation, Vector Quantized VAE | A novel framework for human interaction generation using collaborative masked modeling in the discrete space, which explicitly models spatio-temporal dependencies within and between the interacting individuals. | 12,356 | 2410.10010 | [
-0.05767035111784935,
-0.10201992094516754,
0.05413445830345154,
0.019677022472023964,
0.020162168890237808,
-0.025373123586177826,
-0.06079680845141411,
-0.05586909130215645,
0.048579659312963486,
-0.06813488155603409,
-0.05716230720281601,
-0.11099120229482651,
-0.07044854760169983,
0.05072057619690895,
-0.023043693974614143,
-0.012026124633848667,
0.07654854655265808,
0.12453632056713104,
-0.036129292100667953,
0.0037830667570233345,
0.05779360607266426,
-0.006655827164649963,
-0.0005757326725870371,
-0.00045164814218878746,
0.03201763704419136,
0.00924812164157629,
0.06736636161804199,
-0.09495583921670914,
0.02492370456457138,
0.005098342429846525,
0.043092112988233566,
-0.07521293312311172,
0.010603553615510464,
0.0024199211038649082,
-0.07398410141468048,
0.06411939859390259,
-0.019816430285573006,
0.0365152545273304,
-0.027754373848438263,
-0.09168095141649246,
-0.036370884627103806,
0.04268967732787132,
0.049098845571279526,
0.02054610475897789,
0.07221032679080963,
0.03786984831094742,
-0.009397463873028755,
-0.04585033655166626,
-0.0010101216612383723,
0.04772946983575821,
-0.04009738191962242,
-0.02183002419769764,
0.056278835982084274,
0.05071517452597618,
-0.05996255949139595,
0.018644407391548157,
0.0017259015003219247,
-0.053333256393671036,
0.05788034945726395,
-0.04024890437722206,
-0.05968264490365982,
-0.054961588233709335,
0.040800709277391434,
0.024757781997323036,
-0.027673987671732903,
-0.08601600676774979,
0.043607864528894424,
0.04976348951458931,
-0.005523601081222296,
0.004789997823536396,
-0.07317931950092316,
-0.0022846893407404423,
-0.05192408338189125,
-0.00656555313616991,
-0.0053306263871490955,
0.07503608614206314,
-0.010230050422251225,
-0.09148105978965759,
0.027081109583377838,
-0.05266174301505089,
0.09203078597784042,
-0.027662284672260284,
0.04452625289559364,
-0.01043868437409401,
0.005771306809037924,
0.02312091737985611,
-0.02271159552037716,
0.11721288412809372,
-0.09848768264055252,
0.01301079336553812,
-0.11810359358787537,
0.010821257717907429,
0.06804899126291275,
-0.04791229963302612,
0.004394287243485451,
-0.020581062883138657,
-0.045233942568302155,
-0.0005261520855128765,
0.001896344474516809,
0.08985146880149841,
-0.03257889300584793,
-0.006213725078850985,
-0.015408490784466267,
-0.05382053926587105,
0.031105825677514076,
0.02009103260934353,
0.047237031161785126,
-0.037226513028144836,
0.0100370142608881,
0.031746815890073776,
0.014438717626035213,
-0.03700825572013855,
-0.05698678269982338,
0.009202908724546432,
0.027431931346654892,
0.06384303420782089,
0.02733912318944931,
-0.008090917021036148,
0.07293950766324997,
0.04119906574487686,
0.027178794145584106,
-0.02873295731842518,
-0.05141061916947365,
0.018134012818336487,
-0.02951483242213726,
-0.0950944647192955,
0.0251530259847641,
-1.4349693922384491e-33,
0.018495550379157066,
0.002828639931976795,
0.05138680338859558,
0.0893653854727745,
0.04276157170534134,
0.0061262138187885284,
-0.006798291578888893,
-0.030057281255722046,
-0.05545661598443985,
-0.036269716918468475,
0.014379697851836681,
-0.05031563714146614,
-0.035815589129924774,
0.13131187856197357,
-0.018999967724084854,
-0.00440734438598156,
-0.01670551486313343,
0.010945872403681278,
-0.0357368104159832,
0.04052048549056053,
0.006668607704341412,
-0.02459503524005413,
-0.06653162837028503,
0.039761077612638474,
0.013459227979183197,
0.01462023425847292,
0.0569760762155056,
-0.02529686503112316,
0.07462722063064575,
-0.01517892349511385,
-0.062010910362005234,
0.030684806406497955,
0.005258888937532902,
-0.0404670424759388,
0.07301930338144302,
0.02612338960170746,
0.03203250840306282,
-0.07659818977117538,
-0.011798261664807796,
0.02928553894162178,
0.007369315251708031,
-0.025883963331580162,
-0.05019219592213631,
-0.059458862990140915,
-0.10537055134773254,
-0.009359950199723244,
0.05892890319228172,
0.11773894727230072,
-0.10126297175884247,
0.018209006637334824,
0.07743949443101883,
0.06223608925938606,
-0.08355903625488281,
-0.0657113716006279,
-0.007003562059253454,
-0.056558527052402496,
0.06259061396121979,
-0.04435745254158974,
0.06771579384803772,
-0.02055605873465538,
-0.006787607911974192,
0.0401625894010067,
0.026022424921393394,
0.041937366127967834,
0.0067947739735245705,
0.015470581129193306,
-0.028793608769774437,
0.06721704453229904,
0.04846873879432678,
0.013508865609765053,
-0.07194137573242188,
0.043222278356552124,
-0.017810944467782974,
-0.03686990961432457,
-0.004866238217800856,
-0.014990251511335373,
0.02892191894352436,
-0.04589521139860153,
-0.021664021536707878,
0.09167780727148056,
-0.048140257596969604,
-0.09559065103530884,
-0.0321880578994751,
-0.07418949156999588,
-0.05456208065152168,
-0.049605440348386765,
-0.000962776888627559,
0.019592100754380226,
-0.02859051153063774,
-0.040765948593616486,
0.009279327467083931,
-0.02947147749364376,
-0.02854648232460022,
0.08194109052419662,
0.024823099374771118,
-1.0716658276322838e-34,
0.0029017699416726828,
0.1145862489938736,
-0.018729280680418015,
0.026334252208471298,
0.01631850190460682,
-0.011950530111789703,
0.02067568339407444,
-0.016267869621515274,
0.06902812421321869,
-0.0271865613758564,
-0.09770891070365906,
0.018524035811424255,
0.08431489765644073,
0.0035144502762705088,
0.0652974545955658,
0.015818102285265923,
0.03236756846308708,
-0.017708735540509224,
-0.04303102567791939,
0.01499217376112938,
0.049702078104019165,
0.04185681417584419,
-0.09355244785547256,
-0.005873597227036953,
-0.08816046267747879,
0.06336373090744019,
0.028672242537140846,
0.01582673378288746,
0.02243061177432537,
0.06937307119369507,
-0.001548222964629531,
0.008123422972857952,
0.04370242357254028,
-0.04209396615624428,
-0.12030328810214996,
0.029934899881482124,
0.057145606726408005,
0.01068712305277586,
-0.033529315143823624,
-0.01182723231613636,
0.02587522752583027,
-0.06391774117946625,
-0.07451990991830826,
0.06511852890253067,
-0.08332055062055588,
0.01810598187148571,
-0.028935754671692848,
0.008111719973385334,
-0.02550993300974369,
0.01940813660621643,
-0.035294122993946075,
0.10296625643968582,
-0.12659215927124023,
-0.05097212642431259,
-0.05624305084347725,
-0.07531663775444031,
-0.06356256455183029,
-0.011705020442605019,
0.11216235905885696,
-0.07382871955633163,
-0.023401178419589996,
-0.022887161001563072,
0.013252086006104946,
-0.021308330819010735,
0.014706091023981571,
0.007853589951992035,
-0.038745228201150894,
-0.027413014322519302,
0.0043852683156728745,
0.026147516444325447,
0.06966149806976318,
0.0021724458783864975,
-0.004551841877400875,
-0.009524667635560036,
0.0879204273223877,
-0.05758066102862358,
0.029518406838178635,
-0.04075058177113533,
-0.014167836867272854,
-0.025660382583737373,
-0.00034292700001969934,
0.0032564173452556133,
0.05430721119046211,
0.08953309804201126,
0.06670846045017242,
0.10865924507379532,
-0.017731178551912308,
0.1028992310166359,
-0.07407921552658081,
0.10017510503530502,
-0.056816037744283676,
0.11686412245035172,
-0.018924573436379433,
0.03471972793340683,
-0.059291772544384,
-3.886948007902902e-8,
-0.04775787144899368,
0.048326410353183746,
-0.03674653172492981,
0.010223337449133396,
-0.05090839043259621,
0.0021027831826359034,
0.0295590627938509,
0.02733970247209072,
0.01875274069607258,
0.011085773818194866,
-0.0015931229572743177,
-0.025707930326461792,
0.07211700826883316,
-0.006032113917171955,
0.0677751749753952,
0.049433279782533646,
-0.04143309220671654,
-0.003964126575738192,
-0.01717047207057476,
0.09359467774629593,
-0.01670319214463234,
0.058145713061094284,
-0.01483998168259859,
0.014500439167022705,
0.015667026862502098,
0.05527959018945694,
-0.00800340622663498,
0.03560981526970863,
0.005224424414336681,
0.036199335008859634,
0.0003060265153180808,
0.036804135888814926,
-0.0011333966394886374,
-0.028473911806941032,
-0.07029257714748383,
0.0589667446911335,
0.009103558026254177,
-0.046419546008110046,
-0.0033379937522113323,
-0.07605165988206863,
-0.004432293586432934,
0.06388922780752182,
0.020216697826981544,
0.010050103068351746,
-0.03822655975818634,
0.022078661248087883,
0.01848777011036873,
-0.16134406626224518,
0.022082064300775528,
0.010581271722912788,
-0.10665550827980042,
-0.06396320462226868,
-0.041681718081235886,
0.09697386622428894,
0.12412574887275696,
0.06603829562664032,
-0.012452535331249237,
0.014736266806721687,
0.1203770637512207,
0.021987808868288994,
0.012727326713502407,
-0.016167709603905678,
-0.07476097345352173,
0.00972671713680029
] | https://github.com/gohar-malik/intermask | 26 | 0 | 0 | 0 |
Denoising Autoregressive Transformers for Scalable Text-to-Image Generation | https://openreview.net/forum?id=amDkNPVWcn | [
"Jiatao Gu",
"Yuyang Wang",
"Yizhe Zhang",
"Qihang Zhang",
"Dinghuai Zhang",
"Navdeep Jaitly",
"Joshua M. Susskind",
"Shuangfei Zhai"
] | Poster | Diffusion models have become the dominant approach for visual generation. They are trained by denoising a Markovian process which gradually adds noise to the input. We argue that the Markovian property limits the model’s ability to fully utilize the generation trajectory, leading to inefficiencies during training and inference. In this paper, we propose DART, a transformer-based model that unifies autoregressive (AR) and diffusion within a non-Markovian framework. DART iteratively denoises image patches spatially and spectrally using an AR model that has the same architecture as standard language models. DART does not rely on image quantization, which enables more effective image modeling while maintaining flexibility. Furthermore, DART seamlessly trains with both text and image data in a unified model. Our approach demonstrates competitive performance on class-conditioned and text-to-image generation tasks, offering a scalable, efficient alternative to traditional diffusion models. Through this unified framework, DART sets a new benchmark for scalable, high-quality image synthesis. | diffusion models, autoregressive models, Transformer | DART unifies autoregressive and diffusion models for efficient, high-quality image generation. | 12,352 | 2410.08159 | [
-0.08076479285955429,
-0.0783766359090805,
0.03944806009531021,
-0.0026509102899581194,
0.0054829735308885574,
-0.046347495168447495,
-0.04255121573805809,
-0.12428971379995346,
0.0530644953250885,
-0.04611733928322792,
-0.05233633518218994,
-0.007630380801856518,
0.03901916369795799,
0.004983309656381607,
-0.016727065667510033,
0.0024898694828152657,
0.011717518791556358,
0.044511955231428146,
-0.07499812543392181,
-0.08775261044502258,
-0.02262481302022934,
-0.0020158295519649982,
0.031361836940050125,
-0.06407760083675385,
0.11532779783010483,
0.11183234304189682,
0.034847889095544815,
-0.11038002371788025,
-0.003113260492682457,
-0.02596127986907959,
0.016104713082313538,
-0.02411632053554058,
-0.06165070831775665,
-0.013736389577388763,
-0.0445232130587101,
0.06381877511739731,
-0.03793130815029144,
0.055925674736499786,
0.03612780570983887,
-0.03242405876517296,
0.00985636841505766,
0.057344939559698105,
-0.02880695089697838,
-0.019483398646116257,
0.06636722385883331,
-0.005234412383288145,
0.003426759270951152,
-0.055136602371931076,
0.05842951685190201,
0.025681620463728905,
-0.017335740849375725,
-0.01028740406036377,
-0.09840373694896698,
0.02512626349925995,
-0.01682174764573574,
0.0789615586400032,
0.09693016856908798,
0.0016800572630017996,
0.009011921472847462,
-0.03533731773495674,
-0.023381581529974937,
-0.026482239365577698,
0.023784952238202095,
0.0015384950675070286,
-0.029630063101649284,
-0.0340769961476326,
0.1017477810382843,
0.02484407089650631,
0.023493407294154167,
0.011086612939834595,
-0.07082054764032364,
0.04495304822921753,
-0.009389625862240791,
0.040870826691389084,
-0.05255680903792381,
0.018889999017119408,
0.036064278334379196,
-0.09433829039335251,
0.06641532480716705,
-0.07586047798395157,
0.07631006091833115,
-0.03385109081864357,
0.08257141709327698,
0.0042107286863029,
0.07559120655059814,
0.06106005236506462,
0.01913810893893242,
0.009831012226641178,
-0.023745831102132797,
-0.031721945852041245,
-0.06106488034129143,
0.0016082336660474539,
-0.004989962093532085,
-0.016442518681287766,
-0.009728284552693367,
0.03827346861362457,
0.023163501173257828,
-0.07263872772455215,
-0.03743984177708626,
0.12457101792097092,
-0.0014745746739208698,
-0.017558183521032333,
0.012964330613613129,
0.004326903261244297,
0.03995836526155472,
-0.06285935640335083,
0.0876876637339592,
0.028458211570978165,
-0.03352648764848709,
0.011465382762253284,
0.047346509993076324,
-0.001602770877070725,
-0.009384133853018284,
-0.006417376920580864,
0.0496038943529129,
-0.030742255970835686,
0.02025722526013851,
-0.028916513547301292,
-0.03136013448238373,
0.015332840383052826,
0.019777650013566017,
0.0008354958263225853,
-0.0690147876739502,
0.04531630501151085,
-0.025659779086709023,
-0.045156776905059814,
-0.018632015213370323,
8.478894935423893e-34,
0.005318601615726948,
-0.010620780289173126,
0.11504624038934708,
0.024075279012322426,
0.10125476121902466,
-0.009503108449280262,
-0.02689959481358528,
-0.046871066093444824,
-0.02095273695886135,
-0.09360233694314957,
0.011234596371650696,
-0.0602775439620018,
-0.08061289042234421,
0.11975562572479248,
0.081214040517807,
-0.04717118665575981,
-0.04739714413881302,
0.03402392566204071,
0.03285301476716995,
-0.015844736248254776,
-0.002226503798738122,
0.019353797659277916,
-0.012931562028825283,
-0.02824549190700054,
-0.09869823604822159,
0.02904815785586834,
0.08081161975860596,
-0.029628073796629906,
-0.008173057809472084,
-0.006401946768164635,
-0.14192520081996918,
-0.03794263303279877,
0.03938019648194313,
-0.008811896666884422,
0.0033051606733351946,
-0.04982037842273712,
-0.01456600334495306,
-0.04861248657107353,
-0.0015595471486449242,
-0.008475864306092262,
-0.03488372638821602,
0.06023530662059784,
-0.022436033934354782,
-0.05352180451154709,
-0.08789967000484467,
0.024330567568540573,
0.011927083134651184,
0.03623810037970543,
-0.08765097707509995,
-0.0011826398549601436,
0.07857614010572433,
0.003734627505764365,
0.022222667932510376,
-0.013324537314474583,
-0.012919881381094456,
-0.05505157262086868,
0.13019396364688873,
-0.0784958153963089,
0.02267289161682129,
-0.00047747569624334574,
0.042137615382671356,
-0.008957065641880035,
-0.002712592016905546,
0.06458238512277603,
0.07315672934055328,
-0.031503863632678986,
0.018110670149326324,
0.05046829208731651,
-0.021923931315541267,
-0.042809367179870605,
0.0012134913122281432,
0.039939671754837036,
0.010363574139773846,
-0.0751977264881134,
0.04783812165260315,
0.017064891755580902,
0.051366303116083145,
0.05631137639284134,
-0.0312196072191,
0.04255259782075882,
-0.07331562042236328,
-0.02942151576280594,
0.0023277585860341787,
-0.08722297102212906,
0.03368772566318512,
-0.005027849692851305,
0.006050995085388422,
-0.08954369276762009,
-0.03555965796113014,
-0.01765703596174717,
-0.03689277544617653,
-0.011654612608253956,
-0.00023089929891284555,
-0.0009351841290481389,
0.034535400569438934,
-6.929728608808211e-35,
0.022479107603430748,
0.0677916407585144,
-0.03635169938206673,
0.08851704001426697,
-0.03269737586379051,
-0.06353390216827393,
0.029275869950652122,
0.04432157054543495,
0.0801822617650032,
-0.10584872215986252,
-0.11582854390144348,
0.06051144748926163,
-0.009383299387991428,
-0.07371696084737778,
-0.019078878685832024,
-0.09380409121513367,
0.10570788383483887,
-0.037500668317079544,
-0.08175509423017502,
0.005439357366412878,
0.020941440016031265,
0.05243252217769623,
0.018595770001411438,
-0.0037846954073756933,
-0.03683331236243248,
0.035774942487478256,
-0.030157804489135742,
0.08949095755815506,
-0.00915649626404047,
-0.04836298152804375,
0.034060362726449966,
-0.011722858995199203,
0.08309417963027954,
-0.005823991261422634,
-0.12027698755264282,
0.026026086881756783,
0.08079659193754196,
0.017960773780941963,
-0.08050280809402466,
0.09822624176740646,
0.043002523481845856,
-0.0759197324514389,
-0.07379131019115448,
0.024452699348330498,
-0.0003856498224195093,
-0.00869740080088377,
-0.04112491011619568,
0.00985234696418047,
0.018169455230236053,
0.03069244883954525,
0.05060436576604843,
0.03634608909487724,
-0.08986179530620575,
0.045514147728681564,
0.00796213373541832,
-0.09367676079273224,
-0.10978172719478607,
-0.011057165451347828,
0.08922132849693298,
0.03254774957895279,
-0.046697258949279785,
-0.0355510488152504,
-0.03171563148498535,
-0.0915973037481308,
-0.039510082453489304,
0.009314987808465958,
-0.09238298237323761,
0.015443868935108185,
0.04400371387600899,
0.07183361798524857,
0.0450645387172699,
-0.017145244404673576,
0.0023203936871141195,
0.07337311655282974,
0.02646864764392376,
-0.05782410502433777,
0.08440591394901276,
-0.01380860898643732,
0.054534055292606354,
-0.06403444707393646,
-0.011374339461326599,
-0.011365962214767933,
-0.017806120216846466,
0.06705356389284134,
0.03432488813996315,
0.08349990099668503,
-0.00985883641988039,
-0.05063779652118683,
0.06321968138217926,
0.04729017987847328,
0.005521220620721579,
0.05162598192691803,
-0.006604594644159079,
0.04163829982280731,
-0.05778827145695686,
-3.7204163305659677e-8,
-0.020104780793190002,
0.05497923120856285,
0.029646234586834908,
0.006167938467115164,
0.033551931381225586,
-0.043912168592214584,
0.08912529051303864,
0.05428438261151314,
0.08852239698171616,
-0.06954553723335266,
0.10853520780801773,
-0.08124824613332748,
-0.007018109783530235,
-0.0618014857172966,
-0.021398762241005898,
0.060070157051086426,
0.024415573105216026,
0.03574143350124359,
0.001369612873531878,
-0.04538973048329353,
-0.08699311316013336,
0.07141965627670288,
-0.060218553990125656,
-0.03088207170367241,
0.045997414737939835,
-0.025219719856977463,
0.018853435292840004,
-0.0005770877469331026,
0.08081526309251785,
0.003658970817923546,
0.010742716491222382,
0.03381766378879547,
0.044562675058841705,
-0.019804494455456734,
-0.007507392205297947,
0.0026850500144064426,
0.04944082349538803,
0.019339971244335175,
-0.048578936606645584,
-0.0773354172706604,
0.001553201349452138,
0.0848628357052803,
-0.030974946916103363,
0.0035175553057342768,
-0.024002542719244957,
0.00849856249988079,
0.05080565810203552,
-0.08636672049760818,
0.05974354222416878,
0.0002650967799127102,
0.03725307807326317,
-0.03600645437836647,
0.03408359736204147,
-0.007533967960625887,
0.012148404493927956,
-0.02920602820813656,
0.05782304331660271,
0.01704745925962925,
0.062238965183496475,
0.09911447763442993,
0.03542160242795944,
0.009589462541043758,
0.013731672428548336,
0.04275977239012718
] | 0 | 0 | 0 | 0 |
|
Many-Objective Multi-Solution Transport | https://openreview.net/forum?id=Neb17mimVH | [
"Ziyue Li",
"Tian Li",
"Virginia Smith",
"Jeff Bilmes",
"Tianyi Zhou"
] | Poster | Optimizing the performance of many objectives (instantiated by tasks or clients) jointly with a few Pareto stationary solutions (models) is critical in machine learning. However, previous multi-objective optimization methods often focus on a few objectives and cannot scale to many objectives that outnumber the solutions, leading to either subpar performance or ignored objectives. We introduce ''Many-objective multi-solution Transport (MosT)'', a framework that finds multiple diverse solutions in the Pareto front of many objectives. Our insight is to seek multiple solutions, each performing as a domain expert and focusing on a specific subset of objectives while collectively covering all of them. MosT formulates the problem as a bi-level optimization of weighted objectives for each solution, where the weights are defined by an optimal transport between objectives and solutions. Our algorithm ensures convergence to Pareto stationary solutions for complementary subsets of objectives. On a range of applications in federated learning, multi-task learning, and mixture-of-prompt learning for LLMs, MosT distinctly outperforms strong baselines, delivering high-quality, diverse solutions that profile the entire Pareto frontier, thus ensuring balanced trade-offs across many objectives. | Multi-Objective Optimization, Mixture of Experts | null | 12,344 | 2403.04099 | [
0.010328504256904125,
-0.08536049723625183,
-0.006621749605983496,
-0.03490805998444557,
0.08639033138751984,
-0.016239186748862267,
0.04602987319231033,
-0.01966741867363453,
-0.026923906058073044,
0.022068150341510773,
-0.10830790549516678,
-0.09555906057357788,
0.0035924362018704414,
0.10027587413787842,
0.004704797174781561,
0.06990122050046921,
0.137916699051857,
0.03732727840542793,
-0.038262199610471725,
-0.007681224960833788,
0.024487517774105072,
0.046151019632816315,
0.016721058636903763,
0.06411436945199966,
-0.046245038509368896,
-0.018035074695944786,
0.022047972306609154,
-0.028514830395579338,
-0.03091421350836754,
0.004118119832128286,
0.0299671683460474,
0.008485950529575348,
-0.02689037099480629,
0.001436833175830543,
-0.03794535622000694,
0.11323214322328568,
-0.06198974698781967,
0.04478445649147034,
0.023985587060451508,
0.01157661434262991,
0.02335045114159584,
-0.027410035952925682,
-0.010328187607228756,
-0.018989432603120804,
0.044481102377176285,
-0.02178928069770336,
-0.028323883190751076,
-0.06590918451547623,
0.018666142597794533,
0.02043958380818367,
0.022112952545285225,
-0.06172758713364601,
-0.09275104105472565,
0.1127287894487381,
0.004247135017067194,
0.01592436619102955,
0.02118374966084957,
-0.012073651887476444,
0.004739387426525354,
0.02101484127342701,
-0.07170421630144119,
-0.09800735861063004,
-0.02448255755007267,
-0.05594925582408905,
0.04583755508065224,
-0.030747240409255028,
-0.04664204269647598,
0.07074267417192459,
-0.11840014904737473,
0.09653691947460175,
0.04289844632148743,
0.000046578275942010805,
-0.047973278909921646,
-0.02525409311056137,
0.013041340745985508,
0.06744410842657089,
0.08544930815696716,
0.003284105332568288,
-0.018092770129442215,
-0.04490656033158302,
-0.023785728961229324,
0.045986421406269073,
-0.03214746341109276,
0.02520827017724514,
0.1250665932893753,
-0.0884263813495636,
-0.09342385828495026,
-0.02874748222529888,
0.030354725196957588,
0.012270046398043633,
-0.03332161903381348,
0.053767114877700806,
0.03783177211880684,
0.018374651670455933,
0.019888194277882576,
-0.013781982474029064,
-0.006138359196484089,
-0.09295772761106491,
0.03272691369056702,
0.09479275345802307,
-0.06590774655342102,
0.06311000138521194,
0.0607304647564888,
-0.07928605377674103,
-0.020201683044433594,
-0.07417788356542587,
-0.019019396975636482,
0.05895564705133438,
0.04706437513232231,
-0.09868788719177246,
0.028892049565911293,
0.06536232680082321,
0.06160486117005348,
0.05852959305047989,
-0.01861930638551712,
0.00938586425036192,
-0.006067322101444006,
0.006203410681337118,
0.04177169129252434,
0.13937604427337646,
-0.10828526318073273,
-0.03517787903547287,
0.08421127498149872,
0.041174110025167465,
0.005121273919939995,
-0.015089225955307484,
-0.016575969755649567,
2.1123413079085727e-33,
-0.05574885755777359,
-0.012586023658514023,
0.061432644724845886,
0.007002886850386858,
-0.04870598763227463,
-0.04913992062211037,
-0.01075432077050209,
0.011819080449640751,
-0.02661578729748726,
-0.035941142588853836,
0.013967370614409447,
0.09440512955188751,
-0.05701328068971634,
0.04815785586833954,
0.03490470349788666,
-0.08931590616703033,
-0.058005742728710175,
0.13068273663520813,
0.04890364781022072,
-0.0883498415350914,
0.04378543794155121,
-0.0681682601571083,
-0.03807514160871506,
-0.025737151503562927,
0.0685097724199295,
-0.020867109298706055,
0.12177771329879761,
-0.06708553433418274,
-0.013010283000767231,
0.04007124528288841,
-0.07282815128564835,
0.006699266377836466,
-0.00482935132458806,
0.08559836447238922,
-0.02512313425540924,
0.0034137850161641836,
-0.05238449200987816,
0.00023527350276708603,
-0.04273327812552452,
0.007773983757942915,
-0.06535893678665161,
-0.04473685100674629,
0.05689721927046776,
0.023174108937382698,
-0.058399446308612823,
-0.07503259181976318,
-0.022616509348154068,
-0.00910672452300787,
-0.016386039555072784,
-0.06483739614486694,
-0.001232057111337781,
-0.06439372152090073,
-0.0006242805393412709,
-0.05981311947107315,
-0.03759927302598953,
-0.021662797778844833,
-0.05986780673265457,
0.014714370481669903,
0.0571853332221508,
-0.024610742926597595,
0.020230410620570183,
0.0210573673248291,
-0.05128393694758415,
-0.022260479629039764,
0.08399487286806107,
0.07803229242563248,
0.00672713341191411,
0.04735938832163811,
0.08456934243440628,
-0.057428501546382904,
-0.006238721311092377,
0.026196978986263275,
0.06351310014724731,
-0.11277823150157928,
0.09519101679325104,
-0.00009352434426546097,
0.0888528898358345,
-0.023467937484383583,
-0.032807666808366776,
-0.027030494064092636,
0.00530273886397481,
0.07919792085886002,
-0.018709121271967888,
-0.04798814654350281,
0.05492439493536949,
-0.012704245746135712,
0.020052028819918633,
-0.07276539504528046,
-0.10025464743375778,
0.015478909015655518,
-0.05393803492188454,
0.020116223022341728,
-0.04230343922972679,
0.02139565348625183,
0.006515074986964464,
-9.149182018071916e-34,
-0.02697439305484295,
-0.024372523650527,
-0.035539451986551285,
0.054878637194633484,
0.06106202304363251,
-0.035425715148448944,
-0.08465467393398285,
-0.07657714933156967,
-0.060714833438396454,
0.030040595680475235,
-0.03769589588046074,
-0.04707949608564377,
0.02951669506728649,
-0.028928812593221664,
-0.053769294172525406,
0.04479256644845009,
-0.026016831398010254,
0.02362458035349846,
0.030219996348023415,
-0.012364588677883148,
-0.009693988598883152,
0.08790753036737442,
0.020407628268003464,
0.03930244967341423,
-0.02599724940955639,
0.02437196671962738,
-0.051790688186883926,
0.07142170518636703,
0.01194075495004654,
0.0744689404964447,
0.007625145372003317,
-0.050711579620838165,
-0.02966086007654667,
-0.01311948336660862,
-0.025389034301042557,
0.10717324912548065,
-0.019344551488757133,
0.03208349645137787,
-0.003177803009748459,
0.12480904161930084,
0.03072500042617321,
-0.09590557962656021,
-0.013578717596828938,
0.007119413465261459,
0.00815583486109972,
0.004355650395154953,
0.00968039408326149,
-0.040212128311395645,
0.05685298144817352,
0.02283080667257309,
-0.023146888241171837,
0.06493879854679108,
-0.04270319268107414,
-0.001737814163789153,
-0.051945723593235016,
-0.02518819086253643,
0.005331107880920172,
-0.05012446641921997,
0.05986590310931206,
-0.042277880012989044,
0.05676795914769173,
0.030728181824088097,
0.03705238550901413,
-0.019775139167904854,
0.03097856044769287,
-0.0021691513247787952,
0.04689602553844452,
-0.024009985849261284,
-0.08972650021314621,
0.02396169677376747,
-0.04412034898996353,
-0.039075836539268494,
0.03173051029443741,
-0.018550993874669075,
-0.014164258725941181,
0.007591735105961561,
0.07482653856277466,
-0.013367686420679092,
-0.04845839738845825,
-0.029372896999120712,
0.06402405351400375,
-0.012701637111604214,
0.006395123898983002,
0.0711853876709938,
0.025525495409965515,
-0.032938066869974136,
0.0784689411520958,
0.09256497025489807,
0.027134940028190613,
0.008883983828127384,
0.00870357733219862,
0.006633275654166937,
0.03049631603062153,
-0.03592187166213989,
-0.024220222607254982,
-4.153926624894666e-8,
-0.02655830979347229,
0.001487742061726749,
0.019516319036483765,
-0.058120906352996826,
0.05208281800150871,
0.00274207117035985,
-0.07938619703054428,
0.0901421383023262,
0.041529297828674316,
-0.021171603351831436,
0.029672101140022278,
0.03515243902802467,
-0.03499987721443176,
-0.020882826298475266,
-0.06720855832099915,
0.008661126717925072,
-0.04247231036424637,
-0.022916533052921295,
-0.02265198901295662,
-0.043006714433431625,
0.04350624978542328,
0.02570435032248497,
0.03629763796925545,
-0.06711212545633316,
0.06821724027395248,
-0.04623368754982948,
-0.0387679822742939,
-0.003935463260859251,
0.10827872902154922,
0.002834963146597147,
-0.04387908801436424,
0.011095557361841202,
-0.013729061931371689,
0.009019779041409492,
-0.020806757733225822,
0.09530550241470337,
0.056974537670612335,
0.03168261796236038,
-0.03855583071708679,
-0.02263656258583069,
0.0016752600204199553,
0.021920286118984222,
-0.07669670879840851,
0.012730371206998825,
0.03953003138303757,
0.00798315554857254,
-0.03985133767127991,
0.003239378798753023,
0.03984668850898743,
0.011149145662784576,
-0.005407790653407574,
-0.04703856632113457,
0.07767455279827118,
0.11532751470804214,
0.0636601522564888,
-0.017735537141561508,
-0.041419245302677155,
-0.10430449992418289,
0.09307655692100525,
0.08320513367652893,
0.04154094681143761,
0.05095492675900459,
-0.08177514374256134,
0.05116838589310646
] | 0 | 0 | 0 | 0 |
|
An Auditing Test to Detect Behavioral Shift in Language Models | https://openreview.net/forum?id=h0jdAboh0o | [
"Leo Richter",
"Xuanli He",
"Pasquale Minervini",
"Matt Kusner"
] | Poster | As language models (LMs) approach human-level performance, a comprehensive understanding of their behavior becomes crucial.
This includes evaluating capabilities, biases, task performance, and alignment with societal values. Extensive initial evaluations, including red teaming and diverse benchmarking, can establish a model’s behavioral profile. However, subsequent fine-tuning or deployment modifications may alter these behaviors in unintended ways. We present an efficient statistical test to tackle Behavioral Shift Auditing (BSA) in LMs, which we define as detecting distribution shifts in qualitative properties of the output distributions of LMs. Our test compares model generations from a baseline model to those of the model under scrutiny and provides theoretical guarantees for change detection while controlling false positives. The test features a configurable tolerance parameter that adjusts sensitivity to behavioral changes for different use cases. We evaluate our approach using two case studies: monitoring changes in (a) toxicity and (b) translation performance. We find that the test is able to detect meaningful changes in behavior distributions using just hundreds of examples. | AI alignment, model auditing, model evaluations, red teaming, sequential hypothesis testing | We present a method for efficiently detecting behavioral changes in language models through output comparisons. | 12,343 | 2410.19406 | [
0.022713204845786095,
-0.09946661442518234,
-0.0010516364127397537,
0.03650661185383797,
0.07801254093647003,
0.01526450552046299,
0.01656104251742363,
-0.07255031913518906,
0.06111392006278038,
-0.014128431677818298,
-0.027294855564832687,
-0.1143006682395935,
0.027806909754872322,
0.05365433171391487,
0.02109082043170929,
-0.03907150775194168,
0.07227254658937454,
0.044580668210983276,
-0.04894820600748062,
-0.016118912026286125,
0.03017541952431202,
0.0573221780359745,
0.02793683297932148,
0.052835460752248764,
-0.024690138176083565,
-0.03809712827205658,
-0.055755238980054855,
-0.04922890663146973,
-0.022584719583392143,
-0.05380914732813835,
0.003319468116387725,
0.04520251974463463,
0.008395764976739883,
0.0585247166454792,
-0.005700526759028435,
-0.008803388103842735,
-0.04263525828719139,
-0.014570591039955616,
0.07190553098917007,
-0.04153580218553543,
-0.08843803405761719,
-0.0843643769621849,
0.0575285367667675,
-0.07276000082492828,
-0.0005579199641942978,
-0.0695868581533432,
-0.015201239846646786,
-0.011573711410164833,
-0.09846872091293335,
0.04394729062914848,
-0.04571971669793129,
-0.025603171437978745,
0.07579255104064941,
0.02577657625079155,
-0.04087775573134422,
-0.020221827551722527,
0.12249772250652313,
0.10492187738418579,
0.007630770094692707,
-0.017893601208925247,
-0.045579832047224045,
-0.0689006894826889,
-0.0665859654545784,
0.05313895270228386,
-0.07767099142074585,
0.039211101830005646,
-0.047145504504442215,
0.021212099120020866,
-0.028391888365149498,
0.015578304417431355,
-0.07753986120223999,
-0.007739319000393152,
0.0009548361413180828,
-0.005795279052108526,
0.03683577477931976,
0.025060610845685005,
0.05601714178919792,
0.05458827316761017,
0.07664617151021957,
-0.13084152340888977,
0.024587370455265045,
0.017005985602736473,
0.04399773105978966,
0.045803215354681015,
0.11189454793930054,
0.014143186621367931,
-0.014108255505561829,
0.03167976811528206,
0.03741208836436272,
0.0466543585062027,
-0.04269685223698616,
-0.012504290789365768,
0.054833557456731796,
0.053190916776657104,
-0.01193232461810112,
0.01679554581642151,
-0.03987216204404831,
-0.04302476719021797,
0.049165740609169006,
0.057631030678749084,
0.03120635636150837,
0.08810409903526306,
-0.0027152036782354116,
-0.05660075694322586,
-0.007556052878499031,
-0.05153454467654228,
0.041582826524972916,
-0.013189841993153095,
-0.003381364978849888,
-0.0014221188612282276,
0.03544822335243225,
0.09100443869829178,
0.0341525599360466,
-0.004246963188052177,
0.035775259137153625,
-0.04361533746123314,
-0.004781530238687992,
-0.027328073978424072,
-0.0920526385307312,
0.07534462213516235,
0.057904332876205444,
0.02642395906150341,
0.05130968987941742,
-0.0473112016916275,
0.05515648424625397,
0.050460297614336014,
-0.04477875679731369,
6.3452211601127805e-34,
0.08951914310455322,
0.06259368360042572,
-0.022675063461065292,
0.06716016680002213,
0.05884024128317833,
-0.041557226330041885,
-0.006848267745226622,
0.030280277132987976,
0.005636159330606461,
-0.05436839535832405,
-0.0060792481526732445,
0.08222796022891998,
-0.01836656965315342,
0.022956738248467445,
0.060519810765981674,
0.030721427872776985,
-0.03419593721628189,
0.09035886824131012,
-0.025396963581442833,
0.035253893584012985,
0.013024341315031052,
-0.09608732908964157,
0.03347543627023697,
-0.02794233150780201,
0.013012460432946682,
0.0034135859459638596,
0.0029998659156262875,
-0.038493093103170395,
-0.07277481257915497,
0.03020104020833969,
-0.11863598972558975,
-0.04255431145429611,
0.01797615736722946,
-0.0017451776657253504,
0.028491130098700523,
-0.07509498298168182,
-0.0150604872033,
-0.026018209755420685,
0.013993612490594387,
-0.054874613881111145,
-0.01977628841996193,
0.03807976469397545,
0.019358495250344276,
0.027218040078878403,
-0.00017912642215378582,
-0.03863121569156647,
-0.017926696687936783,
-0.025592608377337456,
-0.0008173107635229826,
-0.050108522176742554,
0.052653852850198746,
0.0709700956940651,
0.0356229767203331,
0.00934586487710476,
0.0029470042791217566,
0.09005168080329895,
0.033949676901102066,
-0.006603137589991093,
-0.023162895813584328,
0.0190287996083498,
0.05116211622953415,
0.03887869790196419,
-0.014979160390794277,
0.03531434386968613,
0.07633494585752487,
-0.011454390361905098,
-0.03143065795302391,
0.011782335117459297,
-0.009675423614680767,
-0.007981504313647747,
0.009328987449407578,
-0.05048637092113495,
-0.04891442134976387,
0.04745106026530266,
-0.009737999178469181,
-0.03800429031252861,
0.04695611447095871,
-0.03853990510106087,
0.043361369520425797,
-0.07105159014463425,
0.06055974215269089,
-0.023852413520216942,
0.043543875217437744,
-0.0766843929886818,
-0.048032574355602264,
-0.11041799187660217,
0.011917982250452042,
-0.054475087672472,
-0.02005668729543686,
0.05449794977903366,
0.010187983512878418,
-0.031048068776726723,
0.017747800797224045,
0.06851467490196228,
0.027245638892054558,
-2.9601991182047875e-33,
-0.10160722583532333,
-0.027467384934425354,
0.020018965005874634,
0.03223460912704468,
-0.06842363625764847,
-0.08321914821863174,
0.054187044501304626,
0.07953137904405594,
-0.012529833242297173,
-0.03310977295041084,
-0.05509728565812111,
-0.06601876020431519,
-0.032418083399534225,
0.04909732565283775,
-0.01920139417052269,
-0.05107302591204643,
0.03739389404654503,
0.034216005355119705,
0.06889224797487259,
0.035161618143320084,
0.06406594067811966,
0.1330515444278717,
-0.09785661846399307,
0.031639885157346725,
-0.02497241273522377,
-0.017541445791721344,
-0.06644878536462784,
0.01915356144309044,
0.022108078002929688,
-0.01898980140686035,
-0.047056589275598526,
0.10307714343070984,
-0.018571915104985237,
0.05238906294107437,
0.010105201974511147,
-0.0033687211107462645,
0.009403524920344353,
-0.055895861238241196,
0.0008763138903304935,
0.06424077600240707,
0.028158437460660934,
0.011739895679056644,
-0.06311994791030884,
-0.03756354749202728,
0.053995877504348755,
0.01763284206390381,
-0.07804924249649048,
-0.018111849203705788,
-0.05826208367943764,
-0.06597493588924408,
-0.005190958268940449,
-0.021710289642214775,
-0.004055608529597521,
0.045735228806734085,
-0.050717372447252274,
-0.06797689944505692,
-0.06538636237382889,
-0.11304295063018799,
-0.06167709827423096,
-0.005664965137839317,
-0.06301724910736084,
0.035563837736845016,
0.019166331738233566,
-0.03183051198720932,
-0.014911409467458725,
-0.0231150072067976,
0.03930939733982086,
-0.0001289638166781515,
0.015329107642173767,
-0.006926367059350014,
0.1348443329334259,
-0.09112567454576492,
-0.05029335245490074,
-0.009545482695102692,
0.0024288382846862078,
-0.0998370423913002,
-0.04871710017323494,
-0.11864954233169556,
-0.046549923717975616,
-0.06166842207312584,
0.001826109946705401,
-0.004649525973945856,
0.03659290447831154,
-0.006649933289736509,
-0.060166072100400925,
0.09227046370506287,
-0.02110501006245613,
0.10095434635877609,
0.011751742102205753,
0.009576444514095783,
-0.034864392131567,
-0.03491763770580292,
-0.09549605846405029,
-0.018936550244688988,
-0.061970748007297516,
-4.913349727075911e-8,
-0.015115111134946346,
0.02593926712870598,
-0.013416099362075329,
0.06367351859807968,
0.007006037048995495,
-0.0038306554779410362,
-0.09944751113653183,
-0.019734948873519897,
0.0500384159386158,
-0.04336332529783249,
-0.01209216844290495,
0.05070594325661659,
-0.01636689342558384,
-0.029062533751130104,
-0.012597509659826756,
0.0482761524617672,
-0.02434496581554413,
0.0882735401391983,
-0.019516263157129288,
0.02055816911160946,
0.006082302425056696,
0.11475998908281326,
0.013788879849016666,
-0.011054123751819134,
-0.020769119262695312,
-0.07684647291898727,
-0.0547032468020916,
0.04095287621021271,
0.08422575891017914,
-0.04880627989768982,
0.047592245042324066,
0.014292888343334198,
-0.008457171730697155,
0.02393226884305477,
0.007768080569803715,
0.05004604533314705,
-0.03907680884003639,
-0.03515394777059555,
0.09840815514326096,
0.08141448348760605,
0.05186382681131363,
0.06616799533367157,
-0.08649921417236328,
0.051754653453826904,
-0.09306132793426514,
-0.07393264025449753,
-0.07123664766550064,
-0.06993770599365234,
0.03542718663811684,
-0.13773423433303833,
0.07421279698610306,
-0.035497672855854034,
-0.05582841858267784,
0.06574686616659164,
-0.009997202083468437,
0.031004661694169044,
-0.0019502983195707202,
-0.006598967127501965,
0.0500359833240509,
0.00167614605743438,
0.1095736175775528,
-0.011643176898360252,
0.019444195553660393,
-0.04322751611471176
] | https://github.com/richterleo/Auditing_Test_for_LMs | 3 | 0 | 0 | 0 |
Radar: Fast Long-Context Decoding for Any Transformer | https://openreview.net/forum?id=ZTpWOwMrzQ | [
"Yongchang Hao",
"Mengyao Zhai",
"Hossein Hajimirsadeghi",
"Sepidehsadat Hosseini",
"Frederick Tung"
] | Poster | Transformer models have demonstrated exceptional performance across a wide range of applications. Though forming the foundation of Transformer models, the dot-product attention does not scale well to long-context data since its time requirement grows quadratically with context length. In this work, we propose Radar, a training-free approach that accelerates inference by dynamically searching for the most important context tokens. For any pre-trained Transformer, Radar can reduce the decoding time complexity without training or heuristically evicting tokens. Moreover, we provide theoretical justification for our approach, demonstrating that Radar can reliably identify the most important tokens with high probability. We conduct extensive comparisons with the previous methods on a wide range of tasks. The results demonstrate that Radar achieves the state-of-the-art performance across different architectures with reduced time complexity, offering a practical solution for efficient long-context processing of Transformers. The code is publicly available at https://github.com/BorealisAI/radar-decoding. | Long-context decoding, Large language models, Inference acceleration, Random features | null | 12,325 | 2503.10571 | [
-0.07520780712366104,
-0.012556030414998531,
-0.016462210565805435,
0.010821840725839138,
0.09521210193634033,
-0.06950700283050537,
0.0308068860322237,
-0.0072913821786642075,
-0.02435486763715744,
-0.05748656392097473,
-0.06852488219738007,
-0.004510614555329084,
-0.005886811763048172,
0.034070368856191635,
0.05441142991185188,
-0.01787734217941761,
0.004898969549685717,
0.06278485059738159,
-0.06425531953573227,
-0.061980899423360825,
0.0022987390402704477,
0.00001459619033994386,
0.016788626089692116,
-0.046427927911281586,
0.033970482647418976,
0.06483499705791473,
0.020607884973287582,
-0.09422352910041809,
-0.003281662706285715,
0.03216859698295593,
0.0619996078312397,
-0.005207972601056099,
-0.0012088778894394636,
0.019717523828148842,
-0.02789161168038845,
0.01663385145366192,
0.033382367342710495,
0.03605382516980171,
0.009364607743918896,
-0.03646951541304588,
0.05329909175634384,
0.02113308571279049,
-0.020482579246163368,
0.04526522383093834,
0.0141611872240901,
0.03902845457196236,
0.09119611978530884,
-0.037510160356760025,
0.01209288276731968,
-0.05899249389767647,
-0.03347252309322357,
0.0038319199811667204,
-0.0740584135055542,
0.055247317999601364,
-0.023381393402814865,
0.05015100911259651,
0.019658315926790237,
0.015652833506464958,
-0.020426908507943153,
0.04107924923300743,
-0.11095236241817474,
-0.07064471393823624,
0.010640066117048264,
-0.040901605039834976,
-0.09873434156179428,
0.010266452096402645,
0.06436662375926971,
0.018456442281603813,
-0.01641758158802986,
-0.004824476782232523,
-0.02323850244283676,
0.0853131040930748,
-0.03720491752028465,
0.0843513160943985,
0.026610927656292915,
-0.008030683733522892,
0.08979372680187225,
-0.041703589260578156,
0.01436570379883051,
-0.10803800821304321,
0.004992127418518066,
0.021658211946487427,
0.03467511013150215,
0.03356686979532242,
0.12918898463249207,
-0.02181166596710682,
-0.04429319500923157,
0.009160180576145649,
0.0008038158994168043,
0.049293942749500275,
-0.045069921761751175,
-0.15747523307800293,
0.017362669110298157,
-0.0299531240016222,
0.01692483387887478,
0.019306635484099388,
0.04824356362223625,
-0.054592106491327286,
-0.051626503467559814,
0.11953994631767273,
-0.044853709638118744,
-0.058537814766168594,
-0.003248062450438738,
-0.015532305464148521,
-0.0575333833694458,
-0.02779633179306984,
0.010956577956676483,
0.07536634057760239,
0.000775620574131608,
0.02596396766602993,
0.07475822418928146,
0.05213405191898346,
-0.024975771084427834,
-0.044046904891729355,
0.035124655812978745,
-0.013879794627428055,
-0.06417251378297806,
-0.05059407278895378,
-0.015772351995110512,
0.09648235142230988,
0.032024819403886795,
-0.021435320377349854,
-0.0712248682975769,
0.06232408806681633,
-0.017498470842838287,
0.017585165798664093,
-0.04753914847970009,
1.1725451456986698e-33,
-0.07800065726041794,
0.09708209335803986,
0.06241701915860176,
0.02329310029745102,
-0.031000468879938126,
0.01919955015182495,
0.07605064660310745,
0.07182315737009048,
0.05215322971343994,
0.08593575656414032,
0.044062141329050064,
0.043799880892038345,
-0.0661337822675705,
0.04755678027868271,
0.004765649326145649,
-0.03110840544104576,
-0.016521934419870377,
0.05006202682852745,
-0.04361949488520622,
-0.03938915580511093,
0.027000147849321365,
-0.045764047652482986,
-0.036069560796022415,
-0.035377275198698044,
-0.0013626384316012263,
-0.0360189750790596,
0.019398944452404976,
-0.07560080289840698,
0.027530405670404434,
-0.0024166738148778677,
-0.1028326004743576,
0.034078698605298996,
-0.005069539416581392,
0.021787350997328758,
0.07391467690467834,
-0.0918545350432396,
-0.014422775246202946,
-0.027828915044665337,
-0.05461383983492851,
-0.022483760491013527,
-0.017108388245105743,
-0.0023479792289435863,
-0.029008742421865463,
-0.03553992509841919,
-0.03853195905685425,
-0.1126367524266243,
-0.004473806358873844,
-0.029503080993890762,
-0.01011738833039999,
0.016488492488861084,
0.09534159302711487,
0.019195130094885826,
-0.0762774646282196,
0.009515326470136642,
0.1019580140709877,
-0.0034508586395531893,
0.07942639291286469,
0.010505400598049164,
0.06954088062047958,
0.04292551055550575,
-0.038829222321510315,
-0.061117298901081085,
-0.022099148482084274,
0.04515979066491127,
0.0015467464691028,
0.002128691878169775,
0.0038437233306467533,
0.056624893099069595,
-0.026292545720934868,
-0.02818489447236061,
0.007598972879350185,
-0.003085412783548236,
0.00182424730155617,
-0.05773396044969559,
0.042091190814971924,
0.04141627252101898,
0.04804854840040207,
0.06938576698303223,
-0.027811909094452858,
-0.015099196694791317,
0.036360934376716614,
0.034577369689941406,
-0.021077774465084076,
0.0022780559957027435,
-0.006240236572921276,
0.03505125641822815,
-0.0155577277764678,
-0.13421162962913513,
-0.0763729065656662,
-0.08511914312839508,
-0.06753211468458176,
0.019459394738078117,
-0.05421218276023865,
-0.04347417503595352,
-0.06961850821971893,
-9.2471512070183e-34,
0.010909746401011944,
0.042851824313402176,
-0.020881520584225655,
-0.011521384119987488,
-0.024223266169428825,
-0.08961515873670578,
-0.025329913944005966,
0.04722806438803673,
-0.014193468727171421,
-0.015327873639762402,
-0.07217471301555634,
-0.10878859460353851,
0.042259104549884796,
-0.02234327234327793,
0.07345254719257355,
0.03699769079685211,
0.015030773356556892,
0.04839801415801048,
-0.015697572380304337,
0.017762767150998116,
0.020252099260687828,
0.06530016660690308,
-0.07146124541759491,
-0.0356663316488266,
-0.08969534933567047,
-0.0063181109726428986,
-0.08599856495857239,
0.10719393193721771,
0.06734222173690796,
-0.026596827432513237,
-0.09217430651187897,
-0.022025875747203827,
0.029972629621624947,
-0.05430128425359726,
-0.03773147985339165,
0.02474633976817131,
0.016274938359856606,
0.02294689044356346,
-0.023750852793455124,
0.0060509066097438335,
0.04920857772231102,
-0.03347541391849518,
0.04054524376988411,
0.06504878401756287,
-0.08388011157512665,
-0.010536608286201954,
0.011035667732357979,
-0.019471438601613045,
-0.007260262034833431,
0.05044148862361908,
0.023713989183306694,
-0.01932671293616295,
-0.09113263338804245,
-0.0589260570704937,
-0.0125424200668931,
-0.045824117958545685,
-0.02222827635705471,
-0.032000213861465454,
0.036574240773916245,
-0.08678531646728516,
0.013555147685110569,
0.03211652860045433,
0.04921026527881622,
-0.026563262566924095,
0.057185374200344086,
-0.0347345769405365,
0.011458907276391983,
-0.06194864958524704,
0.019512401893734932,
0.08393386006355286,
0.024180052801966667,
-0.047164496034383774,
-0.051817070692777634,
-0.000924936612136662,
-0.05229863524436951,
0.034247156232595444,
0.05074988678097725,
-0.07966726273298264,
-0.04258972406387329,
-0.026192104443907738,
-0.023373126983642578,
-0.03639121353626251,
-0.05245158448815346,
-0.005581121426075697,
0.016463546082377434,
0.13615508377552032,
0.06966801732778549,
-0.020016053691506386,
0.05242045596241951,
0.07747973501682281,
-0.05286677926778793,
0.10266447067260742,
0.033714283257722855,
0.08772624284029007,
0.017809182405471802,
-3.769346790249983e-8,
-0.0925818681716919,
0.024338243529200554,
-0.06582111865282059,
-0.013390429317951202,
0.061807647347450256,
-0.0035905081313103437,
0.07383058965206146,
0.0958574041724205,
-0.039613574743270874,
-0.005413951352238655,
0.061951618641614914,
-0.10686636716127396,
-0.01964564435184002,
-0.027553588151931763,
-0.0006119953468441963,
0.055823616683483124,
-0.007335414178669453,
-0.04738136753439903,
-0.03370675817131996,
0.00423863809555769,
0.030427977442741394,
0.13879568874835968,
0.020687943324446678,
-0.022344043478369713,
0.006344054825603962,
-0.04856852814555168,
0.030034177005290985,
0.06710219383239746,
0.05080251768231392,
0.016962848603725433,
-0.0824522078037262,
0.050996143370866776,
-0.02274065464735031,
-0.032106947153806686,
0.004447635728865862,
0.006613411474972963,
0.11141063272953033,
-0.02038058452308178,
-0.0050273314118385315,
-0.039306677877902985,
0.08662282675504684,
0.08107095956802368,
-0.10559051483869553,
0.07266812771558762,
-0.008953929878771305,
0.031391751021146774,
0.01571541465818882,
-0.04953128099441528,
0.016558369621634483,
-0.006592568475753069,
0.06476610153913498,
-0.015046515502035618,
-0.07748449593782425,
-0.0010153643088415265,
-0.006364447996020317,
0.07675136625766754,
0.019957972690463066,
-0.055989645421504974,
0.02237468585371971,
0.12278115004301071,
-0.05889827013015747,
0.07418667525053024,
-0.026885777711868286,
0.053117536008358
] | https://github.com/BorealisAI/radar-decoding | 7 | 0 | 0 | 0 |
Can Knowledge Editing Really Correct Hallucinations? | https://openreview.net/forum?id=hmDt068MoZ | [
"Baixiang Huang",
"Canyu Chen",
"Xiongxiao Xu",
"Ali Payani",
"Kai Shu"
] | Poster | Large Language Models (LLMs) suffer from hallucinations, referring to the non-factual information in generated content, despite their superior capacities across tasks. Meanwhile, knowledge editing has been developed as a new popular paradigm to correct erroneous factual knowledge encoded in LLMs with the advantage of avoiding retraining from scratch. However, a common issue of existing evaluation datasets for knowledge editing is that they do not ensure that LLMs actually generate hallucinated answers to the evaluation questions before editing. When LLMs are evaluated on such datasets after being edited by different techniques, it is hard to directly adopt the performance to assess the effectiveness of different knowledge editing methods in correcting hallucinations. Thus, the fundamental question remains insufficiently validated: Can knowledge editing really correct hallucinations in LLMs? We proposed HalluEditBench to holistically benchmark knowledge editing methods in correcting real-world hallucinations. First, we rigorously construct a massive hallucination dataset with 9 domains, 26 topics and more than 6,000 hallucinations. Then, we assess the performance of knowledge editing methods in a holistic way on five dimensions including Efficacy, Generalization, Portability, Locality, and Robustness. Through HalluEditBench, we have provided new insights into the potentials and limitations of different knowledge editing methods in correcting hallucinations, which could inspire future improvements and facilitate progress in the field of knowledge editing. | LLMs, Knowledge Editing, Hallucination, Benchmark | We proposed HalluEditBench to holistically benchmark knowledge editing methods in correcting real-world hallucinations on five dimensions including Efficacy, Generalization, Portability, Locality, and Robustness. | 12,320 | 2410.16251 | [
0.012276945635676384,
-0.0053214943036437035,
-0.012329190038144588,
0.048093754798173904,
-0.00744227971881628,
0.03088264912366867,
-0.06224521994590759,
-0.0276891328394413,
0.01386923436075449,
-0.007073653396219015,
-0.07737646996974945,
-0.00017037047655321658,
0.06291675567626953,
0.01782001182436943,
-0.010989741422235966,
0.05117719620466232,
0.04599735140800476,
0.07591556012630463,
-0.08417560905218124,
-0.030696304515004158,
0.03918968513607979,
0.028732815757393837,
0.07502980530261993,
-0.014998416416347027,
0.05448205769062042,
-0.004098175093531609,
-0.011960593983530998,
-0.029432836920022964,
-0.018648840487003326,
-0.04167841002345085,
-0.01018450502306223,
0.12596546113491058,
-0.03454611450433731,
0.007150198798626661,
-0.03527907282114029,
0.07903960347175598,
-0.04620055481791496,
0.10229792445898056,
0.01957869529724121,
-0.05028272047638893,
-0.013505293056368828,
-0.0013089505955576897,
-0.0018217201577499509,
0.03334188088774681,
0.02856285125017166,
-0.00617782911285758,
-0.06256787478923798,
0.037116020917892456,
-0.06363237649202347,
0.012072952464222908,
-0.07295327633619308,
-0.05974292382597923,
-0.015073109418153763,
0.04676184058189392,
-0.04653746634721756,
0.025553466752171516,
0.019955415278673172,
0.03671357035636902,
-0.003427523421123624,
-0.07797524333000183,
-0.0471685416996479,
-0.04974234476685524,
-0.030950570479035378,
-0.007524126209318638,
-0.023086300119757652,
-0.010882114991545677,
-0.024720000103116035,
0.031212393194437027,
-0.020434819161891937,
0.0012040071887895465,
-0.005571644753217697,
0.01789749041199684,
-0.05441254749894142,
0.0815320685505867,
-0.015711411833763123,
0.0005996034597046673,
-0.03189218044281006,
-0.052979253232479095,
-0.00403447775170207,
-0.0800861120223999,
0.029495304450392723,
0.055347103625535965,
0.03779846802353859,
-0.00011969641491305083,
0.06438729912042618,
-0.007416362874209881,
0.03148370981216431,
0.05287369340658188,
-0.04851946979761124,
0.0660015344619751,
-0.009144086390733719,
-0.09722336381673813,
0.04537537693977356,
-0.03698427975177765,
0.06497018039226532,
0.029105503112077713,
-0.009870342910289764,
-0.024806102737784386,
-0.016191653907299042,
0.020158419385552406,
0.049001775681972504,
0.06539978086948395,
-0.07225129753351212,
-0.1292538344860077,
-0.013800187036395073,
-0.052213240414857864,
0.07532735168933868,
0.0739966407418251,
0.010591701604425907,
-0.07428847253322601,
0.0029087357688695192,
0.06676431000232697,
0.011338437907397747,
-0.0516451857984066,
0.024269048124551773,
-0.019592225551605225,
0.014781509526073933,
0.041309501975774765,
-0.026066679507493973,
0.07486651092767715,
0.02825099602341652,
0.049644626677036285,
0.14630568027496338,
0.03464670479297638,
-0.0103292316198349,
-0.06904832273721695,
-0.07248318195343018,
2.272181907326518e-33,
0.09004167467355728,
0.045590028166770935,
-0.009518882259726524,
0.03632761165499687,
0.05270583555102348,
0.02878185361623764,
-0.08783100545406342,
-0.008606623858213425,
0.05366111919283867,
-0.03455333039164543,
-0.006723810452967882,
0.013915425166487694,
-0.015399807132780552,
0.010745454579591751,
0.04937093332409859,
0.09994463622570038,
-0.09452390670776367,
0.04635319113731384,
-0.09631505608558655,
0.010298294946551323,
0.06608662754297256,
-0.009342154487967491,
-0.03668355569243431,
-0.024715695530176163,
-0.0576145201921463,
0.04174220189452171,
0.03154364228248596,
-0.02931966818869114,
0.004181355703622103,
0.007131188176572323,
-0.112450510263443,
0.0003908534999936819,
0.043223101645708084,
0.02077450230717659,
0.005437775515019894,
-0.003427586518228054,
0.03425950929522514,
-0.09608878195285797,
0.027809884399175644,
-0.0542016439139843,
0.0022939234040677547,
0.047906458377838135,
-0.02076001651585102,
0.015008107759058475,
-0.0549420565366745,
0.010713424533605576,
0.01342522818595171,
-0.0078134061768651,
-0.062322188168764114,
-0.03193311393260956,
0.07881899178028107,
0.060659125447273254,
-0.011289886198937893,
0.033126961439847946,
-0.049866050481796265,
0.058166176080703735,
0.04800198972225189,
-0.06935881823301315,
0.06074732542037964,
0.048306357115507126,
0.09681368619203568,
0.1155429556965828,
-0.018473893404006958,
0.06691557168960571,
-0.016657941043376923,
0.010212114080786705,
-0.003713370766490698,
0.01727687008678913,
0.016906434670090675,
-0.12864497303962708,
-0.09507569670677185,
-0.06380167603492737,
-0.008249280042946339,
0.022570282220840454,
-0.02456032484769821,
-0.07151699811220169,
0.0029633138328790665,
-0.020447131246328354,
-0.025949331000447273,
-0.018742775544524193,
0.053449392318725586,
-0.03716907650232315,
-0.038878463208675385,
-0.07339102774858475,
0.00040984281804412603,
-0.09616737067699432,
0.0024308946449309587,
-0.05200323835015297,
-0.07937812805175781,
-0.038817036896944046,
0.044124457985162735,
-0.04326805844902992,
0.015095793642103672,
-0.023979434743523598,
-0.027674244716763496,
-2.991884935773187e-33,
-0.0590926818549633,
-0.08123776316642761,
0.03816601634025574,
0.08657388389110565,
-0.03894956409931183,
-0.06439884006977081,
-0.04714033752679825,
0.013835828751325607,
0.008134318515658379,
-0.11442246288061142,
-0.0341389961540699,
-0.02142777480185032,
-0.00016691218479536474,
-0.04734981060028076,
-0.02729347161948681,
-0.07467028498649597,
0.002227845136076212,
0.019408943131566048,
-0.04950481653213501,
0.08183291554450989,
-0.036748167127370834,
0.031195063143968582,
-0.09876667708158493,
0.04828806221485138,
-0.028762437403202057,
0.07399406284093857,
0.017042484134435654,
0.02132445015013218,
0.020959196612238884,
0.005752337630838156,
-0.008709621615707874,
0.018338516354560852,
-0.042324431240558624,
-0.01599816232919693,
-0.034380100667476654,
0.002602308988571167,
0.06611713767051697,
-0.046951789408922195,
-0.0923105999827385,
-0.018458735197782516,
0.08432632684707642,
0.03457803279161453,
-0.059423964470624924,
0.016717754304409027,
-0.018569784238934517,
-0.01778947003185749,
-0.11916685104370117,
-0.01272793859243393,
0.04447139427065849,
0.023081809282302856,
0.004818271845579147,
-0.052191030234098434,
-0.07101548463106155,
-0.014319191686809063,
0.015250389464199543,
-0.10503536462783813,
-0.08165614306926727,
-0.054033584892749786,
0.0741470530629158,
0.0024594010319560766,
-0.09516553580760956,
0.08063188940286636,
-0.015620224177837372,
-0.013193625025451183,
0.047865353524684906,
0.0027168297674506903,
-0.07777074724435806,
0.09285883605480194,
0.02372124418616295,
0.022421108558773994,
-0.018216002732515335,
-0.025395069271326065,
-0.0060261813923716545,
0.016961773857474327,
0.05006937310099602,
-0.050185032188892365,
0.04305833950638771,
-0.07010256499052048,
-0.03348172828555107,
-0.09502812474966049,
-0.02673099935054779,
-0.035943157970905304,
0.02176334336400032,
0.07446442544460297,
0.03539721295237541,
0.029436180368065834,
0.029771167784929276,
0.10431157052516937,
-0.017505258321762085,
0.047855228185653687,
-0.024063117802143097,
0.02678968571126461,
-0.005387228913605213,
0.04767123609781265,
-0.06556760519742966,
-3.799478420773994e-8,
-0.02237439714372158,
-0.00362710515037179,
-0.004716728813946247,
0.05930976942181587,
0.025146404281258583,
-0.08821411430835724,
-0.05739946290850639,
0.12083223462104797,
-0.026723137125372887,
-0.09968241304159164,
0.07656987756490707,
0.02122993767261505,
-0.032137252390384674,
-0.06759537756443024,
-0.0057372963055968285,
0.06113012507557869,
0.013855397701263428,
0.1215658038854599,
-0.007601005025207996,
-0.02782581001520157,
0.10927043855190277,
0.06772845983505249,
0.003709308337420225,
-0.04732318967580795,
0.06991198658943176,
-0.0026880933437496424,
-0.006321700755506754,
0.08088168501853943,
-0.030916549265384674,
0.005709272809326649,
0.05653418228030205,
-0.01918611116707325,
0.047999829053878784,
0.09237440675497055,
0.030141746625304222,
-0.01323248352855444,
0.016988618299365044,
-0.0054898555390536785,
0.07931698858737946,
-0.04501008614897728,
0.0632629245519638,
0.01787586882710457,
0.003006819635629654,
0.0643327534198761,
-0.04125456511974335,
-0.11318723112344742,
-0.0066443332470953465,
-0.046228643506765366,
0.0300923939794302,
-0.06888582557439804,
-0.020999621599912643,
0.07524586468935013,
0.025042811408638954,
0.02473895438015461,
0.10342148691415787,
0.10071709752082825,
0.022549524903297424,
0.020911796018481255,
0.0117271663621068,
-0.06332940608263016,
0.09080307930707932,
-0.0252990759909153,
-0.03791900351643562,
0.04248150438070297
] | https://github.com/llm-editing/HalluEditBench | 13 | 0 | 0 | 0 |
SysCaps: Language Interfaces for Simulation Surrogates of Complex Systems | https://openreview.net/forum?id=1R5BcYS8EC | [
"Patrick Emami",
"Zhaonan Li",
"Saumya Sinha",
"Truc Nguyen"
] | Poster | Surrogate models are used to predict the behavior of complex energy systems that are too expensive to simulate with traditional numerical methods.
Our work introduces the use of language descriptions, which we call "system captions" or SysCaps, to interface with such surrogates.
We argue that interacting with surrogates through text, particularly natural language, makes these models more accessible for both experts and non-experts.
We introduce a lightweight multimodal text and timeseries regression model and a training pipeline that uses large language models (LLMs) to synthesize high-quality captions from simulation metadata.
Our experiments on two real-world simulators of buildings and wind farms show that our SysCaps-augmented surrogates have better accuracy on held-out systems than traditional methods while enjoying new generalization abilities, such as handling semantically related descriptions of the same test system.
Additional experiments also highlight the potential of SysCaps to unlock language-driven design space exploration and to regularize training through prompt augmentation. | surrogate models, multimodal text and timeseries models, language-interfaced regression | Augmenting surrogate regression models of complex systems simulations with natural language interfaces. | 12,317 | 2405.19653 | [
-0.0421505831182003,
-0.03946148604154587,
0.04023328423500061,
0.046007368713617325,
0.07075904309749603,
-0.10547938197851181,
-0.009349573403596878,
0.013174531981348991,
0.035965174436569214,
-0.007093263790011406,
-0.06108107417821884,
-0.12007459998130798,
0.010473194532096386,
0.01453347783535719,
0.04646948724985123,
-0.0518198162317276,
0.03487908095121384,
0.039782002568244934,
-0.017525697126984596,
-0.0074362135492265224,
0.1133686900138855,
0.04783836379647255,
0.020960284397006035,
-0.005279860459268093,
0.085137739777565,
0.052789170295000076,
-0.07627331465482712,
0.024151049554347992,
0.02284695766866207,
-0.0011629884829744697,
0.019586343318223953,
0.006847048178315163,
-0.018833059817552567,
-0.0032351261470466852,
0.0006890791701152921,
0.08818650245666504,
-0.005369469989091158,
-0.02395617961883545,
-0.011776745319366455,
-0.07344718277454376,
-0.12213967740535736,
-0.05286451056599617,
0.021919872611761093,
0.009599626064300537,
0.10141950845718384,
0.00932725053280592,
0.03532969206571579,
-0.06929919868707657,
-0.025588274002075195,
0.014964453876018524,
-0.1114838644862175,
-0.08519555628299713,
0.0680684968829155,
-0.02580367401242256,
-0.00479382649064064,
-0.004949013702571392,
0.08682362735271454,
-0.03977048397064209,
0.017289822921156883,
-0.04099145531654358,
-0.05440596491098404,
-0.10030331462621689,
-0.028408760204911232,
0.015003505162894726,
-0.021250970661640167,
0.01882227323949337,
0.03260476887226105,
0.08529134094715118,
0.010166524909436703,
0.003400401910766959,
-0.06014588475227356,
0.020209074020385742,
0.01908361166715622,
0.02294672466814518,
-0.009434201754629612,
0.054665230214595795,
-0.015488538891077042,
-0.08573544770479202,
0.09279092401266098,
-0.0969119593501091,
0.052269138395786285,
-0.027380947023630142,
0.013573586009442806,
0.031054597347974777,
0.06450556963682175,
0.02047763019800186,
0.039931509643793106,
0.08641320466995239,
0.05243665352463722,
0.012457120232284069,
-0.02854173071682453,
-0.06909991055727005,
0.08194831013679504,
0.03615759685635567,
0.025446826592087746,
0.05158226564526558,
0.01958935521543026,
-0.09796121716499329,
-0.01939861848950386,
0.08906899392604828,
0.01864362321794033,
0.0625552088022232,
0.03373152017593384,
-0.010288489982485771,
0.016937296837568283,
-0.03233328089118004,
0.016076378524303436,
-0.008215172216296196,
0.018126370385289192,
-0.06097732111811638,
0.013506803661584854,
0.05915517359972,
-0.051805462688207626,
0.02988400310277939,
0.04579910635948181,
-0.040104422718286514,
0.039184585213661194,
-0.11505488306283951,
0.021383579820394516,
0.05138177424669266,
-0.017034796997904778,
0.031071219593286514,
-0.010005425661802292,
0.043834079056978226,
0.1048893928527832,
0.014735348522663116,
-0.005202779546380043,
5.562310698396443e-34,
-0.0038370799738913774,
0.03479085862636566,
0.048503272235393524,
0.11701612919569016,
0.14267286658287048,
-0.06453970819711685,
0.007688217796385288,
0.02390645444393158,
-0.03741974011063576,
0.012103429064154625,
-0.011417062021791935,
0.040136467665433884,
-0.011906618252396584,
0.08282306790351868,
0.066203773021698,
0.028100593015551567,
-0.04871034994721413,
0.08204484730958939,
-0.047372642904520035,
-0.02119274064898491,
-0.016743624582886696,
-0.023889295756816864,
-0.00148819200694561,
-0.04952937364578247,
0.06861481815576553,
0.02970646694302559,
0.06712879985570908,
-0.01339472085237503,
0.008175150491297245,
-0.002294721780344844,
-0.06546775996685028,
0.01599128544330597,
0.026163648813962936,
0.005202633794397116,
0.022991163656115532,
-0.028075985610485077,
0.05196528881788254,
-0.0653393566608429,
-0.012346116825938225,
0.01503802090883255,
-0.025513552129268646,
-0.001342588453553617,
-0.03326430544257164,
0.034141190350055695,
-0.0439319834113121,
-0.03850170597434044,
-0.007914516143500805,
0.01642731763422489,
-0.0760623887181282,
-0.07056155055761337,
-0.009282462298870087,
0.07775650173425674,
-0.028886541724205017,
-0.03108322061598301,
0.003485986264422536,
-0.04585579037666321,
0.05762946233153343,
-0.014365985058248043,
0.006046175956726074,
0.007064338773488998,
-0.0653597041964531,
0.04363614320755005,
0.0239982008934021,
0.029298841953277588,
-0.007529245689511299,
0.045916493982076645,
0.002244624076411128,
0.038252439349889755,
0.05173679441213608,
0.013823446817696095,
-0.06465752422809601,
-0.015457418747246265,
-0.05062519758939743,
0.01975901797413826,
0.05736047402024269,
0.02057753875851631,
0.05052297189831734,
-0.016694337129592896,
-0.0320560559630394,
0.07061304152011871,
0.061001457273960114,
-0.03657301515340805,
0.01663445308804512,
-0.03172582387924194,
0.0070769949816167355,
-0.1528214067220688,
-0.017482569441199303,
-0.0904526486992836,
-0.02787182666361332,
-0.11670105904340744,
-0.07899042218923569,
-0.03387049585580826,
-0.052877943962812424,
-0.05134480446577072,
-0.06014144420623779,
-1.7688752665807394e-33,
-0.038390930742025375,
0.023590734228491783,
-0.07860873639583588,
0.06858693063259125,
-0.08614227175712585,
-0.0581846721470356,
-0.03857811167836189,
-0.023761775344610214,
0.00918289739638567,
-0.08450574427843094,
-0.06070612370967865,
-0.054418593645095825,
0.052653346210718155,
-0.01240260899066925,
0.05951584130525589,
-0.025818655267357826,
-0.03951448202133179,
-0.07923626899719238,
-0.017274295911192894,
0.01573312282562256,
0.008130147121846676,
0.0982074961066246,
-0.040105003863573074,
0.00820788461714983,
-0.044579293578863144,
0.00787447951734066,
-0.017929229885339737,
0.04719628766179085,
0.031240247189998627,
-0.037530992180109024,
-0.01768689975142479,
-0.026139115914702415,
0.045634571462869644,
-0.021653391420841217,
-0.009520015679299831,
0.07275162637233734,
0.10386359691619873,
-0.0484759584069252,
-0.031576693058013916,
0.05159816890954971,
0.0991135984659195,
-0.05556758865714073,
-0.0484682098031044,
-0.03221091255545616,
-0.0775003656744957,
-0.02761780470609665,
-0.07231755554676056,
-0.05582600086927414,
-0.02545170858502388,
-0.03592206910252571,
0.01761852577328682,
-0.0373530350625515,
-0.07020977884531021,
-0.02591114118695259,
-0.032282765954732895,
-0.11689440160989761,
-0.06293357163667679,
-0.07049067318439484,
0.008884289301931858,
-0.03161978721618652,
-0.02854938432574272,
-0.0470857247710228,
0.03577767685055733,
-0.0141092948615551,
-0.021825341507792473,
-0.11340096592903137,
-0.06293022632598877,
-0.06568153202533722,
-0.015550903044641018,
-0.033634252846241,
0.030897032469511032,
-0.030837204307317734,
-0.036666691303253174,
0.06022053584456444,
-0.02068515308201313,
0.037091128528118134,
0.021456502377986908,
-0.04227486252784729,
-0.031460825353860855,
-0.08707533031702042,
0.053934987634420395,
0.014088934287428856,
0.08518059551715851,
0.05323309451341629,
0.047444071620702744,
0.06389936804771423,
0.00197543459944427,
0.11832916736602783,
0.02412775717675686,
0.018780434504151344,
-0.052348438650369644,
0.009858379140496254,
0.0007961403462104499,
0.17030580341815948,
-0.05493221431970596,
-3.9859592959601287e-8,
-0.048154983669519424,
0.10715967416763306,
0.05314832925796509,
0.016574885696172714,
-0.02103697508573532,
0.053345583379268646,
-0.04105665534734726,
-0.01419630367308855,
-0.020732801407575607,
-0.03365792706608772,
0.051679208874702454,
0.00531636131927371,
0.022309012711048126,
-0.012957043945789337,
-0.005156739614903927,
0.09843919426202774,
-0.00369349610991776,
0.021038787439465523,
-0.012832843698561192,
0.003022256074473262,
0.04080357030034065,
0.08195837587118149,
-0.008262824267148972,
-0.024882888421416283,
0.06060563772916794,
0.008452396839857101,
-0.061567578464746475,
0.060746513307094574,
0.09576503932476044,
0.003065279917791486,
0.031982772052288055,
-0.0112903518602252,
-0.0017842751694843173,
0.014227142557501793,
0.013334881514310837,
0.06070283427834511,
0.029394950717687607,
-0.029118528589606285,
0.04334413632750511,
-0.003236762946471572,
0.017979077994823456,
-0.007280711550265551,
-0.10926235467195511,
0.0401557981967926,
0.02053588628768921,
0.03139113634824753,
-0.10980366915464401,
-0.11395296454429626,
0.0527631938457489,
-0.017095332965254784,
0.02908731810748577,
-0.025977719575166702,
-0.026948638260364532,
0.0959223136305809,
0.04484613612294197,
0.017357217147946358,
-0.03456408903002739,
-0.0019045722438022494,
0.0574205219745636,
0.07632867991924286,
0.03469746559858322,
0.01779775507748127,
-0.06466878950595856,
0.023730143904685974
] | 0 | 0 | 0 | 0 |
|
Generative Classifiers Avoid Shortcut Solutions | https://openreview.net/forum?id=oCUYc7BzXQ | [
"Alexander Cong Li",
"Ananya Kumar",
"Deepak Pathak"
] | Poster | Discriminative approaches to classification often learn shortcuts that hold in-distribution but fail even under minor distribution shift. This failure mode stems from an overreliance on features that are spuriously correlated with the label. We show that generative classifiers, which use class-conditional generative models, can avoid this issue by modeling all features, both core and spurious, instead of mainly spurious ones. These generative classifiers are simple to train, avoiding the need for specialized augmentations, strong regularization, extra hyperparameters, or knowledge of the specific spurious correlations to avoid. We find that diffusion-based and autoregressive generative classifiers achieve state-of-the-art performance on five standard image and text distribution shift benchmarks and reduce the impact of spurious correlations in realistic applications, such as medical or satellite datasets. Finally, we carefully analyze a Gaussian toy setting to understand the inductive biases of generative classifiers, as well as the data properties that determine when generative classifiers outperform discriminative ones. | distribution shift, shortcut, generative models, robustness | We show that generative classifiers are more robust to realistic distribution shifts than discriminative classifiers. | 12,315 | null | [
-0.046222686767578125,
-0.12828049063682556,
0.09342047572135925,
0.02805514819920063,
0.07752150297164917,
0.030026869848370552,
-0.0009444355964660645,
-0.014764238148927689,
-0.028377778828144073,
-0.009605667553842068,
-0.017259087413549423,
-0.043592892587184906,
0.006985352840274572,
0.023779958486557007,
-0.05635010451078415,
-0.08295033872127533,
0.10784263163805008,
0.014012778177857399,
-0.08600233495235443,
-0.019399816170334816,
-0.015340348705649376,
0.031589165329933167,
-0.016459409147500992,
0.03071434237062931,
0.0853554904460907,
0.023506930097937584,
0.025461435317993164,
-0.04994877427816391,
-0.007569877430796623,
-0.038318436592817307,
0.030249258503317833,
0.010592270642518997,
-0.10101689398288727,
0.03201654553413391,
-0.033257268369197845,
0.024729862809181213,
-0.005052550230175257,
0.09859615564346313,
0.061822835355997086,
0.024444980546832085,
0.047675926238298416,
-0.06807966530323029,
0.03090483695268631,
0.02849089540541172,
0.05038503184914589,
0.009814117103815079,
-0.011080695316195488,
-0.03718535974621773,
0.015054845251142979,
0.023146096616983414,
-0.06899812072515488,
-0.007122804876416922,
-0.08006112277507782,
0.08345548063516617,
-0.03466774895787239,
0.019918929785490036,
0.07071930915117264,
0.04765420779585838,
-0.009745389223098755,
0.0164114311337471,
-0.03121081180870533,
-0.12257608026266098,
-0.009227062575519085,
-0.04137507453560829,
-0.011066044680774212,
-0.031109174713492393,
0.05621635913848877,
0.04718613997101784,
-0.007052560802549124,
0.05641116201877594,
0.00038025228423066437,
0.11903363466262817,
-0.04527609050273895,
0.0870329886674881,
0.018461301922798157,
0.023741986602544785,
0.01902749575674534,
-0.010224933736026287,
0.09917527437210083,
-0.12732040882110596,
0.021805942058563232,
0.0474613755941391,
0.04032033681869507,
-0.00043154723243787885,
0.15123090147972107,
0.01934215798974037,
-0.0068296645767986774,
0.07991451025009155,
-0.07048080861568451,
-0.029083697125315666,
0.03185705840587616,
-0.04973463714122772,
0.018932152539491653,
-0.02423221431672573,
-0.027868248522281647,
0.024005940183997154,
0.01620059460401535,
-0.10418782383203506,
0.0033244260121136904,
0.06863473355770111,
-0.03430679813027382,
0.0020782037172466516,
0.007413797080516815,
-0.043471481651067734,
0.04574025049805641,
-0.1105109229683876,
0.035995952785015106,
0.006333900615572929,
0.03586507961153984,
-0.03741367533802986,
-0.0014253725530579686,
0.042875442653894424,
-0.061928294599056244,
-0.03681031987071037,
-0.05424872413277626,
0.01052173227071762,
-0.00894419476389885,
-0.009864247404038906,
-0.011799455620348454,
0.014187045395374298,
-0.06878472119569778,
0.016067374497652054,
-0.0698126032948494,
0.021118411794304848,
-0.01468726433813572,
-0.018820814788341522,
-0.0327921062707901,
4.226702996732621e-33,
0.035984739661216736,
-0.04817603528499603,
-0.005739166401326656,
0.04949801787734032,
0.018600625917315483,
-0.013020503334701061,
-0.0802755355834961,
-0.007293172180652618,
0.02691642940044403,
-0.01553579792380333,
-0.03139929100871086,
0.009037705138325691,
-0.04622985050082207,
0.10713686794042587,
0.04602856934070587,
0.04799967631697655,
-0.09082838147878647,
0.057412125170230865,
0.0037078887689858675,
-0.023312576115131378,
-0.008969665504992008,
-0.06553635746240616,
-0.04802679270505905,
-0.00002852850047929678,
-0.07490114122629166,
0.013158435001969337,
0.004903722554445267,
-0.06015815958380699,
-0.02510143630206585,
0.010789182037115097,
-0.10205601155757904,
-0.012496113777160645,
0.08031201362609863,
-0.029519598931074142,
0.011552161537110806,
-0.06386664509773254,
-0.0037299252580851316,
-0.031438302248716354,
0.010189346969127655,
0.005013280548155308,
-0.10001525282859802,
0.056747350841760635,
-0.03769153729081154,
0.0312853679060936,
-0.02397497557103634,
-0.027484826743602753,
0.09485094994306564,
-0.04392442852258682,
-0.015724269673228264,
-0.07171344012022018,
0.03035869635641575,
0.005265206564217806,
-0.0006253303145058453,
-0.04897841811180115,
-0.05213282257318497,
0.021862411871552467,
0.04713268578052521,
-0.014245509169995785,
-0.00017779655172489583,
-0.043470438569784164,
0.036353081464767456,
0.08069863170385361,
0.030916457995772362,
0.03929153084754944,
-0.021662063896656036,
-0.027947168797254562,
0.022725297138094902,
0.04344119131565094,
0.03683676943182945,
0.10015083849430084,
0.001525483443401754,
0.022985149174928665,
-0.09544900059700012,
-0.12149452418088913,
0.11255747824907303,
0.06398729234933853,
0.002875083591789007,
-0.02874108776450157,
-0.0062448107637465,
0.0007296037510968745,
-0.04836757481098175,
-0.04002952203154564,
0.055969320237636566,
-0.022432249039411545,
-0.038206588476896286,
-0.018701553344726562,
-0.02335282228887081,
-0.07510899007320404,
-0.06269526481628418,
-0.00009552228584652767,
-0.049357954412698746,
0.03366333618760109,
-0.0045723626390099525,
0.0690200924873352,
0.029680009931325912,
-3.983454277297119e-33,
-0.08398246765136719,
0.061022378504276276,
0.026821119710803032,
0.06518552452325821,
-0.025578271597623825,
0.05480645224452019,
-0.017233243212103844,
0.07573992758989334,
-0.011677421629428864,
-0.07440727949142456,
0.034461475908756256,
0.006149265915155411,
-0.022921400144696236,
0.028866101056337357,
-0.02404988370835781,
-0.02292618341743946,
-0.0016592495376244187,
0.029167087748646736,
-0.023662466555833817,
0.044603437185287476,
0.054216329008340836,
0.08160332590341568,
-0.07700028270483017,
0.03902432695031166,
-0.049999479204416275,
0.07133842259645462,
-0.05430031567811966,
0.07298791408538818,
-0.04150610789656639,
-0.0635496973991394,
-0.025219382718205452,
0.037711042910814285,
-0.030834658071398735,
-0.02954183705151081,
-0.1107528880238533,
0.09612299501895905,
-0.03308068588376045,
-0.0053799329325556755,
-0.07979460805654526,
0.0930076465010643,
0.07791896909475327,
-0.017923688516020775,
-0.05495520308613777,
0.009202246554195881,
-0.029615294188261032,
-0.10797025263309479,
0.010183294303715229,
0.06452587991952896,
0.09524630755186081,
0.08915556222200394,
-0.011929393745958805,
0.030704112723469734,
-0.11557448655366898,
0.09413972496986389,
-0.018087314441800117,
-0.0680098682641983,
-0.10515105724334717,
-0.06561938673257828,
0.018159056082367897,
0.031597260385751724,
-0.06128299608826637,
-0.023374641314148903,
0.04720654711127281,
-0.04462536424398422,
-0.0013040457852184772,
-0.0692368820309639,
0.012499247677624226,
0.0123183848336339,
0.029307572171092033,
0.07439818233251572,
0.10552672296762466,
0.0402885265648365,
0.026090996339917183,
-0.026740549132227898,
-0.03538630157709122,
0.012865680269896984,
0.04864494875073433,
0.0182422436773777,
-0.026693833991885185,
-0.011595580726861954,
0.08116582036018372,
-0.04418507590889931,
-0.002869231626391411,
0.06237033009529114,
-0.0448913536965847,
0.1025797575712204,
0.02464916929602623,
-0.012975812889635563,
0.08655063807964325,
0.003716363338753581,
-0.034157369285821915,
0.01178851816803217,
-0.018496107310056686,
0.03525292128324509,
-0.04323368892073631,
-4.4338037241686834e-8,
-0.029993198812007904,
0.04003937542438507,
-0.06191776320338249,
0.027898233383893967,
0.011747363023459911,
-0.06066347658634186,
-0.06011795997619629,
0.1642572283744812,
-0.017808780074119568,
-0.034887995570898056,
0.08768577128648758,
-0.0265701562166214,
-0.029433008283376694,
0.019114172086119652,
0.035689517855644226,
0.010949972085654736,
0.02672325074672699,
0.058368101716041565,
-0.0202634297311306,
0.02323259972035885,
-0.01949402317404747,
-0.04549337550997734,
0.05261462926864624,
-0.09934365749359131,
0.020055832341313362,
-0.04882853478193283,
-0.0050680628046393394,
-0.004545555915683508,
0.09154048562049866,
-0.0071228742599487305,
-0.029140282422304153,
0.0399390384554863,
-0.02842927724123001,
-0.017620772123336792,
0.028751352801918983,
0.13277681171894073,
0.03718693181872368,
-0.019979560747742653,
-0.035181038081645966,
0.037348248064517975,
0.03158104419708252,
0.027208583429455757,
-0.03140488639473915,
0.036615580320358276,
-0.02798132412135601,
-0.004045369569212198,
0.1072627529501915,
-0.037165481597185135,
0.03493719920516014,
0.010082978755235672,
0.00507223280146718,
-0.08352871239185333,
0.024389488622546196,
0.0389995239675045,
-0.010508101433515549,
-0.015706229954957962,
0.0366063117980957,
-0.015186739154160023,
0.027872974053025246,
0.06767783313989639,
-0.010234858840703964,
-0.013896908611059189,
-0.057952336966991425,
0.05049176141619682
] | 0 | 0 | 0 | 0 |
|
Compute-Constrained Data Selection | https://openreview.net/forum?id=4es2oO9tw1 | [
"Junjie Yin",
"Alexander M Rush"
] | Poster | Data selection can reduce the amount of training data needed to finetune LLMs; however, the efficacy of data selection scales directly with its compute. Motivated by the practical challenge of compute-constrained finetuning, we consider the setting in which both the cost of selecting data and training are budgeted for. We first formalize the problem of data selection with a cost-aware utility function,
and model the data selection problem as trading off initial-selection cost for training gain. We run a comprehensive sweep of experiments across multiple tasks, varying compute budget by scaling finetuning tokens, model sizes, and data selection compute. Interestingly we find that many powerful data selection methods are almost never compute-optimal, and that cheaper data selection alternatives dominate both from a theoretical and empirical perspective. For compute-optimal training, we find that perplexity and gradient data selection require training-to-selection model size ratios of 5x and 10x, respectively. | Data Selection, Scaling Laws, Compute-constrained, Compute-optimal Training. | Post-training data selection in a scaling laws framework. | 12,314 | 2410.16208 | [
-0.046912435442209244,
-0.05481329560279846,
0.04703129827976227,
0.08307577669620514,
0.08592214435338974,
-0.018097398802638054,
-0.002647914458066225,
-0.00675104558467865,
-0.014354647137224674,
0.010976145975291729,
-0.07908482104539871,
-0.08052597939968109,
0.005057780537754297,
0.016882553696632385,
-0.006500765681266785,
0.02021719329059124,
0.1230858564376831,
0.08172935992479324,
-0.07310232520103455,
0.003795305034145713,
-0.01827189512550831,
-0.014836345799267292,
0.03628343716263771,
-0.02233065851032734,
0.052271295338869095,
-0.08294053375720978,
0.042913734912872314,
0.011042986996471882,
0.04617564380168915,
-0.01329226978123188,
0.014465803280472755,
0.04395980015397072,
-0.002882167464122176,
-0.023278599604964256,
-0.08203083276748657,
-0.030815042555332184,
-0.03733474016189575,
0.005909958854317665,
-0.005133726634085178,
0.07787534594535828,
0.0043263561092317104,
-0.010345259681344032,
-0.039822615683078766,
0.03686727583408356,
-0.00585129065439105,
-0.0019113451708108187,
0.030780525878071785,
-0.017924966290593147,
-0.041881412267684937,
-0.001985749928280711,
-0.02488364651799202,
0.004138861317187548,
-0.043993301689624786,
0.02374882809817791,
-0.028155582025647163,
-0.03427908569574356,
0.05210422724485397,
-0.04124268516898155,
-0.0016567431157454848,
-0.0003549777902662754,
-0.09331852942705154,
-0.1427963674068451,
-0.07913700491189957,
-0.0037490299437195063,
0.019949734210968018,
-0.029454126954078674,
0.021536441519856453,
0.0010081648360937834,
-0.022841783240437508,
0.008748501539230347,
0.003736939048394561,
0.06653553992509842,
-0.07909888029098511,
0.0495142787694931,
0.03848225995898247,
0.03326971083879471,
0.041973888874053955,
-0.038772787898778915,
0.07358729094266891,
-0.026091760024428368,
-0.039123814553022385,
0.027341678738594055,
0.016763783991336823,
0.041005901992321014,
0.13750843703746796,
-0.03288322314620018,
0.027805274352431297,
0.012111197225749493,
0.1064380556344986,
0.025894643738865852,
-0.016779599711298943,
0.018647581338882446,
-0.0646154135465622,
-0.005067461170256138,
-0.04756013676524162,
0.03215905651450157,
-0.03908226639032364,
-0.10511809587478638,
0.014147567562758923,
0.04427123814821243,
-0.013350199908018112,
0.07201917469501495,
0.07535789906978607,
0.05572940409183502,
0.028596652671694756,
-0.10015156120061874,
0.032107822597026825,
0.051574405282735825,
0.07647868245840073,
-0.10530441254377365,
0.05809856951236725,
0.07650448381900787,
0.023037627339363098,
0.012948302552103996,
0.0011714089196175337,
0.024731505662202835,
-0.031890109181404114,
-0.031657278537750244,
0.04457315802574158,
0.1301347315311432,
-0.06564106792211533,
-0.052025265991687775,
0.0022739062551409006,
0.02302202209830284,
-0.028955182060599327,
-0.03179796412587166,
-0.13233672082424164,
3.662650848698178e-33,
-0.028503060340881348,
-0.012996094301342964,
0.030876362696290016,
-0.06624165922403336,
-0.00547143304720521,
-0.018215227872133255,
0.004267842508852482,
0.026635896414518356,
-0.07758829742670059,
0.07473140209913254,
-0.05016222596168518,
0.037729695439338684,
-0.02819173038005829,
0.08196321129798889,
0.12682701647281647,
-0.032593149691820145,
-0.021795807406306267,
0.1044641062617302,
-0.029386933892965317,
-0.03301934152841568,
0.08112215250730515,
-0.03370116651058197,
0.04749789834022522,
-0.06516646593809128,
0.002471185987815261,
0.005690285935997963,
0.049645956605672836,
0.02082434669137001,
0.023329995572566986,
0.016969824209809303,
-0.10330314189195633,
0.0049377926625311375,
0.08866769075393677,
0.03000544011592865,
-0.003168177092447877,
0.03261019289493561,
-0.042709529399871826,
0.018893839791417122,
-0.0018782084807753563,
0.05642849579453468,
0.020473558455705643,
0.034047920256853104,
0.07730624824762344,
-0.04135676473379135,
-0.07576894015073776,
-0.07768960297107697,
0.025001615285873413,
-0.006641815882176161,
-0.04135821759700775,
-0.09142307937145233,
0.06013961881399155,
-0.009830707684159279,
-0.03985631838440895,
-0.06335438042879105,
-0.009497524239122868,
0.020331570878624916,
0.0066868094727396965,
0.048232097178697586,
0.026078224182128906,
0.06979253888130188,
-0.07779549062252045,
-0.03233226016163826,
-0.02704467438161373,
0.05426988750696182,
0.004501732531934977,
0.028449853882193565,
-0.040375493466854095,
0.015964286401867867,
0.0375819206237793,
-0.04343092441558838,
0.026932911947369576,
-0.0070009068585932255,
0.02229749970138073,
-0.07873757928609848,
0.11798088252544403,
0.0020187802147120237,
0.10018317401409149,
-0.07051809132099152,
0.01047478336840868,
-0.0030434569343924522,
-0.003979047294706106,
0.06431587040424347,
0.012570246122777462,
-0.04995935037732124,
-0.014407265931367874,
-0.07742968201637268,
0.00010988395661115646,
-0.029534827917814255,
-0.021058909595012665,
-0.027446234598755836,
-0.07906212657690048,
0.03310895711183548,
-0.07459266483783722,
-0.03940987214446068,
-0.01938207633793354,
-2.71902937073131e-33,
-0.036051951348781586,
0.007497543469071388,
0.009556311182677746,
0.16818799078464508,
0.0571451298892498,
-0.04427557811141014,
0.015214528888463974,
-0.026844749227166176,
-0.007820208556950092,
-0.05051593482494354,
-0.014641094021499157,
-0.05909213796257973,
0.025186823680996895,
-0.07639937847852707,
0.023475827649235725,
0.015145714394748211,
-0.11609087884426117,
0.01010317262262106,
-0.016826504841446877,
-0.005530351772904396,
0.0005842734826728702,
0.033995121717453,
-0.01785886660218239,
0.00004271346551831812,
-0.02662728913128376,
0.014869960024952888,
-0.08866836875677109,
0.05638653039932251,
0.01688498631119728,
-0.00006886044866405427,
-0.12934458255767822,
-0.017522400245070457,
-0.025547964498400688,
-0.04539337754249573,
-0.016684582456946373,
0.0654589980840683,
0.05957036092877388,
0.10388407111167908,
-0.054257746785879135,
0.14619015157222748,
0.028657888993620872,
0.010444488376379013,
-0.05939153581857681,
0.06707751750946045,
0.06631786376237869,
-0.03834158927202225,
-0.044122397899627686,
-0.017700886353850365,
0.07495895028114319,
0.008294513449072838,
-0.009783382527530193,
0.04205160588026047,
-0.046884384006261826,
0.02880537137389183,
-0.08509291708469391,
-0.04978630319237709,
0.028394535183906555,
-0.01845932938158512,
0.06557536870241165,
-0.009618937969207764,
-0.08114705234766006,
-0.020000867545604706,
0.009841237217187881,
0.033906690776348114,
-0.011894852854311466,
-0.02260177955031395,
0.049798544496297836,
-0.0021087047643959522,
0.0006707482389174402,
0.006562727503478527,
-0.05252815783023834,
-0.00044488534331321716,
0.07817097008228302,
-0.060814596712589264,
-0.048797618597745895,
0.024760764092206955,
0.0446174293756485,
-0.027645407244563103,
0.03505557402968407,
0.03472304344177246,
0.030457304790616035,
0.06544926762580872,
0.0033578858710825443,
-0.01833217591047287,
0.04538682848215103,
0.07683388143777847,
0.08065391331911087,
0.07381055504083633,
-0.00196103029884398,
-0.03455512225627899,
-0.06233391538262367,
-0.025981608778238297,
0.060442496091127396,
-0.0007098284550011158,
-0.05656830221414566,
-4.54370301383733e-8,
-0.034831974655389786,
-0.021348606795072556,
0.016415942460298538,
0.03999845311045647,
0.09126510471105576,
-0.029982132837176323,
-0.02930152788758278,
0.13697238266468048,
-0.015389906242489815,
-0.0001895287714432925,
0.08945483714342117,
0.057125888764858246,
-0.039135247468948364,
0.002780594164505601,
0.024515202268958092,
0.016369204968214035,
0.04435841739177704,
-0.04008252173662186,
-0.015943266451358795,
-0.01824185624718666,
-0.030638061463832855,
0.08926322311162949,
0.045389533042907715,
-0.018669864162802696,
0.08388619869947433,
-0.07652218639850616,
-0.08858157694339752,
0.11573946475982666,
0.0764681026339531,
0.007319518830627203,
-0.02943849191069603,
-0.00882883183658123,
0.01064673624932766,
-0.022707799449563026,
0.06862764805555344,
0.03285645321011543,
-0.06527171283960342,
-0.008558277040719986,
-0.02523083984851837,
-0.0625397264957428,
0.03826246038079262,
0.038346465677022934,
-0.024074699729681015,
-0.015091726556420326,
0.007795378565788269,
0.0031174151226878166,
-0.05360584333539009,
-0.0483894981443882,
0.014904975891113281,
-0.00032508833101019263,
0.0749892145395279,
-0.03087235987186432,
0.005753754172474146,
0.004543313290923834,
0.08617662638425827,
0.0006594301084987819,
-0.04315829649567604,
-0.06834063678979874,
-0.066441610455513,
0.016674034297466278,
0.0080249710008502,
-0.05415194854140282,
-0.09325794130563736,
-0.0716632530093193
] | https://github.com/oseyosey/ccds | 9 | 0 | 0 | 0 |
Generalization through variance: how noise shapes inductive biases in diffusion models | https://openreview.net/forum?id=7lUdo8Vuqa | [
"John Vastola"
] | Poster | How diffusion models generalize beyond their training set is not known, and is somewhat mysterious given two facts: the optimum of the denoising score matching (DSM) objective usually used to train diffusion models is the score function of the training distribution; and the networks usually used to learn the score function are expressive enough to learn this score to high accuracy. We claim that a certain feature of the DSM objective—the fact that its target is not the training distribution's score, but a noisy quantity only equal to it in expectation—strongly impacts whether and to what extent diffusion models generalize. In this paper, we develop a mathematical theory that partly explains this 'generalization through variance' phenomenon. Our theoretical analysis exploits a physics-inspired path integral approach to compute the distributions typically learned by a few paradigmatic under- and overparameterized diffusion models. We find that the distributions diffusion models effectively learn to sample from resemble their training distributions, but with `gaps' filled in, and that this inductive bias is due to the covariance structure of the noisy target used during training. We also characterize how this inductive bias interacts with feature-related inductive biases. | diffusion models, generalization, inductive biases, theory, infinite-width neural networks, generative models, path integral | Diffusion models generalize well partly because of noise built into their training objective | 12,303 | null | [
-0.05791613087058067,
-0.11889293044805527,
0.07955899089574814,
0.07316415011882782,
0.08042161911725998,
-0.035873133689165115,
0.012537924572825432,
-0.038239166140556335,
0.06877448409795761,
-0.07398436963558197,
-0.044063616544008255,
0.007336685433983803,
0.014074254781007767,
0.044718824326992035,
-0.010010414756834507,
-0.0484003871679306,
0.09887191653251648,
0.07034359872341156,
-0.11269511282444,
-0.0221086535602808,
-0.039292916655540466,
0.018117409199476242,
-0.054825980216264725,
0.016509488224983215,
0.054150618612766266,
-0.01468209084123373,
0.030197154730558395,
-0.049724578857421875,
-0.03451090678572655,
-0.010256527923047543,
0.06316819787025452,
-0.03311736136674881,
-0.018636595457792282,
0.004999883938580751,
-0.02838946506381035,
-0.03198909014463425,
-0.08892642706632614,
0.02722446247935295,
0.0667634978890419,
0.03488709777593613,
0.026052096858620644,
0.011369572952389717,
0.04260452091693878,
0.03295530378818512,
0.09041443467140198,
-0.035688240081071854,
-0.019068751484155655,
-0.09034433215856552,
-0.03569917753338814,
0.05715538188815117,
-0.015440879389643669,
-0.0224812813103199,
-0.12419349700212479,
0.06663472205400467,
0.01741378754377365,
0.014462475664913654,
0.08014088124036789,
0.05198030173778534,
0.011439823545515537,
0.009549575857818127,
-0.008194725029170513,
-0.11694493144750595,
-0.056351423263549805,
0.019900012761354446,
-0.018516192212700844,
-0.05090108886361122,
0.04426834359765053,
0.07382708042860031,
-0.03285888209939003,
0.03358745947480202,
0.022451648488640785,
0.09553200006484985,
-0.10893774777650833,
0.028437059372663498,
0.06674200296401978,
0.054955676198005676,
0.024168511852622032,
0.024573642760515213,
0.014027033932507038,
-0.03544245660305023,
0.030260903760790825,
0.037647273391485214,
0.06974279880523682,
-0.061983365565538406,
0.12408316135406494,
-0.01916833408176899,
0.03473449498414993,
-0.016784638166427612,
0.011332735419273376,
-0.01793360710144043,
-0.055455584079027176,
-0.018079344183206558,
-0.1209254339337349,
0.030219852924346924,
0.05194951593875885,
0.07199469208717346,
0.05192682519555092,
-0.012405806221067905,
0.06639846414327621,
0.11123616248369217,
-0.02113916538655758,
0.05164751783013344,
-0.02519473433494568,
-0.018907558172941208,
0.057246845215559006,
-0.07310861349105835,
0.056194499135017395,
-0.012817323207855225,
0.03542802482843399,
-0.036616403609514236,
-0.046074073761701584,
-0.00910197477787733,
-0.017776425927877426,
0.03788861259818077,
-0.002963890554383397,
-0.05710962787270546,
0.06395579129457474,
0.05446286499500275,
-0.009200545027852058,
0.08243230730295181,
-0.039017435163259506,
0.0017013499746099114,
-0.08172845840454102,
0.011583513580262661,
0.001463008113205433,
0.04312914237380028,
-0.06925927102565765,
1.3384840812389865e-33,
0.029892554506659508,
-0.059710294008255005,
0.026707259938120842,
0.009849858470261097,
0.06943714618682861,
-0.0845174714922905,
-0.03832339495420456,
-0.04974571242928505,
0.06631072610616684,
-0.012128914706408978,
-0.08196693658828735,
0.04519578814506531,
-0.059056952595710754,
0.03904813155531883,
0.036461859941482544,
0.027472760528326035,
-0.07289579510688782,
0.09553436934947968,
0.028621146455407143,
-0.007048724684864283,
0.0952032133936882,
-0.035679806023836136,
-0.008054785430431366,
-0.011904893442988396,
-0.006125798914581537,
-0.004556276835501194,
-0.019146962091326714,
-0.020932888612151146,
-0.03342295065522194,
0.027860190719366074,
-0.10240846127271652,
0.008513148874044418,
0.01327729132026434,
0.018480073660612106,
0.04502343386411667,
-0.003128396114334464,
0.014841573312878609,
-0.007535452954471111,
0.002895795274525881,
-0.04932735115289688,
-0.024548513814806938,
0.011334958486258984,
-0.0350942462682724,
-0.00694245146587491,
-0.09951937198638916,
-0.01920313946902752,
-0.014448831789195538,
-0.05316828191280365,
-0.02953324280679226,
-0.06468604505062103,
-0.035396356135606766,
-0.054262924939394,
0.012535125017166138,
-0.050249867141246796,
-0.050673503428697586,
0.06510741263628006,
0.0260321032255888,
-0.05310423672199249,
-0.027938934043049812,
-0.0480068065226078,
0.018565399572253227,
0.05395010486245155,
-0.012963184155523777,
0.0573282390832901,
0.06353729963302612,
-0.003249806584790349,
-0.02190048061311245,
0.041094668209552765,
0.025253791362047195,
-0.025751527398824692,
0.009331568144261837,
0.03381776809692383,
-0.1634739488363266,
-0.10976754128932953,
0.05034560710191727,
-0.05048384144902229,
0.045752834528684616,
0.019485825672745705,
0.01387361716479063,
-0.054077088832855225,
-0.04553157463669777,
-0.026314718648791313,
-0.04764612764120102,
-0.1352585107088089,
-0.02975141443312168,
0.03958679735660553,
-0.05939051881432533,
-0.10183827579021454,
-0.04690049588680267,
-0.09110558778047562,
-0.01975700631737709,
0.006114212796092033,
-0.05171054229140282,
0.014017199166119099,
0.007355781737715006,
-2.2681323292879352e-33,
-0.12165166437625885,
0.10795823484659195,
0.06439981609582901,
0.0705590471625328,
-0.018265822902321815,
0.01893729344010353,
-0.033200059086084366,
0.02006453275680542,
0.05071353539824486,
0.03874993324279785,
-0.07870729267597198,
-0.0052045295014977455,
-0.03875548392534256,
-0.03204325959086418,
-0.0584513396024704,
-0.03138812631368637,
0.06749404966831207,
0.027411719784140587,
-0.0027938683051615953,
-0.021629491820931435,
0.03513941168785095,
0.047622665762901306,
-0.08626081794500351,
-0.015491817146539688,
-0.010954543948173523,
0.024608640000224113,
-0.04308633133769035,
0.06094340607523918,
0.010017400607466698,
0.005680094938725233,
-0.02873842790722847,
0.00556052103638649,
-0.028828009963035583,
-0.06092258542776108,
-0.07791315764188766,
0.10832542926073074,
0.017669769003987312,
0.051670581102371216,
-0.014437993988394737,
0.10147177428007126,
0.03455213084816933,
-0.025474004447460175,
-0.027386916801333427,
0.015678275376558304,
0.04704119265079498,
-0.06349469721317291,
0.018682900816202164,
0.04392784833908081,
0.04490523412823677,
-0.03642650693655014,
0.05715601146221161,
0.06438128650188446,
-0.0623515322804451,
0.10648416727781296,
-0.0922737643122673,
0.012419008649885654,
-0.05355829745531082,
-0.010080304928123951,
0.06315921992063522,
0.03528681769967079,
0.002281318651512265,
0.04462258890271187,
-0.031364597380161285,
-0.00803865771740675,
-0.10961045324802399,
-0.016319051384925842,
-0.08039731532335281,
0.0330238975584507,
0.012963253073394299,
0.027735579758882523,
-0.028358979150652885,
-0.014975360594689846,
-0.016043642535805702,
-0.021675972267985344,
-0.05729181692004204,
-0.07530214637517929,
0.05093130096793175,
-0.029672708362340927,
-0.04801752418279648,
0.009812395088374615,
0.03352711349725723,
-0.00869889184832573,
0.020156776532530785,
-0.03792442008852959,
0.10021085292100906,
0.032325126230716705,
0.02765258029103279,
0.006022430025041103,
0.01887296326458454,
-0.05053842440247536,
-0.016108287498354912,
-0.0017278447048738599,
0.023633984848856926,
0.04322851449251175,
-0.13068966567516327,
-5.04664008360578e-8,
-0.07099149376153946,
0.06861212849617004,
-0.002695953007787466,
0.023342058062553406,
0.06367875635623932,
0.00007169594755396247,
0.04731099680066109,
0.08238254487514496,
-0.03250622749328613,
-0.03239548206329346,
0.05492686107754707,
-0.005915103480219841,
-0.07673488557338715,
0.0404466837644577,
0.023763762786984444,
0.11190568655729294,
-0.024865133687853813,
0.059560760855674744,
-0.006186077371239662,
0.021350134164094925,
0.02155977487564087,
-0.002019731793552637,
-0.003656818764284253,
-0.036065008491277695,
0.129134401679039,
-0.03088896907866001,
-0.03623921051621437,
0.047751571983098984,
0.0513615719974041,
0.07219141721725464,
0.002450368832796812,
0.028596973046660423,
0.013274476863443851,
-0.019995683804154396,
0.03049818053841591,
0.09299619495868683,
0.061193279922008514,
-0.07218922674655914,
-0.023864949122071266,
-0.015258416533470154,
-0.034657951444387436,
0.053405363112688065,
0.027387989684939384,
0.029127048328518867,
-0.013618187978863716,
0.017192862927913666,
0.06027499586343765,
-0.0005280945333652198,
-0.008107922971248627,
0.009997925721108913,
0.07892271876335144,
-0.06394457817077637,
-0.017063505947589874,
0.004741603508591652,
0.06869881600141525,
-0.0016195867210626602,
0.029863866046071053,
-0.04831273853778839,
-0.01201032754033804,
0.05491796135902405,
0.00019035543664358556,
-0.0004942677915096283,
0.017341524362564087,
0.0576978363096714
] | 0 | 0 | 0 | 0 |
|
Learning to Explore and Exploit with GNNs for Unsupervised Combinatorial Optimization | https://openreview.net/forum?id=vaJ4FObpXN | [
"Utku Umur ACIKALIN",
"Aaron M Ferber",
"Carla P Gomes"
] | Poster | Combinatorial optimization (CO) problems are pervasive
across various domains, but their NP-hard nature often necessitates problem-specific
heuristic algorithms. Recent advancements in deep learning have led to the development of learning-based heuristics, yet these approaches often struggle with limited search capabilities.
We introduce Explore-and-Exploit GNN ($X^2$GNN, pronounced x-squared GNN),
a novel unsupervised neural framework that combines exploration and exploitation for combinatorial search optimization:
i) Exploration - $X^2$GNN generates multiple solutions simultaneously, promoting diversity in the search space;
(ii) Exploitation - $X^2$GNN employs neural stochastic iterative refinement to exploit partial existing solutions, guiding the search toward promising regions and helping escape local optima.
By balancing exploration and exploitation, $X^2$GNN achieves superior performance and generalization on several graph CO problems including Max Cut, Max Independent Set, and Max Clique. Notably, for large Max Clique problems, $X^2$GNN consistently generates solutions within 1.2\% of optimality, while other state-of-the-art learning-based approaches struggle to reach within 22\% of optimal. Moreover, $X^2$GNN consistently generates better solutions than Gurobi on large graphs for all three problems under reasonable time budgets. Furthermore, $X^2$GNN exhibits exceptional generalization capabilities. For the Maximum Independent Set problem, $X^2$GNN outperforms state-of-the-art methods even when trained on smaller or out-of-distribution graphs compared to the test set. Our framework offers a more effective and flexible approach to neural combinatorial optimization, addressing a key challenge in the field and providing a promising direction for future research in learning-based heuristics for combinatorial optimization. | combinatorial optimization, unsupervised learning, graph neural networks | X^2GNN iteratively refines (exploit) a solution pool (explore) using GNN for combinatorial optimization, generalizing across problem distribution and size. % , and outperforming ML and traditional OR baselines. | 12,302 | null | [
-0.03180275484919548,
-0.0492744967341423,
0.03292522951960564,
0.013953277841210365,
0.07425867766141891,
0.005444441456347704,
0.02200634963810444,
-0.09412762522697449,
-0.015686998143792152,
-0.04528799653053284,
0.016379309818148613,
-0.048062317073345184,
-0.014997526071965694,
0.08274910598993301,
-0.07011904567480087,
0.035983502864837646,
0.11267630755901337,
0.009492351673543453,
-0.04849982634186745,
-0.006668301299214363,
0.015440859831869602,
-0.03487391769886017,
0.01892804354429245,
-0.03493405506014824,
-0.0014167942572385073,
-0.06225961819291115,
-0.025288434699177742,
-0.06926270574331284,
0.10326799750328064,
0.012852184474468231,
0.09894303232431412,
-0.016190769150853157,
-0.028404418379068375,
-0.007146717049181461,
0.009423097595572472,
0.09345357120037079,
-0.11594247072935104,
0.005799918435513973,
0.04183679074048996,
-0.01201443187892437,
0.014514520764350891,
0.02427472360432148,
0.015001893043518066,
0.014455643482506275,
0.09143288433551788,
0.02007070742547512,
-0.016426410526037216,
-0.04365035519003868,
0.03260098397731781,
-0.06777045875787735,
0.017350776121020317,
-0.04839755594730377,
-0.03810517117381096,
0.031366508454084396,
0.11660700291395187,
0.0018146991496905684,
-0.06675513833761215,
-0.008935194462537766,
0.024298468604683876,
-0.013175759464502335,
0.03667110204696655,
-0.10675826668739319,
0.007265506777912378,
-0.0709688737988472,
-0.004810701124370098,
-0.05362103879451752,
0.03737291693687439,
0.05604062229394913,
-0.007729764096438885,
0.086290143430233,
0.12502692639827728,
0.11107492446899414,
-0.04103892669081688,
-0.0133648831397295,
0.13560858368873596,
0.14079447090625763,
0.06746424734592438,
-0.010022329166531563,
0.002306644106283784,
-0.0012175472220405936,
0.04226168990135193,
-0.00023738504387438297,
0.03072257898747921,
0.03415176272392273,
0.062494441866874695,
-0.05585963651537895,
-0.04296473041176796,
-0.002418539021164179,
0.08292622119188309,
-0.005617940798401833,
-0.09447437524795532,
0.03465837240219116,
0.010378080420196056,
-0.0033162350300699472,
-0.004411704372614622,
-0.07190795242786407,
0.08936485648155212,
-0.007138307671993971,
-0.052136778831481934,
0.05514635518193245,
-0.005259675905108452,
0.054056547582149506,
0.05713306739926338,
-0.07329005748033524,
0.01235134806483984,
-0.01728399656713009,
0.018267884850502014,
0.06475953012704849,
0.052417028695344925,
-0.06678451597690582,
-0.012267069891095161,
0.09603051096200943,
0.044735491275787354,
0.028286080807447433,
0.00029166400781832635,
-0.0011088114697486162,
0.005211427807807922,
0.026837196201086044,
0.059867218136787415,
0.11633984744548798,
-0.06623846292495728,
0.002110176021233201,
-0.0704893097281456,
0.017307594418525696,
-0.06222835183143616,
-0.004434049129486084,
-0.046675559133291245,
5.972070764019815e-33,
0.02320338413119316,
0.030856747180223465,
0.04225480183959007,
-0.03037922829389572,
0.06713740527629852,
-0.04996199160814285,
0.01669556088745594,
0.058187782764434814,
-0.04526260867714882,
0.04940195754170418,
-0.0352550745010376,
0.04268479719758034,
-0.09045468270778656,
0.08485473692417145,
0.03181379288434982,
-0.07644106447696686,
-0.01578556001186371,
-0.0016300342977046967,
-0.03209611028432846,
-0.06962898373603821,
0.05561457946896553,
-0.04094402864575386,
0.007986198179423809,
0.006482917349785566,
0.08186528086662292,
-0.06416790187358856,
0.039418209344148636,
-0.09433695673942566,
0.032257020473480225,
0.0087836142629385,
-0.14564798772335052,
0.0033219787292182446,
0.02487112581729889,
0.10999920964241028,
0.01430757250636816,
0.044554661959409714,
-0.00030932168010622263,
0.009061230346560478,
-0.021080534905195236,
0.02727660909295082,
-0.03207163140177727,
0.01714545115828514,
0.07746066153049469,
-0.04392913728952408,
-0.024659451097249985,
-0.08885940909385681,
0.045777007937431335,
-0.01641264371573925,
-0.0747179388999939,
-0.050004325807094574,
-0.046586234122514725,
0.048617295920848846,
-0.038228411227464676,
0.021474620327353477,
0.08637794107198715,
-0.008839067071676254,
0.12401439249515533,
0.04468684270977974,
0.04328795522451401,
0.00042299769120290875,
0.04824408143758774,
-0.014253532513976097,
-0.049787506461143494,
0.04945908486843109,
0.011718625202775002,
0.06452412903308868,
-0.05892923101782799,
-0.022696908563375473,
0.08596735447645187,
-0.012220237404108047,
-0.01965165324509144,
-0.010091735050082207,
0.008784253150224686,
-0.0653819739818573,
0.06806249171495438,
0.0049581583589315414,
0.024412119761109352,
-0.03254452720284462,
-0.009368494153022766,
-0.036455024033784866,
-0.002443756442517042,
0.02312164567410946,
-0.06931238621473312,
-0.04268600791692734,
-0.07024255394935608,
-0.04551341012120247,
0.028603490442037582,
-0.11373842507600784,
-0.015541944652795792,
-0.04860507324337959,
-0.06061858311295509,
0.01026065181940794,
0.012428874149918556,
-0.0432138629257679,
-0.01093225833028555,
-4.5632316010096876e-33,
-0.01464110892266035,
0.016180939972400665,
-0.056180667132139206,
0.0011777427280321717,
0.04020099714398384,
-0.005589415784925222,
-0.02170497365295887,
-0.11987798660993576,
-0.06194670870900154,
-0.03263930603861809,
-0.06007032468914986,
-0.003206043504178524,
0.017373956739902496,
0.028964051976799965,
-0.043978434056043625,
0.0222508292645216,
-0.01741369441151619,
0.072167307138443,
-0.00448490958660841,
-0.01481570303440094,
-0.015563465654850006,
0.1292697787284851,
-0.05626172199845314,
0.0029174385126680136,
-0.029042230919003487,
0.02786395512521267,
-0.012807486578822136,
0.015221121720969677,
0.026966696605086327,
0.08089884370565414,
-0.09925130009651184,
-0.021059392020106316,
-0.1289282888174057,
0.030550220981240273,
0.02401369996368885,
0.050625137984752655,
0.012972402386367321,
0.027078233659267426,
-0.01913822442293167,
0.060699112713336945,
0.03297434374690056,
-0.03643598034977913,
-0.01933012716472149,
0.09677180647850037,
0.022657062858343124,
0.050411589443683624,
-0.00009144461364485323,
0.0006145909428596497,
0.019811004400253296,
0.04271058738231659,
0.009695862419903278,
0.02416747435927391,
-0.09415259212255478,
0.009418714791536331,
-0.050784412771463394,
-0.05284740775823593,
-0.007814678363502026,
0.033928994089365005,
0.044096413999795914,
0.023996399715542793,
0.0004966374253854156,
-0.034625060856342316,
-0.06045997142791748,
0.024125484749674797,
-0.08473021537065506,
-0.04717714339494705,
-0.06566713005304337,
0.05393381789326668,
-0.02032473124563694,
0.021970761939883232,
-0.0998874232172966,
-0.05041535943746567,
0.04176802933216095,
-0.018210021778941154,
-0.10954523831605911,
0.07565967738628387,
0.009145626798272133,
-0.019225796684622765,
-0.019355446100234985,
-0.0239546038210392,
0.014583294279873371,
-0.008847491815686226,
0.0715077668428421,
0.06206731125712395,
0.059320174157619476,
0.027822788804769516,
0.02768275886774063,
0.07776395231485367,
0.018657321110367775,
-0.026574360206723213,
-0.04567265138030052,
-0.0432610884308815,
-0.035704657435417175,
0.036107875406742096,
0.015610174275934696,
-4.073734771736781e-8,
-0.10477231442928314,
0.0636337473988533,
0.044939447194337845,
-0.023908015340566635,
0.05016220733523369,
-0.036973170936107635,
-0.03317895904183388,
0.041699107736349106,
-0.031033935025334358,
0.008120681159198284,
0.0177913848310709,
-0.008901714347302914,
0.007079703267663717,
-0.026096161454916,
-0.1077653020620346,
0.026225026696920395,
0.022348716855049133,
-0.03660259768366814,
0.029702411964535713,
-0.01949213445186615,
-0.04610311985015869,
-0.004759355913847685,
-0.008309731259942055,
-0.03279571607708931,
-0.04302574321627617,
-0.10418873280286789,
-0.06498164683580399,
0.02384185791015625,
0.03163382411003113,
0.030446231365203857,
-0.04969802871346474,
0.0009729656740091741,
0.06185754016041756,
-0.036870867013931274,
-0.005213056690990925,
0.09506221860647202,
-0.04836444929242134,
0.008605421520769596,
-0.017807677388191223,
-0.09928975254297256,
0.008418792858719826,
0.09177862107753754,
-0.02015618048608303,
0.0030098287388682365,
-0.05428642779588699,
0.03940451517701149,
0.03546140715479851,
-0.004701436031609774,
0.03837030008435249,
-0.012260675430297852,
0.011454446241259575,
-0.013573451898992062,
0.0287396851927042,
0.05600980296730995,
0.11920350044965744,
0.008938830345869064,
-0.0726117417216301,
-0.03580893576145172,
0.0321938581764698,
0.055173493921756744,
-0.027196533977985382,
0.0453929603099823,
-0.08366529643535614,
-0.01460634171962738
] | 0 | 0 | 0 | 0 |
|
Fitting Networks with a Cancellation Trick | https://openreview.net/forum?id=C06kww3Qky | [
"Jiashun Jin",
"Jingming Wang"
] | Poster | The degree-corrected block model (DCBM), latent space model (LSM), and $\beta$-model are all popular network models. We combine their modeling ideas and propose the logit-DCBM as a new model. Similar as the $\beta$-model and LSM, the logit-DCBM contains nonlinear factors, where fitting the parameters is a challenging open problem. We resolve this problem by introducing a cancellation trick. We also propose R-SCORE as a recursive community detection algorithm, where in each iteration, we first use the idea above to update our parameter estimation, and then use the results to remove the nonlinear factors in the logit-DCBM so the renormalized model approximately satisfies a low-rank model, just like the DCBM. Our numerical study suggests that R-SCORE significantly improves over existing spectral approaches in many cases. Also, theoretically, we show that the Hamming error rate of R-SCORE is faster than that of SCORE in a specific sparse region, and is at least as fast outside this region. | Network analysis, DCBM, logit-DCBM, community detection, SCORE | null | 12,287 | 2502.16728 | [
-0.03042888455092907,
-0.11571192741394043,
0.002749539678916335,
0.08315294235944748,
0.060329779982566833,
-0.003381958929821849,
-0.03133991360664368,
-0.03405909240245819,
0.026568492874503136,
0.004557416774332523,
0.01015884056687355,
0.043442294001579285,
0.0535951666533947,
-0.007903809659183025,
-0.05260753259062767,
-0.006362718064337969,
0.010290592908859253,
0.08019910752773285,
-0.02157430723309517,
-0.04614364355802536,
-0.0835961252450943,
0.06033897399902344,
-0.04299301281571388,
0.008560121059417725,
0.13639108836650848,
-0.05735492706298828,
-0.009300614707171917,
-0.014419769868254662,
-0.061081912368535995,
-0.012702647596597672,
0.03404742106795311,
0.027095047757029533,
0.03896228224039078,
0.05312017351388931,
-0.045549746602773666,
0.02376447431743145,
-0.021346597000956535,
-0.02474255859851837,
-0.0033415001817047596,
0.06432078778743744,
0.0753970518708229,
0.041909728199243546,
0.0628509446978569,
-0.0007485165260732174,
0.0713798850774765,
-0.05004582554101944,
-0.055327337235212326,
0.017803406342864037,
-0.024001430720090866,
-0.03487809747457504,
0.004425135441124439,
0.027358565479516983,
-0.010493360459804535,
0.08023528754711151,
0.012274662964046001,
-0.02607535384595394,
0.01687033474445343,
-0.021367114037275314,
0.03783454746007919,
-0.0325532890856266,
0.04112758859992027,
-0.0475827157497406,
-0.07688561826944351,
0.001956320833414793,
0.05898923799395561,
0.028591260313987732,
-0.015222539193928242,
0.07311189919710159,
0.016857655718922615,
0.026409901678562164,
-0.03968682885169983,
0.04785703867673874,
-0.09856966137886047,
0.06726706773042679,
0.05624231696128845,
0.07586818188428879,
0.06488239765167236,
0.06866173446178436,
0.09894134849309921,
-0.0020570955239236355,
0.019410593435168266,
0.007819403894245625,
-0.014711308293044567,
-0.038895383477211,
0.06164092198014259,
-0.02363322488963604,
0.02693161927163601,
-0.023869801312685013,
0.07094921916723251,
-0.021617893129587173,
-0.06283606588840485,
0.07311610132455826,
0.0046040755696594715,
-0.016890855506062508,
-0.039351943880319595,
-0.004657666198909283,
0.06679301708936691,
-0.06744692474603653,
-0.00403797160834074,
0.1458110511302948,
-0.050702620297670364,
0.03476029634475708,
0.03367920592427254,
-0.11516322195529938,
0.03483065962791443,
0.007422978058457375,
0.06042352318763733,
0.14535365998744965,
0.015905670821666718,
-0.1381445825099945,
-0.04899134114384651,
0.0498088076710701,
-0.023177850991487503,
0.05147041752934456,
0.08351349830627441,
-0.06035301089286804,
0.12715037167072296,
0.005204282701015472,
0.027974314987659454,
0.051189716905355453,
-0.012891515158116817,
0.010736670345067978,
0.03653373569250107,
-0.10725634545087814,
0.04717360809445381,
0.07839835435152054,
-0.1256488412618637,
1.8076981715685532e-33,
0.04411790147423744,
0.0915692076086998,
0.015581750310957432,
-0.0032079797238111496,
0.0635279193520546,
0.046494193375110626,
-0.05201689898967743,
-0.002629652852192521,
0.008486342616379261,
0.05754535272717476,
-0.0765075832605362,
0.031049298122525215,
-0.02771591581404209,
0.033647239208221436,
0.034610506147146225,
0.02076800726354122,
0.02307526022195816,
-0.003568662330508232,
-0.0234704427421093,
-0.030499126762151718,
0.08681852370500565,
-0.037298258394002914,
0.011831458657979965,
-0.05045253038406372,
0.019566092640161514,
-0.05086289346218109,
0.02485046535730362,
-0.034753456711769104,
-0.03918585926294327,
0.015226221643388271,
-0.027483386918902397,
0.07061269134283066,
-0.0544339157640934,
0.07522415369749069,
0.024033429101109505,
-0.014037376269698143,
0.07999186962842941,
-0.04191731661558151,
0.018457967787981033,
-0.1250815987586975,
-0.04478813335299492,
-0.010175195522606373,
-0.10374632477760315,
-0.08007986843585968,
-0.05253153666853905,
-0.0432325042784214,
0.015938693657517433,
-0.07303230464458466,
-0.0024381563998758793,
0.02473462000489235,
0.026931393891572952,
0.03471586853265762,
-0.03737201914191246,
0.01301916129887104,
-0.09964437037706375,
-0.03770393133163452,
0.028757089748978615,
0.04734988138079643,
0.05866168811917305,
0.07598371803760529,
-0.0004155241185799241,
-0.0001844043581513688,
-0.006510323379188776,
-0.04057712107896805,
0.023213667795062065,
0.01805860735476017,
-0.01732795126736164,
-0.020564552396535873,
0.0934089943766594,
-0.020376749336719513,
-0.00414257263764739,
0.024856099858880043,
-0.052079081535339355,
0.0014678728766739368,
-0.022970758378505707,
-0.042955558747053146,
-0.04362906515598297,
-0.08861485868692398,
-0.002223015297204256,
0.010095900855958462,
-0.06462039053440094,
0.0112653449177742,
-0.03271198272705078,
-0.05350789800286293,
-0.04019602760672569,
-0.08343431353569031,
0.029000386595726013,
-0.017921265214681625,
-0.07199320942163467,
-0.09103719145059586,
-0.05455806478857994,
-0.000372265960322693,
0.02487630397081375,
0.015121204778552055,
0.02917172759771347,
-7.84645968922731e-34,
-0.06807009875774384,
-0.008809234015643597,
0.02147834748029709,
0.04567890986800194,
0.028120724484324455,
-0.13473093509674072,
0.0041305855847895145,
-0.014193834736943245,
0.020429054275155067,
-0.04999956861138344,
0.06626725196838379,
-0.0336359478533268,
0.02096601389348507,
0.05145645886659622,
0.1500314325094223,
0.01736205443739891,
-0.05560960993170738,
0.005001717712730169,
-0.015386493876576424,
0.009137803688645363,
-0.02028755284845829,
0.0735694169998169,
-0.10373423248529434,
0.01497782301157713,
0.01739192008972168,
0.023150209337472916,
-0.039319466799497604,
0.021809693425893784,
0.04289998114109039,
-0.04326881840825081,
-0.013450621627271175,
0.05610162392258644,
-0.09136039018630981,
-0.03632849454879761,
0.03791552409529686,
0.11274032294750214,
-0.0079962108284235,
0.0673065111041069,
-0.028646904975175858,
0.013600352220237255,
0.07921092957258224,
-0.01567402482032776,
-0.12718039751052856,
-0.02076214738190174,
0.026800280436873436,
0.030313709750771523,
0.04541347548365593,
0.04207300767302513,
-0.002922021085396409,
0.05117601156234741,
-0.03286835923790932,
-0.03586331382393837,
0.052445270121097565,
0.08821385353803635,
-0.01809956692159176,
0.028068793937563896,
-0.04751982539892197,
0.07898546755313873,
0.06538256257772446,
0.021285971626639366,
-0.021162696182727814,
-0.10741659253835678,
0.013011039234697819,
0.03916483744978905,
-0.06827004253864288,
-0.12736597657203674,
0.049949485808610916,
0.040241539478302,
-0.014538769610226154,
0.01677573099732399,
0.014751502312719822,
0.05721048638224602,
0.022541791200637817,
0.06872060894966125,
-0.013793401420116425,
-0.0460139736533165,
0.0432015061378479,
-0.011499876156449318,
-0.009194457903504372,
-0.03887355327606201,
0.013527210801839828,
0.06466273963451385,
-0.003574213245883584,
-0.028880586847662926,
0.06944802403450012,
0.003027206053957343,
0.06176774576306343,
-0.006400390528142452,
0.016145208850502968,
-0.0278388150036335,
-0.08179695159196854,
-0.061864059418439865,
-0.015318932943046093,
0.015972578898072243,
-0.04985036700963974,
-5.45785461270043e-8,
-0.06955202668905258,
0.014553623273968697,
-0.07652610540390015,
0.032392311841249466,
0.036336082965135574,
0.028020143508911133,
0.008170879445970058,
0.013256331905722618,
-0.02492137998342514,
0.005798597354441881,
0.09892980009317398,
0.028902685269713402,
-0.02367286942899227,
-0.0406024232506752,
-0.08236975222826004,
0.027508853003382683,
-0.0261986181139946,
0.05736057832837105,
-0.006160815712064505,
0.03370857611298561,
0.04237945377826691,
0.010261215269565582,
0.0031035065185278654,
0.04200524464249611,
0.0878618061542511,
-0.04726225510239601,
-0.04689629003405571,
0.06044973433017731,
0.03591848537325859,
-0.007555379532277584,
0.036252375692129135,
0.024908114224672318,
0.04093227535486221,
0.030172260478138924,
-0.019283974543213844,
0.14960011839866638,
0.019249368458986282,
0.005110903177410364,
-0.030333738774061203,
0.0064317695796489716,
-0.040034085512161255,
-0.0016896432498469949,
-0.04197905957698822,
-0.001719899009913206,
-0.008774091489613056,
-0.03938736766576767,
0.0028345806058496237,
-0.03212983161211014,
0.017591243609786034,
-0.10172779113054276,
-0.006638957187533379,
-0.026621920987963676,
-0.08009538054466248,
-0.031625986099243164,
0.014974944293498993,
0.003995301201939583,
-0.009075832553207874,
-0.02841382659971714,
0.051919709891080856,
0.030442334711551666,
-0.03780296817421913,
-0.05181587114930153,
-0.06883285194635391,
-0.056913670152425766
] | 0 | 0 | 0 | 0 |
|
Do LLM Agents Have Regret? A Case Study in Online Learning and Games | https://openreview.net/forum?id=qn9tBYQHGi | [
"Chanwoo Park",
"Xiangyu Liu",
"Asuman E. Ozdaglar",
"Kaiqing Zhang"
] | Poster | Large language models (LLMs) have been increasingly employed for (interactive) decision-making, via the development of LLM-based autonomous agents. Despite their emerging successes, the performance of LLM agents in decision-making has not been fully investigated through quantitative metrics, especially in the multi-agent setting when they interact with each other, a typical scenario in real-world LLM-agent applications. To better understand the limits of LLM agents in these interactive environments, we propose to study their interactions in benchmark decision-making settings in online learning and game theory, through the performance metric of regret. We first empirically study the no-regret behaviors of LLMs in canonical non-stochastic online learning problems, as well as the emergence of equilibria when LLM agents interact through playing repeated games. We then provide some theoretical insights into the no-regret behaviors of LLM agents, under certain assumptions on the supervised pre-training and the rationality model of human decision-makers who generate the data. Notably, we also identify (simple) cases where advanced LLMs such as GPT-4 fail to be no-regret. To further promote the no-regret behaviors, we propose a novel unsupervised training loss of regret-loss, which, in contrast to the supervised pre-training loss, does not require the labels of (optimal) actions. Finally, we establish the statistical guarantee of generalization bound for regret-loss minimization, and more importantly, the optimization guarantee that minimizing such a loss may automatically lead to known no-regret learning algorithms, when single-layer self-attention models are used. Our further experiments demonstrate the effectiveness of our regret-loss, especially in addressing the above “regrettable” cases. | LLM agents, online learning, repeated games | null | 12,286 | 2403.16843 | [
0.04373643919825554,
-0.07879286259412766,
-0.03809075802564621,
-0.022495873272418976,
0.026272766292095184,
0.033337805420160294,
0.04884723946452141,
0.02734610065817833,
0.13516278564929962,
0.009643684141337872,
-0.026372363790869713,
0.0067170667462050915,
0.06818961352109909,
0.01630082167685032,
0.011129495687782764,
-0.0360938236117363,
0.0995420441031456,
0.04348278045654297,
-0.06966222077608109,
-0.05669347196817398,
0.030884746462106705,
-0.022849561646580696,
0.07819612324237823,
-0.04317643120884895,
0.033313408493995667,
-0.05890516936779022,
-0.017839394509792328,
0.020620333030819893,
-0.024662597104907036,
0.013137858361005783,
0.009480887092649937,
0.03934400901198387,
-0.0021957922726869583,
-0.01090223714709282,
-0.08689188957214355,
0.0019313544034957886,
-0.10170543193817139,
-0.002307336078956723,
-0.009194248355925083,
0.0013828304363414645,
-0.01943466253578663,
-0.02862578257918358,
-0.021206431090831757,
0.0033260018099099398,
0.05241234973073006,
0.0006659927312284708,
-0.030782952904701233,
-0.052557237446308136,
-0.06541619449853897,
-0.008317931555211544,
-0.062390103936195374,
-0.042126789689064026,
-0.04435649886727333,
-0.021369652822613716,
-0.006654659286141396,
0.04053545370697975,
-0.031892985105514526,
0.05170704424381256,
0.019073842093348503,
-0.04976733773946762,
-0.008078920654952526,
-0.17165391147136688,
-0.037758272141218185,
0.0016641032416373491,
-0.01098884828388691,
0.03541846573352814,
-0.0037606130354106426,
0.04078196361660957,
-0.013886824250221252,
0.04377147555351257,
0.028163691982626915,
0.027780625969171524,
-0.04304775223135948,
-0.01732759177684784,
-0.0389537550508976,
-0.038686830550432205,
0.032640811055898666,
-0.01137623842805624,
0.027618709951639175,
0.046166304498910904,
0.004119498189538717,
0.035402633249759674,
-0.007206781301647425,
-0.005691005382686853,
0.02470047026872635,
-0.0020119142718613148,
0.04647090658545494,
0.053664520382881165,
0.14506784081459045,
-0.05035115405917168,
-0.03886290639638901,
-0.00024908044724725187,
-0.009864780120551586,
0.03505706414580345,
0.01055191457271576,
0.04286697134375572,
-0.04777102544903755,
-0.043728578835725784,
-0.025246139615774155,
0.07939055562019348,
0.016094619408249855,
0.06335222721099854,
-0.06320875883102417,
0.04773390665650368,
0.006962527520954609,
0.0059433141723275185,
-0.010907822288572788,
0.0487031452357769,
0.06550478935241699,
-0.03740840405225754,
-0.01203098427504301,
-0.00460665300488472,
0.08047400414943695,
0.08853857964277267,
-0.015858817845582962,
0.026624461635947227,
0.056527476757764816,
0.020905815064907074,
-0.024403180927038193,
0.07839613407850266,
0.013979102484881878,
0.0138612175360322,
0.11302034556865692,
-0.035371050238609314,
0.014301392249763012,
0.005843194201588631,
-0.0762905701994896,
-1.6638731087087106e-34,
0.11399585753679276,
-0.03778194263577461,
0.0330386720597744,
-0.029738374054431915,
0.02904088795185089,
0.03348521515727043,
0.0662577897310257,
0.03900803625583649,
-0.03698718547821045,
0.0015975231071934104,
0.021643241867423058,
0.03172938525676727,
-0.05911865830421448,
0.07894650846719742,
0.09695179015398026,
0.009397625923156738,
-0.03412889689207077,
0.00048515843809582293,
0.03759681060910225,
-0.0694412812590599,
0.09619385004043579,
0.022272460162639618,
0.07284574210643768,
-0.06494347751140594,
0.002587212948128581,
0.040110912173986435,
-0.03819351643323898,
-0.0501236692070961,
0.056676965206861496,
-0.012902077287435532,
-0.09465868771076202,
-0.0011193979298695922,
-0.0362824872136116,
0.04893728345632553,
0.03189226984977722,
0.02646423876285553,
-0.14835356175899506,
0.008872044272720814,
-0.0043976460583508015,
-0.038350362330675125,
-0.04945803061127663,
-0.011009622365236282,
0.05729755759239197,
-0.05656549334526062,
-0.13369880616664886,
-0.0771016925573349,
0.01988905295729637,
-0.03771345317363739,
-0.03633470460772514,
-0.04017271101474762,
0.0412253737449646,
0.06171319633722305,
-0.033225689083337784,
0.03497471287846565,
-0.06779364496469498,
0.05605284497141838,
-0.018436698243021965,
0.04173676297068596,
-0.10803742706775665,
0.006818280555307865,
-0.01723559945821762,
-0.022931542247533798,
0.00952918827533722,
0.04644322767853737,
-0.0015103378100320697,
0.0880037471652031,
-0.05024508759379387,
0.06968697905540466,
0.10303054749965668,
-0.08451695740222931,
0.025582464411854744,
-0.004315952304750681,
0.02104632370173931,
-0.10587956756353378,
0.0677492618560791,
-0.030257077887654305,
0.06588686257600784,
-0.017992518842220306,
0.08007072657346725,
-0.03527870029211044,
0.04688279703259468,
-0.0028532803989946842,
-0.1079440787434578,
-0.06507333368062973,
-0.050391677767038345,
-0.06711829453706741,
0.009287224151194096,
-0.08758331090211868,
0.028683552518486977,
0.061358071863651276,
-0.1578482836484909,
-0.005441374145448208,
-0.04452347755432129,
0.05062447860836983,
-0.013933599926531315,
-2.3729166304042032e-33,
-0.0773303434252739,
-0.13896611332893372,
0.0038017691113054752,
0.10137009620666504,
0.020786764100193977,
-0.03611202538013458,
-0.05377126485109329,
-0.027331802994012833,
0.013952433131635189,
-0.05458628386259079,
-0.017024096101522446,
0.08260209858417511,
0.06458031386137009,
0.044105298817157745,
0.012538897804915905,
-0.06541088968515396,
0.0160029586404562,
-0.0038189832121133804,
0.014958033338189125,
-0.020225893706083298,
0.040744684636592865,
0.07250979542732239,
-0.1256343424320221,
0.04292625933885574,
0.00820852816104889,
0.04199548810720444,
0.036732159554958344,
-0.006490273401141167,
-0.014314915053546429,
0.04867967963218689,
0.029225140810012817,
0.0023787999525666237,
-0.0741954818367958,
0.0038336883299052715,
-0.008468798361718655,
0.0717029869556427,
0.07770999521017075,
0.05307749658823013,
-0.06643463671207428,
0.09095434099435806,
0.05483097955584526,
-0.08265122771263123,
-0.03974186256527901,
-0.03178720921278,
0.015744661912322044,
0.02338971383869648,
-0.0488327294588089,
-0.015352929942309856,
0.08890361338853836,
0.00047983613330870867,
0.061240606009960175,
0.025309571996331215,
-0.06337893009185791,
0.02481858618557453,
-0.06092777103185654,
-0.07211514562368393,
-0.00047522710519842803,
-0.07364946603775024,
0.03215450793504715,
0.031229771673679352,
-0.03709295764565468,
-0.012901701033115387,
-0.030170615762472153,
0.016718028113245964,
-0.027907174080610275,
-0.003267867723479867,
-0.07568606734275818,
0.01169819850474596,
-0.0012861620634794235,
-0.011732108891010284,
0.014173666015267372,
0.0037433889228850603,
-0.08729634433984756,
-0.06787574291229248,
0.043019287288188934,
0.021421687677502632,
-0.08135232329368591,
0.03686203807592392,
-0.004265699535608292,
-0.029430052265524864,
0.007052340544760227,
0.0472392737865448,
0.04195720702409744,
0.030154798179864883,
0.014783622696995735,
0.000632462790235877,
0.03866831958293915,
0.07111775875091553,
0.03061997890472412,
-0.0567660927772522,
0.03208869323134422,
-0.0010725884931162,
0.025330642238259315,
0.0018747637514024973,
-0.1070348247885704,
-4.3869921029227044e-8,
-0.02161070518195629,
-0.06667855381965637,
0.06181610748171806,
0.03895583748817444,
0.05465339869260788,
-0.007987791672348976,
-0.008728076703846455,
0.05052172765135765,
-0.054980382323265076,
0.03307283669710159,
0.099929578602314,
0.013648021966218948,
-0.07544080913066864,
-0.049088723957538605,
0.04596766084432602,
0.04364221915602684,
-0.012166846543550491,
-0.07019279897212982,
-0.005812312476336956,
0.0017115193186327815,
0.06430968642234802,
0.000008349430572707206,
-0.020252838730812073,
-0.062077876180410385,
-0.022603580728173256,
-0.06253672391176224,
0.013660917058587074,
0.08947991579771042,
0.013140360824763775,
0.10101810842752457,
0.04007669538259506,
-0.023831721395254135,
0.015848621726036072,
0.003429759992286563,
0.03419553488492966,
-0.02313857711851597,
0.016324343159794807,
-0.07741847634315491,
0.04517153650522232,
-0.009429448284208775,
-0.030115507543087006,
0.01015013363212347,
-0.019799649715423584,
0.03799942880868912,
0.02145812101662159,
0.01627320796251297,
-0.11006998270750046,
-0.09503551572561264,
0.008632197976112366,
-0.07417550683021545,
-0.030569618567824364,
0.04540475830435753,
-0.0429198294878006,
0.061764251440763474,
0.1060645580291748,
-0.009813575074076653,
0.0027422988787293434,
-0.04013863205909729,
0.04458894580602646,
0.022367320954799652,
0.05080072954297066,
-0.049464333802461624,
-0.09252110123634338,
-0.004324143752455711
] | 0 | 0 | 0 | 0 |
|
Seq-VCR: Preventing Collapse in Intermediate Transformer Representations for Enhanced Reasoning | https://openreview.net/forum?id=30oIfmrcFO | [
"Md Rifat Arefin",
"Gopeshh Subbaraj",
"Nicolas Gontier",
"Yann LeCun",
"Irina Rish",
"Ravid Shwartz-Ziv",
"Christopher Pal"
] | Poster | Decoder-only Transformers often struggle with complex reasoning tasks, particularly arithmetic reasoning requiring multiple sequential operations. In this work, we identify representation collapse in the model’s intermediate layers as a key factor limiting their reasoning capabilities. To address this, we propose Sequential Variance-Covariance Regularization (Seq-VCR), which enhances the entropy of intermediate representations and prevents collapse. Combined with dummy pause tokens as substitutes for chain-of-thought (CoT) tokens, our method significantly improves performance in arithmetic reasoning problems. In the challenging 5 × 5 integer multiplication task, our approach achieves 99.5% exact match accuracy, outperforming models of the same size (which yield 0% accuracy) and GPT-4 with five-shot CoT prompting (44%). We also demonstrate superior results on arithmetic expression and longest increasing subsequence (LIS) datasets. Our findings highlight the importance of preventing intermediate layer representation collapse to enhance the reasoning capabilities of Transformers and show that Seq-VCR offers an effective solution without requiring explicit CoT supervision. | LLMs, Representation Learning, Reasoning, Representation Collapse | null | 12,284 | null | [
-0.06932051479816437,
-0.021598346531391144,
0.07182082533836365,
-0.033636998385190964,
-0.006300958804786205,
0.022922536358237267,
-0.005866115912795067,
0.04041521996259689,
0.05470224469900131,
-0.07339150458574295,
-0.08480308204889297,
-0.009095259010791779,
-0.005307008512318134,
0.0634959265589714,
-0.007435214705765247,
0.024551020935177803,
0.003851513843983412,
0.10860839486122131,
-0.07591243088245392,
-0.1398514360189438,
0.010965406894683838,
0.05495933070778847,
-0.009192223660647869,
0.033411137759685516,
0.05651221424341202,
-0.02723887376487255,
-0.024743450805544853,
-0.03815072402358055,
0.08844377100467682,
0.06431034207344055,
0.03480494022369385,
0.004678245168179274,
-0.01108156330883503,
0.035099614411592484,
-0.01640092022716999,
0.0693337693810463,
-0.047581326216459274,
-0.008478536270558834,
-0.034698598086833954,
0.014312977902591228,
-0.0032854045275598764,
-0.024433594197034836,
0.025829290971159935,
0.004679860081523657,
0.03328699618577957,
-0.004232346545904875,
0.03022535890340805,
-0.05169744789600372,
0.009127883240580559,
-0.018262475728988647,
0.004023282788693905,
0.06597302854061127,
-0.05704678222537041,
0.06858540326356888,
-0.04028699919581413,
-0.037387456744909286,
-0.016750900074839592,
0.022639160975813866,
0.012088187038898468,
-0.04295976832509041,
-0.08584514260292053,
-0.057046886533498764,
-0.023180263116955757,
-0.05407975614070892,
-0.07705517113208771,
0.02222718670964241,
0.09073606133460999,
0.028952140361070633,
-0.013536781072616577,
0.04002220928668976,
-0.0400407612323761,
0.09354574233293533,
-0.10355120152235031,
0.06831081956624985,
0.032088059931993484,
0.09062351286411285,
0.015708040446043015,
-0.11445087194442749,
0.04265717417001724,
-0.0667567104101181,
0.032233819365501404,
-0.01844303123652935,
0.049403637647628784,
0.022147340700030327,
0.07420375943183899,
-0.03426305577158928,
0.03165922686457634,
0.05189407244324684,
0.034920584410429,
0.021008970215916634,
-0.04210929200053215,
-0.08913248777389526,
0.09888066351413727,
0.0539817251265049,
0.07491213083267212,
0.06880903244018555,
0.00006864882016088814,
-0.012559588998556137,
0.02441207319498062,
0.07558928430080414,
-0.01742975041270256,
0.08612549304962158,
0.01064840704202652,
-0.0444103442132473,
-0.006790379993617535,
-0.028629863634705544,
0.0524134635925293,
0.08084183931350708,
0.0038199389819055796,
-0.10674717277288437,
0.0468585230410099,
0.0389285646378994,
0.08106756955385208,
0.002478189766407013,
0.02206828072667122,
0.03467274457216263,
0.009252042509615421,
-0.0011415643384680152,
-0.006658374797552824,
-0.00036820044624619186,
0.03495119884610176,
0.06100252643227577,
-0.008950741030275822,
0.07353255897760391,
0.0405443049967289,
-0.005820679944008589,
-0.017042813822627068,
1.0155587935013149e-33,
-0.009164949879050255,
0.04798981919884682,
0.11265171319246292,
0.0671142190694809,
0.008718892000615597,
0.045252662152051926,
-0.01863105222582817,
0.059567518532276154,
-0.07970677316188812,
0.11320048570632935,
-0.057072367519140244,
0.018363293260335922,
-0.06628026068210602,
0.022379960864782333,
0.04326729103922844,
0.01599004492163658,
-0.03358958289027214,
0.018954860046505928,
-0.08429935574531555,
-0.10386499017477036,
0.0339457206428051,
-0.05247629061341286,
-0.00069722457556054,
0.02506854757666588,
0.004038164392113686,
0.005099293310195208,
0.03782220929861069,
-0.07093097269535065,
-0.014637337997555733,
-0.025811253115534782,
-0.15877743065357208,
0.0016609005397185683,
-0.0016606950666755438,
0.031047431752085686,
0.05591152235865593,
-0.07364784926176071,
0.06752905249595642,
-0.07606124877929688,
0.061677973717451096,
-0.010294376872479916,
-0.043152399361133575,
0.05035942792892456,
-0.010661059059202671,
0.00032298031146638095,
-0.046921174973249435,
-0.07642263174057007,
-0.0007151734316721559,
0.01548015046864748,
0.02260092832148075,
-0.022348396480083466,
0.03589851036667824,
0.07163773477077484,
-0.05787456035614014,
-0.041231896728277206,
0.024339662864804268,
-0.05396445095539093,
0.03944659233093262,
0.016098419204354286,
0.05780010297894478,
0.09542792290449142,
-0.020437132567167282,
0.0580039918422699,
-0.012736627832055092,
0.04457392543554306,
-0.04213707894086838,
0.03045251965522766,
-0.04010918736457825,
0.0426289327442646,
0.03210359066724777,
0.014653578400611877,
-0.04576808586716652,
0.030853619799017906,
-0.012396561913192272,
-0.09478142857551575,
0.06969911605119705,
-0.018544243648648262,
0.04800304025411606,
-0.10560419410467148,
-0.023082716390490532,
0.03903670981526375,
0.0042971400544047356,
0.04020819440484047,
0.0006522596813738346,
0.01512135099619627,
-0.026660453528165817,
-0.055250924080610275,
0.02168164774775505,
-0.06824562698602676,
-0.05686161667108536,
-0.08261507749557495,
-0.007776598446071148,
-0.024925472214818,
0.041199907660484314,
-0.0546199232339859,
-0.05365677922964096,
-5.319724021052744e-34,
-0.058844223618507385,
0.01019025407731533,
-0.0420600026845932,
0.012960403226315975,
-0.0699348896741867,
-0.14329762756824493,
0.029629282653331757,
-0.06915556639432907,
-0.0022605513222515583,
-0.06571611016988754,
0.01182567235082388,
-0.034760430455207825,
0.04564350098371506,
-0.016080299392342567,
0.024487588554620743,
-0.005182620137929916,
-0.07080956548452377,
0.006272892002016306,
0.0002903536369558424,
0.07287485897541046,
0.03904656320810318,
0.1071421355009079,
-0.13525526225566864,
-0.04468896985054016,
-0.00013595737982541323,
0.0662526860833168,
-0.03548156097531319,
0.05413387343287468,
0.1298893243074417,
0.013327055610716343,
-0.02807164005935192,
-0.046689409762620926,
0.06177542358636856,
-0.01642731763422489,
-0.08120683580636978,
-0.016791440546512604,
0.02437625639140606,
-0.005068451166152954,
-0.02638920582830906,
-0.001863438985310495,
0.011958343908190727,
-0.08917716145515442,
-0.028589414432644844,
0.04217652976512909,
0.010819305665791035,
-0.023266160860657692,
-0.006887711584568024,
0.040303125977516174,
-0.014026659540832043,
-0.002473636996001005,
-0.025069015100598335,
0.028324337676167488,
-0.10817426443099976,
0.06324964016675949,
0.009217239916324615,
-0.042720310389995575,
-0.05200236663222313,
-0.009423372335731983,
0.01876336894929409,
-0.03703375160694122,
0.002231125719845295,
0.034697264432907104,
0.1217963695526123,
-0.08729598671197891,
0.02745339833199978,
0.030833659693598747,
-0.03597772866487503,
-0.012310299091041088,
0.0030898461118340492,
-0.03399765118956566,
0.023735286667943,
-0.013884657062590122,
0.011960091069340706,
-0.030179359018802643,
0.02251305617392063,
-0.04864922910928726,
-0.02109125256538391,
-0.0031800742726773024,
-0.019833233207464218,
0.009582921862602234,
-0.026670265942811966,
-0.04183250293135643,
0.05993185192346573,
0.02661449834704399,
-0.007817122153937817,
0.009838893078267574,
0.03984801471233368,
0.0831688717007637,
0.03716764226555824,
0.06074080988764763,
-0.08315396308898926,
0.01902850903570652,
0.0713857114315033,
0.10099080950021744,
-0.06735158711671829,
-4.370594197666833e-8,
-0.05742260441184044,
0.05114486441016197,
-0.035940833389759064,
-0.03405727818608284,
0.05420875549316406,
-0.05448161065578461,
-0.013728806748986244,
0.08964141458272934,
-0.03503463789820671,
-0.004145598504692316,
0.0812610536813736,
-0.08485431224107742,
-0.04469708353281021,
-0.038152094930410385,
0.017692280933260918,
0.09595450013875961,
0.05620108172297478,
-0.02786806970834732,
-0.03104306384921074,
-0.01070578582584858,
-0.0761592835187912,
0.0654502734541893,
-0.04443499073386192,
0.006715547759085894,
-0.03341761231422424,
-0.07383976876735687,
-0.03368513286113739,
0.14808344841003418,
0.11623957008123398,
-0.00045661989133805037,
0.005693491548299789,
0.011627551168203354,
0.057948999106884,
-0.058300264179706573,
0.007755241822451353,
0.03746568039059639,
0.03555731847882271,
0.042778130620718,
0.012909666635096073,
-0.09495555609464645,
0.02345220185816288,
-0.015057496726512909,
-0.06356807798147202,
0.05764739587903023,
-0.006287550088018179,
-0.024138320237398148,
-0.027754168957471848,
-0.08030038326978683,
0.010547860525548458,
0.05096249282360077,
0.04438527300953865,
-0.02595609799027443,
-0.02378866635262966,
0.08095884323120117,
0.058309730142354965,
0.025702908635139465,
-0.028997596353292465,
-0.05614905431866646,
-0.03957744687795639,
0.013918334618210793,
-0.02486954815685749,
0.010896896943449974,
-0.026489021256566048,
0.01126345619559288
] | 0 | 0 | 0 | 0 |
|
Underdamped Diffusion Bridges with Applications to Sampling | https://openreview.net/forum?id=Q1QTxFm0Is | [
"Denis Blessing",
"Julius Berner",
"Lorenz Richter",
"Gerhard Neumann"
] | Poster | We provide a general framework for learning diffusion bridges that transport prior to target distributions. It includes existing diffusion models for generative modeling, but also underdamped versions with degenerate diffusion matrices, where the noise only acts in certain dimensions. Extending previous findings, our framework allows to rigorously show that score-matching in the underdamped case is indeed equivalent to maximizing a lower bound on the likelihood. Motivated by superior convergence properties and compatibility with sophisticated numerical integration schemes of underdamped stochastic processes, we propose *underdamped diffusion bridges*, where a general density evolution is learned rather than prescribed by a fixed noising process. We apply our method to the challenging task of sampling from unnormalized densities without access to samples from the target distribution. Across a diverse range of sampling problems, our approach demonstrates state-of-the-art performance, notably outperforming alternative methods, while requiring significantly fewer discretization steps and almost no hyperparameter tuning. | Variational Inference, Sampling, Diffusion Models | We provide a general framework for learning diffusion bridges, including degenerate noise and in particular underdamped versions, which are applied in the context of sampling. | 12,283 | 2503.01006 | [
-0.08322589099407196,
-0.09759040921926498,
0.058095838874578476,
0.04469788447022438,
0.050117701292037964,
-0.07577984035015106,
-0.05305895581841469,
-0.09457550942897797,
0.010197529569268227,
-0.03251691907644272,
-0.05424979701638222,
-0.011804361827671528,
-0.0010130765149369836,
-0.027774330228567123,
-0.013188608922064304,
-0.015364840626716614,
0.06160986050963402,
0.10180562734603882,
-0.019421573728322983,
0.0410836897790432,
-0.03136882558465004,
0.07346086204051971,
-0.021123351529240608,
-0.02494662255048752,
0.0706762969493866,
-0.03490104898810387,
0.06456779688596725,
-0.06250972300767899,
0.01299160998314619,
-0.06262440979480743,
0.0701923668384552,
0.03166733682155609,
-0.10963322222232819,
-0.03399277478456497,
0.011376614682376385,
-0.010296786203980446,
-0.0380266048014164,
0.01095703337341547,
0.03923392295837402,
0.03648507595062256,
0.06813468039035797,
0.048470914363861084,
0.013933737762272358,
0.008232633583247662,
0.04728381708264351,
-0.05547862499952316,
0.017450327053666115,
-0.08369405567646027,
0.018949292600154877,
0.04093950614333153,
0.032119281589984894,
0.0017748330719769,
-0.05035791173577309,
0.05934610590338707,
-0.03551952540874481,
-0.0280939768999815,
0.04979712888598442,
-0.026457827538251877,
-0.003221989842131734,
-0.0064729913137853146,
-0.0953366756439209,
-0.09422909468412399,
-0.028441499918699265,
-0.027227263897657394,
-0.06058606132864952,
-0.03647167235612869,
0.008303599432110786,
0.10454780608415604,
0.02342204377055168,
0.11894728243350983,
-0.023948466405272484,
0.06508860737085342,
-0.040719833225011826,
0.024688009172677994,
0.029382161796092987,
-0.0038422327488660812,
-0.001654113526456058,
-0.02356267161667347,
0.041145097464323044,
-0.08464917540550232,
0.06452616304159164,
0.017568273469805717,
0.024237703531980515,
-0.05767027661204338,
0.05101548880338669,
-0.023592131212353706,
0.007179951760917902,
0.05889609456062317,
0.06978128850460052,
-0.03269203007221222,
-0.06180175393819809,
0.11747818440198898,
-0.13563337922096252,
-0.008676383644342422,
-0.01058007963001728,
-0.01086434256285429,
0.0791914239525795,
-0.0039966171607375145,
0.059427451342344284,
0.10990872979164124,
-0.008233155123889446,
0.09392480552196503,
0.016937393695116043,
-0.0019667204469442368,
0.006028446834534407,
-0.11179064959287643,
0.02127767726778984,
0.01585445925593376,
0.031828612089157104,
-0.06264594197273254,
0.006511560641229153,
-0.04545477405190468,
0.06754307448863983,
0.05319368839263916,
0.005889147520065308,
-0.12106692790985107,
0.06087861582636833,
-0.02127758599817753,
-0.027184581384062767,
0.03501694276928902,
-0.03655039891600609,
0.037904590368270874,
-0.08945494145154953,
0.061245597898960114,
-0.038039568811655045,
0.04907074570655823,
-0.04594558849930763,
2.457048414131485e-33,
0.00792090967297554,
-0.05505136400461197,
0.010800355114042759,
0.03769916296005249,
0.03805786743760109,
0.004733514040708542,
-0.005848850589245558,
-0.051402609795331955,
0.007653719745576382,
0.0120280422270298,
-0.01694430224597454,
0.00578019954264164,
-0.11515408009290695,
0.04355796426534653,
0.002833031350746751,
-0.0487426221370697,
-0.04197373241186142,
0.1283416599035263,
0.009531737305223942,
-0.03475449979305267,
0.036006391048431396,
-0.021032145246863365,
-0.04764332249760628,
-0.03173266723752022,
-0.008249436505138874,
0.032849669456481934,
0.014392390847206116,
-0.040054578334093094,
0.07270240783691406,
0.022793037816882133,
-0.09236595034599304,
-0.03218444436788559,
0.0029295284766703844,
-0.0019080835627391934,
0.051358360797166824,
0.039719101041555405,
0.005430661141872406,
0.012828820385038853,
-0.08424128592014313,
-0.07032601535320282,
-0.03119540959596634,
0.007294332142919302,
0.01905420422554016,
-0.02888970635831356,
-0.14877501130104065,
-0.03774862736463547,
0.029502123594284058,
-0.05173817276954651,
0.03682577982544899,
-0.05499471351504326,
0.004781169351190329,
-0.07885447144508362,
-0.06180446222424507,
-0.04383375123143196,
-0.07283517718315125,
0.022585149854421616,
0.058289915323257446,
-0.05136018246412277,
-0.0068277702666819096,
0.00016259380208794028,
-0.011706070974469185,
-0.004065000917762518,
-0.03398391604423523,
-0.013537122868001461,
0.11588374525308609,
0.054337307810783386,
-0.027576949447393417,
0.0492553748190403,
0.05535057932138443,
-0.0225791297852993,
0.009517846629023552,
-0.012366055510938168,
-0.09421436488628387,
-0.077640600502491,
0.02466573938727379,
0.009101235307753086,
0.02989419549703598,
0.04790322482585907,
0.05194329842925072,
0.027200544252991676,
-0.04007820412516594,
-0.028941377997398376,
-0.07988382130861282,
-0.0940861776471138,
0.046509455889463425,
-0.01631160080432892,
-0.020459339022636414,
-0.06937225908041,
-0.09799257665872574,
-0.06997375190258026,
-0.06369415670633316,
-0.026502300053834915,
-0.02801256626844406,
-0.013998662121593952,
0.006745499093085527,
-4.5870147904626996e-33,
-0.05501016229391098,
0.044594816863536835,
0.04918213561177254,
0.05785994976758957,
-0.006126162596046925,
-0.0012349803000688553,
-0.016983218491077423,
0.017718655988574028,
0.07034651935100555,
-0.043426547199487686,
-0.06721040606498718,
-0.022264555096626282,
-0.0362989567220211,
-0.06162625923752785,
0.006267533637583256,
-0.02838888391852379,
0.04582640901207924,
0.07782625406980515,
-0.03264782205224037,
0.0033370228484272957,
-0.015497743152081966,
0.06012158840894699,
-0.02295360341668129,
0.04880821332335472,
-0.038220059126615524,
0.031003814190626144,
0.020611431449651718,
0.08134366571903229,
0.004599904641509056,
-0.03263446316123009,
-0.05074925348162651,
-0.04037317633628845,
0.0024251965805888176,
-0.03906131535768509,
-0.07535339891910553,
0.1298302561044693,
0.07576461881399155,
0.09160172939300537,
-0.027900174260139465,
0.036139097064733505,
0.042380791157484055,
-0.024736441671848297,
-0.05566041171550751,
0.01735592819750309,
0.08030398190021515,
-0.03256484866142273,
-0.08015535026788712,
0.022151652723550797,
0.04821530357003212,
0.028148621320724487,
0.043957844376564026,
0.0754823386669159,
-0.028527310118079185,
0.056165363639593124,
-0.0717439353466034,
-0.052357472479343414,
-0.03306959569454193,
-0.049901582300662994,
0.08446061611175537,
-0.0009950194507837296,
-0.045288074761629105,
0.002280130051076412,
-0.02159653604030609,
-0.038439080119132996,
-0.07654745876789093,
-0.018501490354537964,
-0.04005527123808861,
-0.01893731579184532,
-0.011646861210465431,
0.07327867299318314,
-0.03607964888215065,
-0.018018405884504318,
0.007680915761739016,
-0.014147310517728329,
0.017455533146858215,
-0.10770373791456223,
0.0753643736243248,
-0.012835265137255192,
-0.018828634172677994,
-0.07170828431844711,
0.02404635213315487,
0.00196597701869905,
0.018020685762166977,
0.013674238696694374,
0.1083500012755394,
-0.020748615264892578,
0.07617305964231491,
-0.0435834676027298,
0.02423369511961937,
-0.037182051688432693,
0.013441619463264942,
0.02100982517004013,
0.06319708377122879,
0.03246084973216057,
-0.13657845556735992,
-4.4520486852661634e-8,
-0.06777991354465485,
0.08430912345647812,
0.026243211701512337,
0.014611655846238136,
0.1194700226187706,
0.03472137078642845,
0.00114323734305799,
0.04389601945877075,
-0.01711236499249935,
-0.07638717442750931,
0.07920198142528534,
-0.015828201547265053,
-0.025670351460576057,
-0.02020297572016716,
-0.0017338086618110538,
0.05426177382469177,
-0.015092366375029087,
0.0035345929209142923,
-0.007517644204199314,
-0.023763900622725487,
-0.07203493267297745,
0.057979680597782135,
-0.004826230928301811,
0.0009045136393979192,
0.09703049808740616,
-0.01914897933602333,
-0.03384559229016304,
-0.0413818359375,
0.09061110019683838,
0.010062137618660927,
-0.008129047229886055,
0.05523541197180748,
0.03750873729586601,
0.04626663029193878,
0.03746756538748741,
0.027499137446284294,
0.059631358832120895,
0.0015725126722827554,
-0.04041576758027077,
-0.06753634661436081,
-0.060863394290208817,
0.06929807364940643,
-0.043184276670217514,
-0.02163134329020977,
0.017777668312191963,
0.07481534034013748,
0.022470494732260704,
0.0097665311768651,
0.0029571475461125374,
0.06853487342596054,
0.034978237003088,
-0.04321080446243286,
0.05492677167057991,
-0.014122073538601398,
0.044434841722249985,
-0.011208992451429367,
0.003134222701191902,
-0.006721838377416134,
0.029332809150218964,
0.058551184833049774,
0.05303898826241493,
-0.006225373595952988,
-0.0148355383425951,
0.11721755564212799
] | https://github.com/DenisBless/UnderdampedDiffusionBridges | 2 | 0 | 0 | 0 |
BoneMet: An Open Large-Scale Multi-Modal Murine Dataset for Breast Cancer Bone Metastasis Diagnosis and Prognosis | https://openreview.net/forum?id=YH4M1Tbxfz | [
"Tiankuo Chu",
"Fudong Lin",
"Shubo Wang",
"Jason Jiang",
"Wiley Jia-Wei Gong",
"Xu Yuan",
"Liyun Wang"
] | Poster | Breast cancer bone metastasis (BCBM) affects women’s health globally, calling
for the development of effective diagnosis and prognosis solutions. While deep
learning has exhibited impressive capacities across various healthcare domains, its
applicability in BCBM diseases is consistently hindered by the lack of an open,
large-scale, deep learning-ready dataset. As such, we introduce the Bone Metastasis
(BoneMet) dataset, the first large-scale, publicly available, high-resolution medical
resource, which is derived from a well-accepted murine BCBM model. The unique
advantage of BoneMet over existing human datasets is repeated sequential scans
per subject over the entire disease development phases. The dataset consists of
over 67 terabytes of multi-modal medical data, including 2D X-ray images, 3D
CT scans, and detailed biological data (e.g., medical records and bone quantitative
analysis), collected from more than five hundreds mice spanning from 2019 to
2024. Our BoneMet dataset is well-organized into six components, i.e., Rotation
X-Ray, Recon-CT, Seg-CT, Regist-CT, RoI-CT, and MiceMediRec. We further
show that BoneMet can be readily adopted to build versatile, large-scale AI models
for managing BCBM diseases in terms of diagnosis using 2D or 3D images, prognosis of bone deterioration, and sparse-angle 3D reconstruction for safe long-term
disease monitoring. Our preliminary results demonstrate that BoneMet has the
potentials to jump-start the development and fine-tuning of AI-driven solutions
prior to their applications to human patients. To facilitate its easy access and
wide dissemination, we have created the BoneMet package, providing three APIs
that enable researchers to (i) flexibly process and download the BoneMet data
filtered by specific time frames; and (ii) develop and train large-scale AI models for
precise BCBM diagnosis and prognosis. The BoneMet dataset is officially available on Hugging Face Datasets at https://huggingface.co/datasets/BoneMet/BoneMet. The BoneMet package is available on the Python Package Index (PyPI) at https://pypi.org/project/BoneMet. Code and tutorials are available at https://github.com/Tiankuo528/BoneMet. | Medical Dataset, Breast Cancer Bone Metastasis, Diagnosis, Prognosis, Sparse CT reconstruction, CT, X-ray, Large language model, AI for Science | null | 12,268 | null | [
-0.058619167655706406,
-0.07892053574323654,
0.04532402381300926,
-0.06462399661540985,
-0.004250353667885065,
-0.06355233490467072,
-0.027217134833335876,
0.026420142501592636,
-0.033289507031440735,
-0.048107877373695374,
-0.09412403404712677,
-0.024072058498859406,
-0.027214691042900085,
0.012527981773018837,
-0.0792912170290947,
0.01809277944266796,
-0.05970624089241028,
0.032306697219610214,
-0.10613946616649628,
0.05791125446557999,
0.015339486300945282,
0.03140995651483536,
0.07393274456262589,
0.07659175246953964,
0.019828755408525467,
-0.0445067435503006,
0.05595570057630539,
-0.037005335092544556,
-0.04853241518139839,
0.017143355682492256,
0.049150947481393814,
0.037922441959381104,
-0.022362075746059418,
0.00641368143260479,
-0.03461848944425583,
0.014552297070622444,
-0.04595528170466423,
0.06325869262218475,
-0.10019153356552124,
0.0029960665851831436,
0.04911092296242714,
0.007857994176447392,
0.012015871703624725,
0.10564979910850525,
0.09226265549659729,
-0.009232636541128159,
0.0016864389181137085,
-0.048162300139665604,
0.10801754146814346,
0.09969876706600189,
-0.08131973445415497,
-0.014319214969873428,
-0.04672365263104439,
0.16441163420677185,
0.07076188176870346,
-0.022976486012339592,
-0.052258580923080444,
-0.056393738836050034,
-0.05720863863825798,
0.0047196559607982635,
0.038426004350185394,
-0.08479589223861694,
0.0035480812657624483,
-0.011285712011158466,
0.05453934147953987,
0.04250658303499222,
0.0030100299045443535,
0.014719667844474316,
-0.07934372127056122,
-0.06542037427425385,
0.05358681455254555,
0.07387176156044006,
0.012337599880993366,
0.06026418134570122,
-0.005679983180016279,
0.04899542033672333,
0.09581338614225388,
0.015443511307239532,
0.10903651267290115,
-0.046201590448617935,
-0.02023601159453392,
0.062030624598264694,
0.056481558829545975,
-0.003516166005283594,
0.08559155464172363,
-0.06091798096895218,
0.006650902330875397,
0.06112007424235344,
-0.11087125539779663,
0.013681869022548199,
0.028733104467391968,
0.05285140872001648,
-0.0265986155718565,
-0.028621619567275047,
-0.017416339367628098,
0.059057846665382385,
0.024668356403708458,
-0.03947226330637932,
-0.013125178404152393,
0.06293871998786926,
-0.07080382108688354,
0.04884548857808113,
0.04422590881586075,
0.0067680999636650085,
0.026364969089627266,
-0.06124229356646538,
0.06756332516670227,
0.05530219152569771,
-0.02528902143239975,
-0.04505334794521332,
0.08909469097852707,
0.025270503014326096,
-0.013274814002215862,
0.012431175448000431,
0.09305550903081894,
-0.01224288996309042,
-0.02517545223236084,
-0.005036037880927324,
0.0022592623718082905,
-0.011420035734772682,
-0.0894840806722641,
-0.049212921410799026,
-0.038887470960617065,
0.015588798560202122,
-0.019592242315411568,
-0.022911785170435905,
-0.07166103273630142,
4.543041383508361e-33,
0.06789860874414444,
-0.002916238969191909,
0.06260527670383453,
0.015674035996198654,
-0.016088826581835747,
-0.001704962458461523,
-0.045889925211668015,
-0.01897622086107731,
-0.08935356140136719,
0.059668149799108505,
-0.030106237158179283,
0.055547356605529785,
-0.027400823310017586,
0.022677753120660782,
0.011136477813124657,
-0.017824780195951462,
-0.010944907553493977,
0.04030318558216095,
-0.009453405626118183,
0.005239912308752537,
0.0845608189702034,
-0.023787859827280045,
-0.0024036893155425787,
-0.03610469028353691,
-0.03367455303668976,
0.03151683509349823,
-0.037322916090488434,
0.032120414078235626,
-0.015062550082802773,
0.02757435105741024,
-0.11654164642095566,
-0.008478745818138123,
0.06338455528020859,
0.061604127287864685,
-0.0021500929724425077,
-0.07056035101413727,
0.0012824643636122346,
-0.001722687971778214,
0.053106218576431274,
0.005534040741622448,
0.08321285247802734,
-0.004367819055914879,
0.03647002950310707,
-0.01015633437782526,
0.00022009341046214104,
-0.021153897047042847,
0.020421326160430908,
-0.05197988823056221,
-0.03654702007770538,
-0.06945019960403442,
0.04384668543934822,
-0.02440977282822132,
-0.03939981013536453,
-0.05374792590737343,
-0.039164651185274124,
0.04129919409751892,
-0.04606446251273155,
-0.04145587608218193,
0.009405424818396568,
0.04828251898288727,
0.016372447833418846,
-0.008827737532556057,
0.004015250131487846,
0.08524366468191147,
-0.02673356607556343,
0.011520623229444027,
-0.0021717227064073086,
-0.012712309136986732,
-0.009918441995978355,
0.06530996412038803,
0.02529238536953926,
-0.0020010038278996944,
0.013198795728385448,
-0.052248191088438034,
0.015564220026135445,
0.04961872845888138,
0.07820555567741394,
-0.012695906683802605,
-0.10732576251029968,
0.0866536796092987,
0.03740110993385315,
0.020804159343242645,
-0.12418947368860245,
0.07085152715444565,
-0.03607403114438057,
-0.04371841624379158,
0.002635902725160122,
-0.05961098521947861,
0.01759014092385769,
-0.041067544370889664,
-0.08439534157514572,
0.010790972039103508,
-0.05130591243505478,
-0.08652926236391068,
-0.06796152144670486,
-3.1416361202412307e-33,
0.03141631931066513,
0.01583622209727764,
-0.0023635677061975002,
-0.03327913582324982,
-0.0008739414624869823,
-0.009933492168784142,
0.0040952228009700775,
-0.011801243759691715,
0.013958805240690708,
-0.04495788365602493,
0.07887905091047287,
-0.04461890831589699,
-0.03735078498721123,
-0.10579065978527069,
-0.026097740978002548,
0.027951369062066078,
0.006505101919174194,
-0.02970176190137863,
-0.06991972029209137,
-0.06462021917104721,
0.020944319665431976,
0.11578452587127686,
-0.04800531640648842,
0.04454360157251358,
-0.08800023794174194,
0.02531985193490982,
-0.007829045876860619,
0.04075506329536438,
0.011624614708125591,
-0.002931789727881551,
-0.08870945870876312,
-0.02382710762321949,
-0.030915778130292892,
-0.03157282993197441,
-0.07911963760852814,
0.09527958929538727,
0.029136914759874344,
-0.08399337530136108,
-0.013622057624161243,
-0.050758201628923416,
0.056352656334638596,
0.01395415049046278,
-0.08764591068029404,
0.0154146458953619,
-0.024028150364756584,
-0.0266328863799572,
0.021791158244013786,
0.07359138131141663,
0.12287183851003647,
0.06758230179548264,
-0.010296701453626156,
-0.026809411123394966,
0.008461102843284607,
-0.06679311394691467,
0.0012288173893466592,
-0.07586246728897095,
-0.023427704349160194,
-0.0897630900144577,
0.037463001906871796,
0.019626354798674583,
-0.05000528693199158,
-0.0792389065027237,
-0.047443654388189316,
-0.022890286520123482,
-0.0550132654607296,
0.0481664314866066,
0.06770026683807373,
-0.06671405583620071,
-0.03389906510710716,
0.06058134511113167,
0.010077999904751778,
-0.026823099702596664,
-0.00035831061541102827,
0.046001750975847244,
-0.041518595069646835,
0.010518638417124748,
-0.004369264468550682,
-0.020824676379561424,
0.0009451532387174666,
0.015682177618145943,
0.07389888167381287,
-0.04353046789765358,
0.014487948268651962,
0.06809850037097931,
0.11519879847764969,
0.072664275765419,
0.006967829540371895,
0.05420288071036339,
0.022652361541986465,
0.03083137236535549,
-0.08013741672039032,
0.049099285155534744,
-0.017496800050139427,
0.02493681199848652,
-0.021220678463578224,
-4.028827405022639e-8,
-0.0339675173163414,
0.07035553455352783,
0.04419045150279999,
-0.04028335586190224,
0.038721468299627304,
-0.011464761570096016,
-0.04729326441884041,
0.08772943168878555,
0.05655961483716965,
0.007045174483209848,
0.07513014227151871,
0.0077959708869457245,
-0.06313759088516235,
-0.06667967885732651,
-0.010193921625614166,
0.1040559858083725,
0.062252793461084366,
-0.01814945414662361,
-0.012818505987524986,
-0.0284747164696455,
-0.006551473401486874,
0.02279737964272499,
0.06295416504144669,
-0.04900642856955528,
0.09082389622926712,
-0.057622335851192474,
-0.07763973623514175,
0.0590585432946682,
-0.022940563037991524,
0.062214676290750504,
0.04589691758155823,
-0.04890129715204239,
0.07279060035943985,
0.0011573327938094735,
0.009079286828637123,
0.0343603678047657,
0.01887611113488674,
0.0012999358586966991,
-0.06259561330080032,
-0.05871474742889404,
-0.003346680663526058,
0.04200167953968048,
0.012478114105761051,
0.023040303960442543,
-0.06892366707324982,
-0.010227102786302567,
0.0019246116280555725,
0.05363764241337776,
0.06398030370473862,
0.006417363882064819,
0.010925930924713612,
-0.024522902444005013,
0.05691666156053543,
0.06416105479001999,
0.03323524817824364,
0.12928159534931183,
0.01864754594862461,
-0.05185797065496445,
0.11604198068380356,
0.06405535340309143,
0.029417002573609352,
-0.04424366354942322,
-0.07259051501750946,
-0.0028449869714677334
] | 0 | 0 | 0 | 0 |
|
Time After Time: Deep-Q Effect Estimation for Interventions on When and What to do | https://openreview.net/forum?id=5yDS32hKJc | [
"Yoav Wald",
"Mark Goldstein",
"Yonathan Efroni",
"Wouter A.C. van Amsterdam",
"Rajesh Ranganath"
] | Poster | Problems in fields such as healthcare, robotics, and finance requires reasoning about the value both of what decision or action to take and when to take it. The prevailing hope is that artificial intelligence will support such decisions by estimating the causal effect of policies such as how to treat patients or how to allocate resources over time. However, existing methods for estimating the effect of a policy struggle with \emph{irregular time}. They either discretize time, or disregard the effect of timing policies.
We present a new deep-Q algorithm that estimates the effect of both when and what to do called Earliest Disagreement Q-Evaluation (EDQ). EDQ makes use of recursion for the Q-function that is compatible with flexible sequence models, such as transformers. EDQ provides accurate estimates under standard assumptions. We validate the approach through experiments on survival time and tumor growth tasks. | effect estimation, treatment times, irregular times, sequential decision making | A deep-Q effect estimation method for intervention on times and types of treatments | 12,267 | null | [
-0.12066541612148285,
0.025927526876330376,
0.06427126377820969,
0.0037948614917695522,
-0.004067863803356886,
0.026119545102119446,
-0.000687497085891664,
0.06638368219137192,
0.09051978588104248,
0.006392750423401594,
-0.02931557595729828,
0.016196437180042267,
0.0015879306010901928,
0.08406534790992737,
-0.06913849711418152,
0.028818078339099884,
0.040910229086875916,
-0.006398702040314674,
-0.09864477813243866,
-0.0760422870516777,
0.005257748067378998,
0.06904854625463486,
0.039203353226184845,
-0.032273873686790466,
0.02147255465388298,
-0.032624565064907074,
-0.05089936405420303,
0.002997514558956027,
0.010340102016925812,
-0.016576889902353287,
0.04612163081765175,
-0.03577927127480507,
0.030725659802556038,
-0.032057005912065506,
-0.023661814630031586,
0.08271521329879761,
-0.05451427772641182,
0.007515492383390665,
-0.05394912511110306,
0.027787890285253525,
-0.02501702681183815,
-0.007974390871822834,
0.017557861283421516,
0.04028572142124176,
0.1247827336192131,
-0.020777655765414238,
-0.0067961979657411575,
0.006062882952392101,
-0.033164747059345245,
0.023100001737475395,
-0.04437028616666794,
0.04663577303290367,
0.058324698358774185,
0.06526420265436172,
0.05991889536380768,
0.05204365774989128,
0.034419093281030655,
0.023615000769495964,
-0.0591931976377964,
-0.03715033084154129,
-0.02717144787311554,
-0.09468913823366165,
-0.013137402944266796,
-0.034133825451135635,
-0.013349950313568115,
0.07169892638921738,
0.010236810892820358,
-0.022504888474941254,
-0.08122082054615021,
0.08853618055582047,
-0.01831214874982834,
0.08834940940141678,
-0.047892745584249496,
0.0037751791533082724,
-0.005703657865524292,
0.0024276310577988625,
0.07867079973220825,
0.08918067067861557,
0.08036617934703827,
-0.10318528860807419,
0.03170529007911682,
-0.0458650141954422,
-0.033834513276815414,
-0.008215409703552723,
0.08790841698646545,
-0.045378655195236206,
-0.0024854287039488554,
0.10125987231731415,
0.019265474751591682,
0.014302104711532593,
0.015804463997483253,
-0.04416659101843834,
0.017950912937521935,
-0.00943725649267435,
0.04465436935424805,
0.03595663979649544,
-0.02653535082936287,
-0.09665641188621521,
0.02656269259750843,
0.07353042811155319,
-0.024709351360797882,
-0.008572082035243511,
-0.032892003655433655,
-0.006036275532096624,
0.006168825551867485,
-0.04047517478466034,
0.02679808810353279,
-0.011224644258618355,
0.030886350199580193,
-0.0265909843146801,
0.0193072110414505,
-0.04147955775260925,
0.11912670731544495,
0.08198628574609756,
0.004683495499193668,
-0.000059872436395380646,
0.00794023647904396,
0.030796417966485023,
0.014039321802556515,
0.026084573939442635,
0.012383408844470978,
-0.0727342963218689,
-0.05437956005334854,
-0.001481164013966918,
0.014245349913835526,
0.0934445783495903,
-0.032002005726099014,
-3.636974011314388e-34,
0.0029149854090064764,
-0.0683765560388565,
0.07922300696372986,
0.021886508911848068,
0.023921364918351173,
-0.0392766110599041,
-0.027023786678910255,
-0.023307770490646362,
-0.021028641611337662,
0.01694764755666256,
-0.012841131538152695,
-0.025624124333262444,
-0.05203265696763992,
-0.034323375672101974,
0.011078259907662868,
0.046386949717998505,
-0.0662987232208252,
0.07121436297893524,
-0.055570777505636215,
-0.0303492471575737,
0.06853294372558594,
-0.11575571447610855,
-0.0682201161980629,
-0.022147752344608307,
0.0435858815908432,
-0.022708360105752945,
0.02877194993197918,
0.025181559845805168,
-0.0029990316834300756,
-0.028237199410796165,
-0.14013884961605072,
0.05774220451712608,
-0.0278497114777565,
0.0009416482062079012,
0.006986909080296755,
0.002037693979218602,
-0.0264633446931839,
0.04435049369931221,
0.01563127711415291,
0.020195651799440384,
-0.014864767901599407,
0.020348457619547844,
0.02074481174349785,
-0.03441241756081581,
-0.0252681951969862,
-0.12033391743898392,
0.04618200287222862,
-0.08407291024923325,
-0.02332245372235775,
-0.03909962251782417,
0.007019595708698034,
-0.0010941835353150964,
0.05123238265514374,
-0.12631507217884064,
0.019963443279266357,
-0.00832124799489975,
-0.020899690687656403,
0.059240762144327164,
-0.10408730804920197,
0.05618741363286972,
0.14384089410305023,
-0.0003979023895226419,
-0.0044946507550776005,
0.011967995204031467,
-0.061801500618457794,
0.09598003327846527,
-0.01869647018611431,
0.052646905183792114,
0.06722103804349899,
0.05075610801577568,
0.030385145917534828,
0.03687524050474167,
-0.01382433995604515,
-0.05974270775914192,
0.023546472191810608,
-0.02443876676261425,
0.07582475990056992,
-0.07011041045188904,
-0.01218823529779911,
0.010795003734529018,
0.0410211980342865,
-0.03514271229505539,
-0.04765836149454117,
0.07619091868400574,
0.029555197805166245,
-0.10841699689626694,
0.017487727105617523,
-0.016127927228808403,
-0.03526507690548897,
-0.07994488626718521,
-0.09999064356088638,
-0.016260577365756035,
0.0027562566101551056,
0.06273046880960464,
-0.008504820987582207,
-1.5278269256331826e-33,
-0.050938837230205536,
0.016744889318943024,
0.04278387501835823,
0.049027036875486374,
0.04209866747260094,
-0.06198519468307495,
-0.08469673991203308,
0.007413999177515507,
0.12386036664247513,
-0.04006829112768173,
-0.02439318411052227,
0.014736897312104702,
0.03925324231386185,
-0.008431886322796345,
0.0023984750732779503,
-0.02965863235294819,
-0.019105248153209686,
-0.11665515601634979,
0.012759462930262089,
0.04561324417591095,
0.06395653635263443,
-0.016507970169186592,
-0.06904608011245728,
0.0029934528283774853,
-0.04934750124812126,
0.039581868797540665,
-0.016413934528827667,
0.021011780947446823,
0.009900142438709736,
-0.07164425402879715,
-0.11240533739328384,
-0.1497253179550171,
-0.057023677974939346,
0.013813192024827003,
-0.021593937650322914,
0.030937114730477333,
0.03498314693570137,
-0.04510048031806946,
0.007787782233208418,
-0.01880035176873207,
0.09365331381559372,
-0.05950529873371124,
-0.06492528319358826,
0.004919623956084251,
-0.018278734758496284,
-0.05101493373513222,
-0.004909204319119453,
0.012609001249074936,
0.03534328192472458,
0.03742039203643799,
0.03922192379832268,
0.060891538858413696,
-0.05909489467740059,
0.05968040972948074,
-0.07905901968479156,
-0.057121340185403824,
0.0016682641580700874,
-0.05252951756119728,
-0.01459614560008049,
0.09899888932704926,
-0.10971817374229431,
0.03181682154536247,
-0.05508062243461609,
0.04035786911845207,
-0.054105162620544434,
0.07601337879896164,
-0.019379684701561928,
-0.006428385153412819,
0.04878772422671318,
-0.034287650138139725,
0.06384146958589554,
0.03343772143125534,
-0.04652457311749458,
0.038219209760427475,
0.014589081518352032,
-0.013454003259539604,
-0.016320494934916496,
0.00937362015247345,
0.004387521184980869,
-0.014041891321539879,
0.025949208065867424,
-0.03772460296750069,
0.023180199787020683,
0.0053978594951331615,
-0.1005391776561737,
0.05283942073583603,
0.06591064482927322,
0.02660602517426014,
0.04926125705242157,
0.01800321787595749,
-0.07498837262392044,
-0.01632952131330967,
-0.00478975148871541,
0.024784037843346596,
-0.13530713319778442,
-4.6699380362724696e-8,
0.06386435776948929,
0.014155317097902298,
0.07656393200159073,
0.0059431567788124084,
0.06559624522924423,
-0.06408993154764175,
-0.017266979441046715,
0.045797426253557205,
0.010277838446199894,
-0.017869731411337852,
0.09928429871797562,
0.08487372100353241,
0.07411760836839676,
-0.047858722507953644,
-0.05940196290612221,
0.10267245769500732,
0.03967146947979927,
-0.011185365729033947,
0.001651756465435028,
-0.0030112825334072113,
0.004865264520049095,
0.020183216780424118,
0.013853831216692924,
-0.028728129342198372,
0.0034911225084215403,
-0.05938682705163956,
-0.03357604518532753,
0.05285756662487984,
0.001450020237825811,
0.029650211334228516,
0.021194230765104294,
0.020467743277549744,
-0.010437838733196259,
0.007746112998574972,
-0.056033577769994736,
-0.021730056032538414,
0.005490653216838837,
-0.036843191832304,
0.01002892293035984,
-0.028176898136734962,
0.005039057228714228,
0.10592571645975113,
-0.05901317670941353,
0.0074818735010921955,
0.0026739216409623623,
-0.03880099207162857,
-0.1360001116991043,
-0.07684477418661118,
-0.009744241833686829,
-0.06711561977863312,
0.04538508132100105,
0.027978215366601944,
0.03100827895104885,
0.029306825250387192,
0.10140590369701385,
0.06481581181287766,
0.09792309254407883,
-0.07528093457221985,
-0.0028577072080224752,
0.03657468780875206,
0.01310095377266407,
0.010797932744026184,
-0.02500062808394432,
-0.015446425415575504
] | 0 | 0 | 0 | 0 |
|
GRAIN: Exact Graph Reconstruction from Gradients | https://openreview.net/forum?id=7bAjVh3CG3 | [
"Maria Drencheva",
"Ivo Petrov",
"Maximilian Baader",
"Dimitar Iliev Dimitrov",
"Martin Vechev"
] | Poster | Federated learning claims to enable collaborative model training among multiple clients with data privacy by transmitting gradient updates instead of the actual client data. However, recent studies have shown the client privacy is still at risk due to the, so called, gradient inversion attacks which can precisely reconstruct clients' text and image data from the shared gradient updates. While these attacks demonstrate severe privacy risks for certain domains and architectures, the vulnerability of other commonly-used data types, such as graph-structured data, remain under-explored. To bridge this gap, we present GRAIN, the first exact gradient inversion attack on graph data in the honest-but-curious setting that recovers both the structure of the graph and the associated node features. Concretely, we focus on Graph Convolutional Networks (GCN) and Graph Attention Networks (GAT) -- two of the most widely used frameworks for learning on graphs. Our method first utilizes the low-rank structure of GNN gradients to efficiently reconstruct and filter the client subgraphs which are then joined to complete the input graph. We evaluate our approach on molecular, citation, and social network datasets using our novel metric.
We show that GRAIN reconstructs up to 80\% of all graphs exactly, significantly outperforming the baseline, which achieves up to 20\% correctly positioned nodes. | gradient leakage, gradient inversion, graph neural networks, federated learning, graph convolutional networks, gnn, gcn, attack, privacy, reconstruction | We present GRAIN, the first gradient leakage attack designed specifically for graph neural networks, and show we achieve a high fraction of exact reconstructions, and outperform existing attacks on partial reconstruction. | 12,256 | 2503.01838 | [
-0.14671529829502106,
-0.04056413471698761,
0.05850401520729065,
0.048240743577480316,
0.10280391573905945,
-0.016972055658698082,
-0.04546563699841499,
-0.06777983158826828,
0.047838035970926285,
-0.004525312688201666,
0.035700827836990356,
0.017373744398355484,
0.061078354716300964,
0.030432922765612602,
-0.021172238513827324,
0.029903462156653404,
0.06268645823001862,
0.11325183510780334,
-0.06133892387151718,
-0.056371282786130905,
-0.07978268712759018,
0.009351540356874466,
0.010323813185095787,
0.00841814186424017,
0.06532563269138336,
-0.0001426950329914689,
-0.0015435831155627966,
-0.06850964576005936,
-0.01924852468073368,
-0.006537408567965031,
0.07502344250679016,
-0.028524233028292656,
-0.061006076633930206,
0.04983893036842346,
-0.07277511805295944,
0.0793541818857193,
-0.030981509014964104,
0.006082774605602026,
0.06624127924442291,
-0.024612536653876305,
-0.03444819897413254,
0.006246580742299557,
0.012872844003140926,
0.08543256670236588,
0.02115660533308983,
0.03116416558623314,
0.020150654017925262,
-0.033737193793058395,
0.020724749192595482,
-0.0030063947197049856,
-0.0378839485347271,
-0.04641950502991676,
-0.0915968120098114,
0.027163958176970482,
0.023055972531437874,
-0.0330980122089386,
-0.04378784820437431,
-0.022967126220464706,
0.007713287137448788,
0.07122387737035751,
-0.043324731290340424,
-0.06879159808158875,
-0.031773749738931656,
0.0034534772858023643,
-0.010410516522824764,
0.053776830434799194,
0.016341863200068474,
0.026661798357963562,
0.020929263904690742,
0.04648866504430771,
0.06276865303516388,
0.03906853124499321,
-0.08687703311443329,
-0.033208128064870834,
0.04108276218175888,
0.10183180868625641,
0.005060824565589428,
-0.002567423041909933,
0.0058598024770617485,
-0.09898146241903305,
0.06030814349651337,
0.013543413951992989,
0.10426393151283264,
-0.020810337737202644,
0.061858002096414566,
-0.01758558303117752,
-0.03693130239844322,
0.0217896718531847,
-0.004152107983827591,
0.07080347090959549,
-0.09673716127872467,
-0.0007791330572217703,
0.0038342196494340897,
-0.024223800748586655,
0.05849995091557503,
0.026483595371246338,
-0.010430962778627872,
-0.05032973363995552,
-0.034985266625881195,
0.08952496200799942,
-0.06436029821634293,
0.030254963785409927,
0.0508926585316658,
-0.019307196140289307,
0.01548341941088438,
0.062070559710264206,
0.02885652706027031,
0.029252538457512856,
0.03667733445763588,
-0.004574817139655352,
0.04561057314276695,
0.11519358307123184,
-0.03053288348019123,
0.014864198863506317,
0.05396689474582672,
-0.06105199456214905,
0.004583962261676788,
-0.032479528337717056,
-0.006107233464717865,
0.08125831931829453,
-0.025999218225479126,
0.03458058834075928,
0.0069086989387869835,
-0.03962114453315735,
-0.05985618755221367,
-0.019069863483309746,
-0.05357344448566437,
-5.134944574298135e-34,
0.08668336272239685,
0.057636383920907974,
0.05123930796980858,
-0.007556327152997255,
0.03588512912392616,
0.0016332062659785151,
-0.06558407843112946,
0.005903766490519047,
-0.08034568279981613,
0.02583897113800049,
-0.027728548273444176,
0.006249834783375263,
0.016646383330225945,
0.09171011298894882,
0.005632475484162569,
0.028934059664607048,
0.004659637808799744,
-0.009562241844832897,
0.031944356858730316,
-0.07559429109096527,
-0.00009919350122800097,
-0.05887901782989502,
0.030086826533079147,
0.0129804452881217,
0.012982026673853397,
-0.0469084233045578,
0.03647240251302719,
0.009747307747602463,
0.023297376930713654,
-0.006483028177171946,
-0.09471265226602554,
0.05144454911351204,
0.01684873178601265,
0.08293534815311432,
-0.0028965233359485865,
0.003914728760719299,
0.010415682569146156,
-0.07875682413578033,
0.03002380020916462,
-0.007130508776754141,
0.04555428400635719,
0.037050675600767136,
-0.03823329880833626,
-0.04549449309706688,
-0.07687413692474365,
-0.005489340052008629,
-0.03921416029334068,
-0.012700938619673252,
0.00016204781422857195,
-0.051051292568445206,
-0.0033046556636691093,
-0.0035038411151617765,
-0.057847872376441956,
0.030977727845311165,
-0.06112875044345856,
-0.0015497029526159167,
0.08890853077173233,
0.046340879052877426,
0.034650709480047226,
0.04133841395378113,
0.05282720550894737,
0.043533336371183395,
0.004315802827477455,
0.009231709875166416,
-0.057016149163246155,
0.017823785543441772,
-0.047458816319704056,
0.018842587247490883,
0.029325416311621666,
0.009311680682003498,
-0.020897909998893738,
0.06179723143577576,
-0.04822487384080887,
-0.04396714270114899,
0.014432227239012718,
-0.006130352150648832,
-0.010112826712429523,
-0.029166201129555702,
-0.01603127457201481,
0.027586679905653,
0.0029899454675614834,
-0.09255331754684448,
0.009020885452628136,
-0.0031882966868579388,
-0.16858947277069092,
-0.004154205322265625,
0.018658386543393135,
-0.047397222369909286,
-0.00843634270131588,
0.04156406968832016,
-0.04871851205825806,
-0.07373271882534027,
-0.049357641488313675,
-0.0089877275750041,
-0.0016807038336992264,
9.937442815794418e-34,
-0.05222957953810692,
0.08977758139371872,
-0.040530525147914886,
0.018139008432626724,
-0.00041865481762215495,
-0.08864068984985352,
0.006276415195316076,
0.011305143125355244,
-0.025988027453422546,
0.04982466623187065,
-0.048543717712163925,
-0.06284600496292114,
0.058187030255794525,
-0.0716041699051857,
0.03997857868671417,
0.008894221857190132,
-0.04744504392147064,
0.021109089255332947,
-0.09964825212955475,
-0.028990058228373528,
-0.010597223415970802,
0.08071859180927277,
-0.062387723475694656,
0.04387039691209793,
0.09247693419456482,
0.03867533430457115,
-0.09375713020563126,
0.039596982300281525,
0.06213158741593361,
-0.01924518495798111,
-0.07682126760482788,
0.008318967185914516,
-0.03197738155722618,
-0.1195618212223053,
-0.04534477740526199,
0.05818213149905205,
0.041846565902233124,
0.02830318547785282,
-0.012355705723166466,
0.01439684722572565,
0.04332001134753227,
0.00874361302703619,
-0.1270366907119751,
0.0420214980840683,
0.012161866761744022,
0.0340690053999424,
-0.028723284602165222,
0.01879039965569973,
-0.030172333121299744,
0.02576729841530323,
-0.0039291055873036385,
0.05252540856599808,
0.03888203948736191,
0.016624199226498604,
-0.08220189064741135,
-0.09001440554857254,
0.08253154158592224,
0.06510865688323975,
0.10222148895263672,
0.008714389987289906,
-0.021172184497117996,
-0.06337710469961166,
-0.0551808662712574,
-0.01751924492418766,
-0.013875816948711872,
-0.048628613352775574,
0.002863666508346796,
0.00890709925442934,
-0.01674622856080532,
0.07170715928077698,
0.016512662172317505,
-0.04225419461727142,
-0.010088296607136726,
0.013507944531738758,
0.09241873770952225,
-0.03676718845963478,
0.03398366644978523,
-0.007509376388043165,
-0.05413605272769928,
0.009596487507224083,
0.1282794326543808,
-0.023122865706682205,
0.04103303328156471,
0.07120169699192047,
0.15173836052417755,
0.0007072989828884602,
0.048621661961078644,
0.038127291947603226,
0.009593249298632145,
0.02754185162484646,
-0.049448657780885696,
-0.02711537294089794,
-0.14448414742946625,
0.03740086033940315,
0.06280911713838577,
-3.898534828294942e-8,
-0.12455876916646957,
0.01144986692816019,
-0.027601299807429314,
0.020889999344944954,
0.024042639881372452,
0.007456005085259676,
0.0769721120595932,
0.13023267686367035,
-0.044297024607658386,
0.01896006613969803,
0.08927328884601593,
-0.041224949061870575,
-0.06960565596818924,
-0.07954999804496765,
-0.054173585027456284,
0.06078799441456795,
-0.026246769353747368,
-0.048802074044942856,
0.03265194594860077,
0.03248947858810425,
-0.06093590706586838,
0.005165452603250742,
0.014348901808261871,
0.021630438044667244,
0.0479704774916172,
-0.0574386902153492,
-0.008494282141327858,
0.035460781306028366,
0.04257657751441002,
0.019688919186592102,
-0.08086531609296799,
-0.052674248814582825,
0.10097995400428772,
-0.04034913703799248,
-0.06927838921546936,
0.14575114846229553,
0.04584081470966339,
-0.023242659866809845,
0.007508596405386925,
0.006817623507231474,
-0.003730845870450139,
0.04890719801187515,
0.023011010140180588,
0.0015176027081906796,
-0.0686660185456276,
-0.0004923644009977579,
0.013873285613954067,
0.022756077349185944,
0.080319344997406,
-0.024171022698283195,
-0.0023577858228236437,
-0.07348618656396866,
-0.008584368973970413,
0.06474326550960541,
-0.004833714570850134,
-0.05763310566544533,
-0.04401955008506775,
-0.054961301386356354,
0.07814066112041473,
0.060748957097530365,
0.050753019750118256,
0.026589157059788704,
-0.04027210548520088,
-0.06506280601024628
] | https://github.com/insait-institute/grain | 0 | 0 | 0 | 0 |
ALLaM: Large Language Models for Arabic and English | https://openreview.net/forum?id=MscdsFVZrN | [
"M Saiful Bari",
"Yazeed Alnumay",
"Norah A. Alzahrani",
"Nouf M. Alotaibi",
"Hisham Abdullah Alyahya",
"Sultan AlRashed",
"Faisal Abdulrahman Mirza",
"Shaykhah Z. Alsubaie",
"Hassan A. Alahmed",
"Ghadah Alabduljabbar",
"Raghad Alkhathran",
"Yousef Almushayqih",
"Raneem Alnajim",
"Salman Alsubaihi",
"Maryam Al Mansour",
"Saad Amin Hassan",
"Dr. Majed Alrubaian",
"Ali Alammari",
"Zaki Alawami",
"Abdulmohsen Al-Thubaity",
"et al. (6 additional authors not shown)"
] | Poster | In this work, we present ALLaM: Arabic Large Language Model, a series of large language models to support the ecosystem of Arabic Language Technologies (ALT). ALLaM is carefully trained, considering the values of language alignment and transferability of knowledge at scale. The models are based on an autoregressive decoder-only architecture and are pretrained on a mixture of Arabic and English texts. We illustrate how the second-language acquisition via vocabulary expansion can help steer a language model towards a new language without any major catastrophic forgetting in English. Furthermore, we highlight the effectiveness of using translation data and the process of knowledge encoding within the language model's latent space. Finally, we show that effective alignment with human preferences can significantly enhance the performance of a large language model (LLM) compared to less aligned models of a larger scale. Our methodology enables us to achieve state-of-the-art performance in various Arabic benchmarks, including MMLU Arabic, ACVA, and Arabic Exams. Our aligned models improve both in Arabic and English from its base aligned models. | Large Language Model, English, Arabic, Second Language Acquisition | We trained a foundational Arabic LLM by leveraging English LLMs. | 12,255 | 2407.15390 | [
-0.030840909108519554,
-0.10755730420351028,
0.006382693070918322,
0.006749855354428291,
-0.056850939989089966,
0.04805203154683113,
-0.01559500303119421,
-0.05998482182621956,
0.06461197882890701,
0.020834123715758324,
0.05374441295862198,
-0.02584417164325714,
0.06844446808099747,
-0.007650523446500301,
-0.014599772170186043,
-0.032744672149419785,
0.033275406807661057,
0.06017190217971802,
-0.06387471407651901,
-0.11567331105470657,
-0.014561528339982033,
0.051853328943252563,
0.042175907641649246,
0.012256160378456116,
0.05660971999168396,
0.03468772768974304,
-0.02153283730149269,
0.01565949060022831,
-0.012769476510584354,
-0.08431665599346161,
0.04464659094810486,
0.03678170591592789,
0.01353730820119381,
0.04699130356311798,
-0.05720105767250061,
0.04944250360131264,
-0.07180917263031006,
-0.015601378865540028,
-0.00665208650752902,
-0.01326118502765894,
0.0073557584546506405,
-0.04329675808548927,
0.02157360129058361,
-0.012377255596220493,
0.032396234571933746,
-0.03382612019777298,
0.009069911204278469,
0.007008786778897047,
0.06359950453042984,
0.08178122341632843,
-0.1155472919344902,
0.009691814891994,
0.008003298193216324,
-0.00003850104258162901,
-0.01215716265141964,
-0.023223580792546272,
0.054495226591825485,
0.004535455722361803,
-0.03162194415926933,
-0.04963469132781029,
-0.07986621558666229,
-0.059021707624197006,
-0.053618792444467545,
0.033927176147699356,
-0.049637164920568466,
-0.07822200655937195,
0.06716335564851761,
0.04861314222216606,
-0.026145553216338158,
0.030021199956536293,
-0.05459447205066681,
0.06613282859325409,
0.005766106303781271,
0.03698970749974251,
0.02267017774283886,
0.025899773463606834,
0.012803563848137856,
-0.09052591025829315,
0.00843272265046835,
0.015249684453010559,
-0.016935091465711594,
0.05047275125980377,
0.04762706905603409,
0.011338675394654274,
0.09289739280939102,
-0.01859697513282299,
0.05027936398983002,
0.018414607271552086,
-0.0017809591954573989,
-0.061879485845565796,
-0.033120982348918915,
-0.05671267583966255,
0.05666600540280342,
0.03515520319342613,
0.08187753707170486,
0.07223083078861237,
0.0710856020450592,
-0.060669977217912674,
-0.05312173441052437,
0.08096291124820709,
0.015824206173419952,
0.08289992809295654,
-0.02067100815474987,
0.005947317462414503,
-0.05051996931433678,
-0.08394604176282883,
0.01135024894028902,
0.06202905997633934,
0.04060244932770729,
-0.1234550029039383,
0.0018935624975711107,
0.01701178587973118,
-0.0253007635474205,
0.05745336040854454,
0.0994265079498291,
-0.022378098219633102,
0.041543152183294296,
-0.10350354760885239,
-0.018506472930312157,
-0.0019312659278512,
-0.09803473204374313,
0.03871661052107811,
0.05367949604988098,
-0.042881425470113754,
0.01283208467066288,
0.020614884793758392,
-0.014833034947514534,
-1.1236118872678304e-33,
0.035171058028936386,
0.09077897667884827,
0.03139590844511986,
0.07090108841657639,
-0.017701242119073868,
-0.013657554052770138,
0.012935125268995762,
0.08107338100671768,
-0.02938971109688282,
-0.04455573856830597,
-0.014068946242332458,
0.06639121472835541,
-0.04111367091536522,
0.07645399123430252,
-0.004808145109564066,
-0.02263786457479,
-0.030628446489572525,
0.021814074367284775,
-0.005131636280566454,
0.02586277201771736,
0.04557114467024803,
0.03177698701620102,
0.04363936558365822,
-0.06938211619853973,
-0.012731502763926983,
0.037852056324481964,
0.09406544268131256,
-0.04307270050048828,
-0.08786582946777344,
0.02790961042046547,
-0.1034570261836052,
-0.05032976716756821,
-0.007290449924767017,
-0.013347016647458076,
-0.011172585189342499,
-0.10940554738044739,
-0.032580774277448654,
-0.06218568608164787,
-0.017554931342601776,
-0.09798861294984818,
-0.03585052490234375,
0.04499683901667595,
0.05443260073661804,
-0.02340010367333889,
-0.04091367498040199,
-0.024286380037665367,
0.07452399283647537,
-0.047674089670181274,
0.02800234593451023,
-0.02745075710117817,
0.07706871628761292,
0.030239511281251907,
-0.049950167536735535,
-0.007924024946987629,
-0.04532073810696602,
0.02875388413667679,
0.0128706656396389,
0.020120894536376,
-0.020143333822488785,
0.037592802196741104,
-0.020286627113819122,
0.011164190247654915,
0.05104624480009079,
0.0675434023141861,
0.08572695404291153,
0.008397777564823627,
0.020815525203943253,
0.035877861082553864,
0.0293062012642622,
-0.008799147792160511,
-0.03263888508081436,
-0.029422415420413017,
-0.013264182023704052,
0.058502160012722015,
0.006170670501887798,
0.003075983840972185,
0.05450372025370598,
-0.13779409229755402,
0.08904608339071274,
-0.02085949294269085,
0.01755530759692192,
-0.01597512699663639,
0.0522758848965168,
-0.08188111335039139,
-0.10555578768253326,
-0.04298174008727074,
0.12646712362766266,
-0.10124950855970383,
0.004962073173373938,
0.031169619411230087,
-0.014003313146531582,
0.061052002012729645,
0.012144936248660088,
-0.0003611969586927444,
0.05243167653679848,
7.841275690263561e-35,
0.03851260989904404,
-0.09977108240127563,
-0.01726590096950531,
0.04050203040242195,
-0.019459189847111702,
-0.07048717886209488,
0.07616271823644638,
0.10973764955997467,
-0.04304015636444092,
-0.0591769777238369,
0.031019214540719986,
-0.020952589809894562,
0.0815420150756836,
-0.024593578651547432,
0.03749296814203262,
-0.04723295941948891,
0.042503293603658676,
0.008028514683246613,
0.047249630093574524,
0.04664028435945511,
-0.005401514936238527,
0.05365248769521713,
-0.02026580274105072,
0.018962068483233452,
-0.008334135636687279,
-0.007227006834000349,
-0.023623811081051826,
0.03998039290308952,
-0.06105552241206169,
0.03290736675262451,
0.027326002717018127,
0.006849133409559727,
-0.05749119445681572,
0.024872086942195892,
-0.08557116240262985,
0.04862942546606064,
0.010247036814689636,
-0.03295108303427696,
-0.0752028226852417,
0.10698192566633224,
0.05357182025909424,
0.012370352633297443,
-0.06354547291994095,
-0.06692936271429062,
0.013619782403111458,
0.0003870024811476469,
-0.08199792355298996,
-0.0009520053863525391,
0.01800473779439926,
-0.06732789427042007,
-0.004473973996937275,
0.05661192163825035,
-0.04033375158905983,
-0.02982908859848976,
0.03674447163939476,
-0.030627811327576637,
-0.028174884617328644,
-0.08600127696990967,
0.0449075773358345,
-0.005624644458293915,
-0.022771451622247696,
0.0021016178652644157,
0.08543530851602554,
-0.08067928999662399,
-0.012892120517790318,
-0.03264481946825981,
0.01173892617225647,
0.028362857177853584,
-0.022671382874250412,
0.017275488004088402,
0.04756598174571991,
-0.08321387320756912,
-0.08051051944494247,
0.02064075879752636,
0.0008837413624860346,
-0.025034818798303604,
-0.008717031218111515,
0.02907978557050228,
-0.02360835298895836,
-0.10404054820537567,
0.03270786255598068,
0.022846411913633347,
-0.0572967529296875,
0.0259986724704504,
0.04001312330365181,
0.06441596150398254,
0.06985482573509216,
0.052399635314941406,
-0.012989519163966179,
0.04297993704676628,
-0.02755177579820156,
0.023313818499445915,
0.003695179708302021,
0.024444574490189552,
-0.11497917771339417,
-3.423049932393951e-8,
-0.15860258042812347,
0.03364337608218193,
0.06414739787578583,
0.07888011634349823,
-0.07488124817609787,
-0.08994273096323013,
-0.06833362579345703,
0.027875885367393494,
-0.0015186738455668092,
-0.029924949631094933,
0.05080275610089302,
0.04067431762814522,
-0.03797696530818939,
0.026725661009550095,
-0.05974365770816803,
0.09655705094337463,
0.01885709911584854,
0.03357454761862755,
0.004846854601055384,
-0.010005950927734375,
0.027721773833036423,
0.04619721323251724,
0.017778214067220688,
0.000010646059308783151,
-0.004643538501113653,
-0.014282599091529846,
-0.033091288059949875,
0.0837743729352951,
0.11083575338125229,
0.006258904002606869,
-0.004740550648421049,
-0.03291340917348862,
0.007605351507663727,
-0.08972220122814178,
0.013882279396057129,
0.08063282072544098,
0.023468555882573128,
-0.028848394751548767,
0.015014449134469032,
-0.01253241952508688,
0.04675283655524254,
-0.008995003998279572,
-0.026328908279538155,
0.025239378213882446,
0.025043338537216187,
-0.002770504681393504,
-0.07254579663276672,
-0.09765828400850296,
0.0298995990306139,
-0.07114597409963608,
0.0838087648153305,
-0.04838228225708008,
0.02151799574494362,
0.0728389173746109,
0.10484147071838379,
0.015652760863304138,
-0.006904184352606535,
-0.0846560001373291,
0.09070229530334473,
-0.006116465665400028,
0.0635017603635788,
-0.0463627353310585,
-0.03725424408912659,
0.05527694895863533
] | 0 | 0 | 0 | 0 |
|
Execution-guided within-prompt search for programming-by-example | https://openreview.net/forum?id=PY56Wur7S0 | [
"Gust Verbruggen",
"Ashish Tiwari",
"Mukul Singh",
"Vu Le",
"Sumit Gulwani"
] | Poster | Large language models (LLMs) can generate code from examples without being limited to a DSL, but they lack search, as sampled programs are independent.
In this paper, we use an LLM as a policy that generates lines of code and then join these lines of code to let the LLM implicitly estimate the value of each of these lines in its next iteration.
We further guide the policy and value estimation by executing each line and annotating it with its results on the given examples.
This allows us to search for programs within a single (expanding) prompt until a sound program is found, by letting the policy reason in both the syntactic (code) and semantic (execution) space.
We evaluate within-prompt search on straight-line Python code generation using five benchmarks across different domains (strings, lists, and arbitrary Python programming problems).
We show that the model uses the execution results to guide the search and that within-prompt search performs well at low token budgets.
We also analyze how the model behaves as a policy and value, show that it can parallelize the search, and that it can implicitly backtrack over earlier generations. | programming-by-example, program synthesis, large language models | We propose to sample, combine and execute different lines of code from large language models to perform an execution-guided search for correct programs within one prompt. | 12,254 | null | [
-0.07219690829515457,
-0.006972982548177242,
-0.030696243047714233,
0.020033175125718117,
0.03567195311188698,
-0.08463210612535477,
-0.0024613409768790007,
0.037726689130067825,
-0.028358610346913338,
-0.039021480828523636,
-0.06996310502290726,
-0.05033530294895172,
0.10667393356561661,
-0.05793808400630951,
0.08084287494421005,
-0.03911992162466049,
0.06908253580331802,
-0.005995296407490969,
-0.046175774186849594,
-0.0921534076333046,
0.027834435924887657,
0.05107054114341736,
0.01564604975283146,
0.023255182430148125,
-0.015743419528007507,
-0.025060588493943214,
0.007047145161777735,
-0.050860099494457245,
0.037685636430978775,
0.06071306765079498,
-0.004736547823995352,
0.05055724456906319,
0.007471146993339062,
0.08074413985013962,
0.04372067376971245,
0.0463327057659626,
-0.08460596203804016,
-0.005658974405378103,
-0.06427407264709473,
-0.057383161038160324,
-0.09343258291482925,
0.04847089573740959,
-0.019004739820957184,
0.014630977995693684,
0.036324694752693176,
-0.05629028007388115,
-0.0003611920983530581,
-0.038980670273303986,
-0.034124672412872314,
-0.061027899384498596,
-0.0723147839307785,
-0.028440862894058228,
-0.0377749502658844,
-0.06368429958820343,
-0.03376147896051407,
0.006533300969749689,
0.14713388681411743,
0.0025864560157060623,
-0.0013357560383155942,
-0.06887367367744446,
-0.07751628756523132,
-0.1010689064860344,
-0.056147657334804535,
0.0003657841880340129,
-0.012288580648601055,
-0.10091828554868698,
0.048854995518922806,
0.025004465132951736,
0.06732740998268127,
0.0006583569338545203,
-0.03483733534812927,
0.06612831354141235,
-0.025376297533512115,
0.05436298996210098,
0.00918632373213768,
-0.003345346776768565,
0.01157344225794077,
0.026445159688591957,
0.014003921300172806,
-0.1070329025387764,
0.009954088367521763,
-0.029096923768520355,
-0.029641959816217422,
0.1101694256067276,
0.054827798157930374,
-0.03510122746229172,
0.056022677570581436,
0.01997157745063305,
0.12014836817979813,
0.06367287039756775,
-0.01168439257889986,
-0.0969298854470253,
-0.00358976935967803,
0.010797681286931038,
0.0384015291929245,
0.07243822515010834,
0.01897406578063965,
-0.12587596476078033,
-0.04197028651833534,
0.05597686767578125,
0.006006117910146713,
0.040896542370319366,
0.07605128735303879,
-0.10202500969171524,
0.00028874215786345303,
0.004371402319520712,
0.04505928233265877,
-0.005285195540636778,
0.032332733273506165,
-0.09588588029146194,
0.014862295240163803,
0.03995935618877411,
0.11665375530719757,
-0.02656041644513607,
0.05618699640035629,
-0.007288539316505194,
0.013773394748568535,
-0.036041487008333206,
0.060823868960142136,
0.051403552293777466,
-0.008079816587269306,
0.02903788909316063,
0.03130980581045151,
0.02920331247150898,
0.03929580748081207,
-0.008778885006904602,
-0.06252792477607727,
-2.663025330109282e-33,
0.03945443779230118,
-0.006407877430319786,
0.002486326964572072,
0.012335517443716526,
0.03397957980632782,
-0.042680200189352036,
0.1082044318318367,
0.050197046250104904,
-0.0644301101565361,
0.0019358727149665356,
0.02652072347700596,
-0.00782942958176136,
-0.023298105224967003,
0.038114868104457855,
0.06770931929349899,
0.010906816460192204,
-0.03485044091939926,
0.08641573041677475,
-0.0595351941883564,
0.03162804991006851,
0.03560630604624748,
-0.010503407567739487,
0.014607604593038559,
-0.047993648797273636,
0.09122385084629059,
0.007220536936074495,
-0.019048986956477165,
-0.05813697725534439,
-0.013343523256480694,
0.03389495238661766,
-0.09436026960611343,
-0.06167348474264145,
0.012857392430305481,
0.04062432050704956,
0.02999971993267536,
0.03805866464972496,
-0.01952369138598442,
-0.016573591157794,
0.014109023846685886,
-0.020355496555566788,
0.00782350916415453,
-0.002158517250791192,
0.05677880719304085,
0.021608971059322357,
-0.0660509541630745,
-0.08315584808588028,
-0.1277119666337967,
0.016970982775092125,
-0.02844863012433052,
0.00882203783839941,
0.0991516038775444,
0.08130331337451935,
-0.01361674815416336,
0.02059161104261875,
-0.018938696011900902,
0.01883612386882305,
-0.011985178105533123,
-0.004996545147150755,
0.0018375901272520423,
0.03174596652388573,
0.0306685883551836,
0.018081510439515114,
0.017270434647798538,
0.05769757926464081,
0.00034770931233651936,
0.05254032835364342,
-0.004796520806849003,
0.07785864174365997,
0.08667807281017303,
0.06981261074542999,
-0.056853245943784714,
-0.021277189254760742,
-0.024992801249027252,
-0.03924567252397537,
0.06396004557609558,
-0.019889619201421738,
0.03999492898583412,
-0.12300819903612137,
0.06564453989267349,
-0.04866907373070717,
0.02461906336247921,
-0.01572609692811966,
-0.0028511995915323496,
0.023249834775924683,
0.03785283863544464,
-0.07254695147275925,
0.03427090123295784,
-0.04701923206448555,
-0.05285583809018135,
-0.04041581600904465,
-0.09808006137609482,
-0.05637557804584503,
-0.02025744318962097,
0.005642548203468323,
-0.009302458725869656,
-4.175956079088158e-34,
0.04391477629542351,
-0.07887476682662964,
0.006323505658656359,
0.046223364770412445,
-0.06545213609933853,
-0.09536400437355042,
-0.027269721031188965,
-0.07846391946077347,
-0.0056473021395504475,
-0.045471370220184326,
-0.11063705384731293,
-0.02328231744468212,
0.07678785920143127,
0.03677374869585037,
0.03064255602657795,
0.04007016867399216,
-0.0122681250795722,
0.036431871354579926,
0.0770249292254448,
0.08033545315265656,
-0.03794889152050018,
0.08645658940076828,
-0.08406033366918564,
0.022554706782102585,
-0.08279313147068024,
-0.03596509248018265,
-0.07197314500808716,
0.02080288529396057,
0.006508616264909506,
0.0015826555900275707,
-0.0627126470208168,
0.027075501158833504,
-0.048399921506643295,
-0.013716244138777256,
-0.07252281159162521,
0.04659420624375343,
0.09398305416107178,
0.0006817098474130034,
-0.06896159797906876,
0.08318907767534256,
0.09198036044836044,
-0.04869341850280762,
0.023175746202468872,
0.01958894543349743,
-0.029461568221449852,
0.001236643991433084,
-0.11004509031772614,
0.04769297316670418,
-0.005718512926250696,
0.001226503518410027,
-0.00630187289789319,
0.012061377987265587,
0.0293430183082819,
0.0010961951920762658,
-0.07825162261724472,
-0.0993531346321106,
-0.02838188037276268,
0.014831607230007648,
-0.034087467938661575,
-0.0032207975164055824,
-0.07304258644580841,
-0.030496541410684586,
0.03692952170968056,
-0.03394318372011185,
0.03282487764954567,
0.025450753048062325,
-0.05387946218252182,
-0.06974770128726959,
-0.0032173669897019863,
-0.06500263512134552,
0.08117940276861191,
-0.05900561809539795,
-0.09009958803653717,
0.049134448170661926,
-0.04563258960843086,
0.0641966164112091,
0.015793558210134506,
-0.08778699487447739,
-0.007619162555783987,
-0.013431998901069164,
0.009295899420976639,
0.03970232605934143,
0.05314590036869049,
0.03618283569812775,
-0.09968575835227966,
0.022828608751296997,
0.017992230132222176,
0.10470358282327652,
-0.02850055880844593,
0.03238019347190857,
-0.03229843080043793,
0.0022473582066595554,
0.012244811281561852,
-0.023986028507351875,
0.009502594359219074,
-4.355259974886394e-8,
-0.018011150881648064,
0.048158373683691025,
0.03947790712118149,
0.10508674383163452,
0.07208317518234253,
0.036834005266427994,
-0.032556939870119095,
0.038544006645679474,
0.005164268892258406,
-0.07319558411836624,
0.03483925387263298,
0.030103204771876335,
-0.03770748898386955,
0.009049897082149982,
-0.05232670530676842,
0.1017269641160965,
-0.04162576422095299,
-0.050383664667606354,
-0.03278660774230957,
-0.01183246448636055,
0.02514849789440632,
0.0839654877781868,
-0.014966239221394062,
0.016304729506373405,
-0.03307616338133812,
-0.03258485347032547,
0.050694260746240616,
0.1026802584528923,
0.06832814961671829,
-0.005992010701447725,
0.044629648327827454,
-0.013275368139147758,
-0.0453682467341423,
0.0057116542011499405,
0.054370131343603134,
-0.0002637080615386367,
-0.0020812132861465216,
-0.04089326038956642,
0.005095026455819607,
-0.06591640412807465,
0.039002399891614914,
0.06526213139295578,
-0.014162914827466011,
-0.002021027496084571,
0.051368605345487595,
-0.028418904170393944,
-0.008161579258739948,
-0.0731915831565857,
0.020057454705238342,
-0.07844412326812744,
-0.015723824501037598,
-0.017371229827404022,
-0.052949532866477966,
0.002064979635179043,
0.010418967343866825,
0.07948987931013107,
-0.09926364570856094,
0.0027299572248011827,
0.00467605609446764,
0.06716392934322357,
-0.009664648212492466,
0.03167131915688515,
0.0335632748901844,
0.015415683388710022
] | 0 | 0 | 0 | 0 |
|
Lie Algebra Canonicalization: Equivariant Neural Operators under arbitrary Lie Groups | https://openreview.net/forum?id=7PLpiVdnUC | [
"Zakhar Shumaylov",
"Peter Zaika",
"James Rowbottom",
"Ferdia Sherry",
"Melanie Weber",
"Carola-Bibiane Schönlieb"
] | Poster | The quest for robust and generalizable machine learning models has driven recent interest in exploiting symmetries through equivariant neural networks. In the context of PDE solvers, recent works have shown that Lie point symmetries can be a useful inductive bias for Physics-Informed Neural Networks (PINNs) through data and loss augmentation. Despite this, directly enforcing equivariance within the model architecture for these problems remains elusive. This is because many PDEs admit non-compact symmetry groups, oftentimes not studied beyond their infinitesimal generators, making them incompatible with most existing equivariant architectures. In this work, we propose Lie aLgebrA Canonicalization (LieLAC), a novel approach that exploits only the action of infinitesimal generators of the symmetry group, circumventing the need for knowledge of the full group structure. To achieve this, we address existing theoretical issues in the canonicalization literature, establishing connections with frame averaging in the case of continuous non-compact groups. Operating within the framework of canonicalization, LieLAC can easily be integrated with unconstrained pre-trained models, transforming inputs to a canonical form before feeding them into the existing model, effectively aligning the input for model inference according to allowed symmetries. LieLAC utilizes standard Lie group descent schemes, achieving equivariance in pre-trained models. Finally, we showcase LieLAC's efficacy on tasks of invariant image classification and Lie point symmetry equivariant neural PDE solvers using pre-trained models. | Canonicalization, Equivariance, Invariance, Lie algebra, Partial Differential Equations, Neural Operator, PINN, Neural PDE solver, Lie point symmetries, Frames, Frame Averaging | null | 12,249 | 2410.02698 | [
-0.12867754697799683,
-0.07182660698890686,
0.08259967714548111,
0.06302393972873688,
-0.0111162718385458,
0.02410927414894104,
-0.058785296976566315,
-0.0765562653541565,
0.03519560024142265,
-0.07524257898330688,
0.06953415274620056,
0.03414672985672951,
-0.11516325175762177,
0.024865014478564262,
-0.006542214658111334,
-0.04198291525244713,
-0.04658360406756401,
0.16294120252132416,
-0.11953409761190414,
-0.06577009707689285,
-0.006930705159902573,
-0.008357680402696133,
0.009978457354009151,
0.06532745063304901,
0.027934450656175613,
0.0012415811652317643,
0.023867402225732803,
-0.03449421003460884,
-0.008606265299022198,
-0.033341508358716965,
0.06317409873008728,
0.021145854145288467,
-0.08807624131441116,
-0.05531791225075722,
-0.07696143537759781,
0.06098579987883568,
0.012404941022396088,
-0.000042047486203955486,
-0.05045429989695549,
-0.0860949158668518,
0.010345458984375,
0.007779823616147041,
0.04827480390667915,
0.04772389307618141,
0.06911740452051163,
0.025653155520558357,
0.05846061930060387,
-0.08061811327934265,
-0.01934623531997204,
0.030143244192004204,
0.02272040955722332,
0.07734919339418411,
-0.07511471956968307,
-0.002223593881353736,
0.010584581643342972,
-0.07559941709041595,
-0.003929138649255037,
-0.06274355947971344,
0.03208693861961365,
-0.05073746666312218,
0.022504590451717377,
-0.009506442584097385,
-0.029901549220085144,
0.04243187606334686,
-0.04252438619732857,
0.04638748988509178,
0.0471293143928051,
0.04611346870660782,
0.0026988277677446604,
-0.0012574740685522556,
-0.02350275032222271,
0.004428690765053034,
-0.06659571826457977,
0.037901971489191055,
0.032457802444696426,
0.045961298048496246,
0.06298273056745529,
-0.0008399751386605203,
0.06806798279285431,
0.04610889405012131,
0.02327273227274418,
0.03228720277547836,
0.01997828111052513,
-0.015307501889765263,
0.038597770035266876,
-0.002093441318720579,
-0.05708109587430954,
0.04349550977349281,
0.03839166834950447,
-0.041696686297655106,
0.002882012166082859,
-0.016383327543735504,
0.002281820634379983,
-0.01852327026426792,
0.12408722937107086,
-0.01545388251543045,
-0.0021039671264588833,
-0.013213071972131729,
-0.039501115679740906,
0.10436180233955383,
0.04046853259205818,
-0.023185424506664276,
0.04199708253145218,
0.05514288321137428,
0.04614150524139404,
-0.03188825026154518,
0.038839150220155716,
-0.017036620527505875,
-0.0030171959660947323,
-0.1095089390873909,
-0.003471337491646409,
-0.03644672781229019,
0.022291965782642365,
-0.03911774232983589,
-0.02678246609866619,
-0.05644363537430763,
0.015086688101291656,
0.012862924486398697,
0.0631658062338829,
0.006704350933432579,
-0.0055761937983334064,
0.03378926217556,
-0.0023985435254871845,
0.08599651604890823,
0.057616930454969406,
0.009401031769812107,
-0.1099848821759224,
3.7138087301879484e-33,
0.018565043807029724,
0.04210658743977547,
0.08854798972606659,
0.006037916988134384,
0.047901030629873276,
0.0014287136727944016,
0.07275675237178802,
0.04173227399587631,
0.011772187426686287,
0.06642439216375351,
-0.14367350935935974,
-0.013862280175089836,
-0.008642596192657948,
0.04056154564023018,
-0.03523771092295647,
0.019689258188009262,
-0.026041820645332336,
0.016308769583702087,
0.037843652069568634,
-0.07318002730607986,
0.06554531306028366,
0.018332138657569885,
-0.008864396251738071,
-0.010580748319625854,
-0.08249688148498535,
0.04019801691174507,
-0.003949966747313738,
0.04729996249079704,
-0.08899403363466263,
0.008455231785774231,
-0.009085717611014843,
-0.030992208048701286,
-0.05258399248123169,
0.07377905398607254,
0.025044621899724007,
-0.015813887119293213,
0.025354238227009773,
-0.0521342009305954,
0.0281606987118721,
-0.015830952674150467,
0.06865733116865158,
0.004841633141040802,
-0.04618349298834801,
0.018919870257377625,
-0.06137522682547569,
-0.016885507851839066,
0.07621660828590393,
0.06928638368844986,
-0.04540566727519035,
-0.0688515156507492,
-0.009192798286676407,
0.01873987726867199,
-0.06576500087976456,
-0.0474916510283947,
0.018335388973355293,
0.05241379514336586,
0.02714608609676361,
0.06616826355457306,
-0.013828087598085403,
-0.013225100003182888,
-0.024288319051265717,
0.06739278137683868,
0.00021522041060961783,
-0.048410747200250626,
-0.04165434464812279,
0.026844175532460213,
-0.025699449703097343,
0.03189603611826897,
0.06407766044139862,
-0.021859992295503616,
-0.024278415367007256,
-0.006390645168721676,
-0.11195818334817886,
0.00694094505161047,
0.11152958124876022,
-0.018393486738204956,
0.03190179914236069,
-0.04552155360579491,
-0.035971708595752716,
0.0440269336104393,
-0.04273390769958496,
0.05400506779551506,
-0.07208064943552017,
0.03639046847820282,
-0.05424273759126663,
-0.061145782470703125,
0.01220026146620512,
-0.0899093821644783,
-0.014958027750253677,
-0.12509550154209137,
-0.10891065746545792,
-0.06730896234512329,
-0.024054987356066704,
0.031016796827316284,
0.022043636068701744,
-4.8702622765788226e-33,
-0.011973069980740547,
-0.049974873661994934,
-0.0044754259288311005,
0.053862977772951126,
0.05113222822546959,
-0.09261281043291092,
-0.009898536838591099,
0.05755482241511345,
0.021949591115117073,
-0.031203582882881165,
-0.008301100693643093,
-0.025418080389499664,
-0.030449317768216133,
-0.037959933280944824,
0.06149594485759735,
0.04343637451529503,
-0.03982207551598549,
-0.0677357167005539,
0.027237361297011375,
0.06010930985212326,
-0.01014371681958437,
0.06151757016777992,
-0.02996147610247135,
0.004738073796033859,
0.014241954311728477,
0.01834682561457157,
-0.036665063351392746,
0.025561131536960602,
0.08132974058389664,
0.008585590869188309,
-0.01260257326066494,
-0.06777932494878769,
-0.07902288436889648,
0.034069448709487915,
-0.08353761583566666,
0.0789390429854393,
0.03085371106863022,
0.028522389009594917,
-0.05734586715698242,
0.08036676794290543,
0.020019516348838806,
-0.013674367219209671,
-0.047775574028491974,
0.11355316638946533,
0.07367498427629471,
-0.024276455864310265,
0.07050152122974396,
-0.002843110589310527,
-0.07657536119222641,
0.07638022303581238,
-0.03536802530288696,
0.01137306448072195,
-0.017431039363145828,
-0.0026324368081986904,
-0.08753999322652817,
0.026628516614437103,
0.04499218612909317,
-0.10232359170913696,
0.04066801443696022,
0.006823311559855938,
-0.00042571843368932605,
-0.12056224048137665,
0.0532274954020977,
0.015329044312238693,
-0.019981537014245987,
-0.013609876856207848,
-0.015538596548140049,
0.035229116678237915,
0.048165213316679,
0.008687339723110199,
0.0007908334373496473,
-0.10943464189767838,
-0.05696617811918259,
-0.014184083789587021,
0.043086159974336624,
-0.04246271401643753,
0.008653492666780949,
-0.06670767068862915,
0.03103717789053917,
-0.1072835773229599,
-0.014156865887343884,
-0.05686678737401962,
0.10575803369283676,
0.05570078641176224,
0.10018681734800339,
0.039443232119083405,
0.07515385746955872,
0.024138078093528748,
0.04863101616501808,
-0.009718579240143299,
-0.005883022677153349,
0.029328707605600357,
0.05200142040848732,
0.11435770243406296,
-0.03872143104672432,
-4.630126682059199e-8,
-0.08233137428760529,
0.055026087909936905,
-0.03112061880528927,
0.017272021621465683,
0.001042015734128654,
-0.039741694927215576,
-0.044261425733566284,
0.021043695509433746,
-0.010320954956114292,
0.02066982164978981,
0.051062535494565964,
-0.03354259580373764,
-0.06791257113218307,
-0.08653344959020615,
-0.026400504633784294,
0.09839152544736862,
-0.019609052687883377,
0.05034605786204338,
-0.003705934388563037,
0.014507206156849861,
0.041072387248277664,
0.03189060837030411,
-0.013371058739721775,
-0.060142241418361664,
0.06610553711652756,
-0.08310616761445999,
-0.052734799683094025,
0.02995111607015133,
0.07003592699766159,
0.07004422694444656,
-0.04997910186648369,
-0.004483033902943134,
0.022959934547543526,
-0.01746477745473385,
0.020389163866639137,
0.03207775950431824,
0.005677449516952038,
-0.06492924690246582,
-0.034596044570207596,
-0.06525023281574249,
-0.05068517103791237,
0.09445029497146606,
-0.010011326521635056,
0.01864030584692955,
0.01823022961616516,
0.021285712718963623,
0.04081573709845543,
-0.03182126209139824,
0.026499876752495766,
0.055994078516960144,
0.05957697331905365,
0.00875051785260439,
-0.07823749631643295,
0.03133699297904968,
0.036655016243457794,
0.032111115753650665,
0.03079123981297016,
-0.02789587527513504,
0.054304350167512894,
-0.0067178551107645035,
-0.07146625965833664,
0.031329382210969925,
-0.08171164989471436,
-0.048130400478839874
] | 0 | 0 | 0 | 0 |
|
Extendable and Iterative Structure Learning Strategy for Bayesian Networks | https://openreview.net/forum?id=3n6DYH3cIP | [
"Hamid Kalantari",
"Russell Greiner",
"Pouria Ramazi"
] | Poster | Learning the structure of Bayesian networks is a fundamental yet computationally intensive task, especially as the number of variables grows. Traditional algorithms require retraining from scratch when new variables are introduced, making them impractical for dynamic or large-scale applications. In this paper, we propose an extendable structure learning strategy that efficiently incorporates a new variable $Y$ into an existing Bayesian network graph $\mathcal{G}$ over variables $\mathcal{X}$, resulting in an updated P-map graph $\bar{\mathcal{G}}$ on $\bar{\mathcal{X}} = \mathcal{X} \cup \{Y\}$. By leveraging the information encoded in $\mathcal{G}$, our method significantly reduces computational overhead compared to learning $\bar{\mathcal{G}}$ from scratch. Empirical evaluations demonstrate runtime reductions of up to 1300x without compromising accuracy. Building on this approach, we introduce a novel iterative paradigm for structure learning over $\mathcal{X}$. Starting with a small subset $\mathcal{U} \subset \mathcal{X}$, we iteratively add the remaining variables using our extendable algorithms to construct a P-map graph over the full set. This method offers runtime advantages comparable to common algorithms while maintaining similar accuracy. Our contributions provide a scalable solution for Bayesian network structure learning, enabling efficient model updates in real-time and high-dimensional settings. | Structure learning, Bayesian networks, Causal discovery | null | 12,243 | null | [
-0.03093155100941658,
-0.06724250316619873,
0.057104334235191345,
0.01005682721734047,
0.027359193190932274,
-0.052453067153692245,
0.0062734149396419525,
-0.066559799015522,
-0.035164669156074524,
-0.023373622447252274,
0.011070678010582924,
0.024575287476181984,
0.03474154695868492,
0.06297316402196884,
0.0007213049684651196,
0.06402454525232315,
0.06994950026273727,
0.09704305231571198,
-0.013819033280014992,
-0.05038965865969658,
-0.02043974958360195,
0.006149558816105127,
-0.00015636523312423378,
-0.019880758598446846,
0.057380836457014084,
0.024730145931243896,
0.06836570054292679,
-0.0032957419753074646,
0.012392709963023663,
-0.022893372923135757,
0.014848305843770504,
-0.027635866776108742,
0.008897689171135426,
-0.059272829443216324,
-0.006072575226426125,
0.054565977305173874,
-0.022537730634212494,
0.004743698984384537,
0.02210080623626709,
0.007235750090330839,
0.03420529142022133,
0.03635875880718231,
0.09653203189373016,
0.09259403496980667,
0.03332488238811493,
0.09871619939804077,
-0.03092029318213463,
-0.0030766369309276342,
0.05903425067663193,
0.007246060762554407,
-0.056455519050359726,
-0.054803431034088135,
-0.050033070147037506,
0.00867468398064375,
0.10710165649652481,
0.011912205256521702,
-0.02561667189002037,
-0.010701708495616913,
-0.07596296817064285,
-0.022093242034316063,
-0.040864378213882446,
-0.08732213824987411,
-0.02998085506260395,
-0.0193658284842968,
-0.010502704419195652,
0.06710217148065567,
0.0015050204237923026,
0.10482374578714371,
0.0025637128856033087,
0.10257390141487122,
0.05725198611617088,
0.08035396784543991,
-0.06626325845718384,
0.08762693405151367,
0.022824948653578758,
0.039392728358507156,
0.03589826822280884,
-0.03975682333111763,
0.014709049835801125,
-0.01835317350924015,
-0.0062043434008955956,
0.005416652653366327,
-0.014067926444113255,
0.0022528485860675573,
0.07840943336486816,
-0.06687593460083008,
-0.03000899776816368,
0.008312095887959003,
0.023307695984840393,
0.09494763612747192,
-0.041639141738414764,
-0.031110970303416252,
0.018403256312012672,
-0.06117574870586395,
-0.028670815750956535,
0.012302949093282223,
0.032941944897174835,
-0.05926777049899101,
0.0056230961345136166,
0.1232062429189682,
-0.007395112421363592,
0.05938521400094032,
0.07585105299949646,
-0.06298375874757767,
-0.06591605395078659,
0.009736216627061367,
-0.008715933188796043,
0.08807145059108734,
-0.005944991484284401,
-0.1110980361700058,
-0.04346035048365593,
0.039567798376083374,
-0.05829041451215744,
-0.053480640053749084,
-0.07288769632577896,
-0.02762998268008232,
0.033500730991363525,
-0.013573357835412025,
-0.01991192065179348,
0.05444563180208206,
0.0200868658721447,
0.030935313552618027,
-0.015551803633570671,
0.012546466663479805,
-0.07513223588466644,
-0.047817718237638474,
-0.10510759800672531,
1.485022492709439e-33,
0.08158163726329803,
0.07970452308654785,
0.06803032010793686,
0.040177151560783386,
0.0408475287258625,
0.07514651119709015,
0.01868135668337345,
0.0010813180124387145,
-0.07173722237348557,
0.022407667711377144,
-0.030827118083834648,
0.026682617142796516,
-0.04148104041814804,
0.04863862693309784,
0.08024408668279648,
-0.06351281702518463,
0.012591036036610603,
0.07235228270292282,
-0.015258136205375195,
-0.037259168922901154,
0.07514575868844986,
-0.02953162044286728,
0.012853260152041912,
-0.006451643072068691,
0.02859671786427498,
-0.018770745024085045,
0.039794180542230606,
0.032388605177402496,
-0.0840887501835823,
-0.002894858829677105,
-0.03532092273235321,
0.039014365524053574,
-0.004380952101200819,
0.11111894994974136,
0.0025682139676064253,
-0.08462255448102951,
0.008239864371716976,
-0.07288584113121033,
0.01586640253663063,
-0.08420190215110779,
0.0022870649117976427,
-0.00412231357768178,
0.05040580779314041,
0.015333062037825584,
-0.074496328830719,
-0.11202141642570496,
-0.04296698793768883,
-0.05174117535352707,
-0.027929138392210007,
-0.03655746206641197,
0.11756544560194016,
-0.07606703788042068,
-0.030908137559890747,
0.008053116500377655,
-0.06116247922182083,
0.030822020024061203,
0.033874548971652985,
0.04155244678258896,
-0.004592417739331722,
0.02774539589881897,
0.058382339775562286,
0.06383619457483292,
0.002056404948234558,
0.06962749361991882,
-0.002572210505604744,
0.04956435412168503,
-0.02970331907272339,
-0.0018179093021899462,
0.0615294985473156,
-0.04498185217380524,
-0.024608876556158066,
0.04076667129993439,
-0.04526473581790924,
-0.05928380787372589,
0.10023634135723114,
0.03072410076856613,
-0.003949908073991537,
-0.04401177912950516,
-0.06898390501737595,
-0.018989622592926025,
-0.0688902884721756,
-0.010833299718797207,
0.03177305683493614,
0.06301465630531311,
0.005018946714699268,
-0.0023592545185238123,
0.07490938901901245,
-0.014407595619559288,
-0.012895412743091583,
0.01951265148818493,
-0.11636241525411606,
-0.08326786756515503,
0.007274852599948645,
-0.030337821692228317,
0.01394033245742321,
-1.4950106131070783e-33,
-0.05353372171521187,
-0.02227363921701908,
0.07091325521469116,
0.03941401466727257,
-0.03059009276330471,
-0.1277877688407898,
0.02301144413650036,
-0.0402265302836895,
-0.055041152983903885,
-0.06006299704313278,
-0.09747496992349625,
-0.020457791164517403,
-0.0108109749853611,
-0.04362819716334343,
0.05731239542365074,
0.05585922673344612,
-0.04671740531921387,
-0.012839912436902523,
-0.040840089321136475,
-0.043845679610967636,
-0.029923446476459503,
0.04970938339829445,
-0.012894167564809322,
-0.03471602499485016,
0.02267700619995594,
0.017854388803243637,
0.0505896657705307,
0.10298339277505875,
0.04347793012857437,
0.022866494953632355,
-0.13046737015247345,
-0.0166343804448843,
-0.08004223555326462,
-0.0013652374036610126,
-0.03974008932709694,
0.11720462143421173,
0.036557093262672424,
0.05092300474643707,
-0.008184694685041904,
-0.012010940350592136,
0.002064213389530778,
-0.01794344186782837,
-0.12175916135311127,
-0.014055481180548668,
0.051695942878723145,
0.02297748439013958,
0.008310130797326565,
0.033724017441272736,
-0.06708363443613052,
0.02261982671916485,
-0.0024428635369986296,
0.06320525705814362,
-0.026718303561210632,
0.0038438981864601374,
-0.045881472527980804,
-0.052327439188957214,
-0.04419158399105072,
-0.007962099276483059,
0.0745355486869812,
-0.029167309403419495,
-0.04315461218357086,
-0.09713909029960632,
0.01032155193388462,
0.08081170171499252,
-0.02324572019279003,
-0.05306433513760567,
-0.007898720912635326,
-0.012717747129499912,
-0.018879743292927742,
0.047958843410015106,
0.017791152000427246,
0.022521665319800377,
-0.06770816445350647,
0.0018449625931680202,
0.002415916882455349,
-0.000461509102024138,
0.007322533521801233,
0.04835187643766403,
-0.004728951025754213,
0.013588531874120235,
0.04573822021484375,
-0.006997322663664818,
-0.02284875512123108,
0.0005047396989539266,
0.12715767323970795,
0.03449203073978424,
0.05343012884259224,
0.07374896109104156,
0.0986146405339241,
0.013579864986240864,
-0.06029461324214935,
0.08414552360773087,
-0.06733737885951996,
0.01978154294192791,
-0.08441982418298721,
-4.108489193299647e-8,
-0.14826641976833344,
0.05973505228757858,
0.025816785171628,
-0.005347750149667263,
0.14165759086608887,
0.01137144397944212,
0.04721466451883316,
0.143821120262146,
0.02843673713505268,
-0.037553247064352036,
0.04563703015446663,
-0.03370986506342888,
0.0327913723886013,
-0.0512973889708519,
-0.037549108266830444,
0.00941433198750019,
0.008108843117952347,
-0.004303036257624626,
-0.027651870623230934,
0.03545418381690979,
-0.0009664976969361305,
0.024088915437459946,
-0.025712013244628906,
0.006251705810427666,
0.039480794221162796,
-0.15197256207466125,
-0.000848849886097014,
0.01776937022805214,
0.018110385164618492,
-0.010948811657726765,
-0.05632173642516136,
0.014905384741723537,
-0.015951961278915405,
0.06096600368618965,
0.01110770646482706,
0.06090489774942398,
0.0012524959165602922,
-0.03471091017127037,
-0.008386464789509773,
-0.06009693443775177,
-0.03520785644650459,
-0.02054835855960846,
-0.026434149593114853,
0.05611680448055267,
0.011721334420144558,
-0.0028272527270019054,
0.02294614166021347,
0.06986628472805023,
0.019245395436882973,
-0.0037107542157173157,
0.008522909134626389,
0.01443337369710207,
-0.06029592826962471,
-0.0604943186044693,
-0.02108047343790531,
0.03430274873971939,
0.07001237571239471,
-0.002512611448764801,
-0.005001281388103962,
0.04956028610467911,
-0.09292685240507126,
-0.060493264347314835,
-0.061740126460790634,
-0.022630605846643448
] | 0 | 0 | 0 | 0 |
|
Improving Graph Neural Networks by Learning Continuous Edge Directions | https://openreview.net/forum?id=iAmR7FfMmq | [
"Seong Ho Pahng",
"Sahand Hormoz"
] | Poster | Graph Neural Networks (GNNs) traditionally employ a message-passing mechanism that resembles diffusion over undirected graphs, which often leads to homogenization of node features and reduced discriminative power in tasks such as node classification. Our key insight for addressing this limitation is to assign fuzzy edge directions---that can vary continuously from node $i$ pointing to node $j$ to vice versa---to the edges of a graph so that features can preferentially flow in one direction between nodes to enable long-range information transmission across the graph. We also introduce a novel complex-valued Laplacian for directed graphs with fuzzy edges where the real and imaginary parts represent information flow in opposite directions. Using this Laplacian, we propose a general framework, called Continuous Edge Direction (CoED) GNN, for learning on graphs with fuzzy edges and prove its expressivity limits using a generalization of the Weisfeiler-Leman (WL) graph isomorphism test for directed graphs with fuzzy edges. Our architecture aggregates neighbor features scaled by the learned edge directions and processes the aggregated messages from in-neighbors and out-neighbors separately alongside the self-features of the nodes. Since continuous edge directions are differentiable, they can be learned jointly with the GNN weights via gradient-based optimization. CoED GNN is particularly well-suited for graph ensemble data where the graph structure remains fixed but multiple realizations of node features are available, such as in gene regulatory networks, web connectivity graphs, and power grids. We demonstrate through extensive experiments on both synthetic and real graph ensemble datasets that learning continuous edge directions significantly improves performance both for undirected and directed graphs compared with existing methods. | Graph Neural Networks, Directed Graphs, Graph Laplacian, Continuous Edge Directions, Graph Ensemble Data | We introduce a novel graph Laplacian and graph neural network (GNN) framework that allows learning of continuous edge directions alongside the GNN parameters, leading to improved performance on both undirected and directed graphs. | 12,236 | 2410.14109 | [
-0.02414386160671711,
-0.0500301755964756,
0.06867750734090805,
0.003204071195796132,
0.0029446007683873177,
-0.004231681581586599,
-0.004096657037734985,
-0.10557229816913605,
-0.03046579100191593,
0.007929452694952488,
0.055023808032274246,
-0.0011512221535667777,
0.002712781773880124,
0.08030910044908524,
-0.05438001826405525,
0.026501648128032684,
0.12424604594707489,
-0.009801853448152542,
-0.04853455722332001,
-0.016081470996141434,
0.0751166120171547,
-0.011371754109859467,
-0.020260103046894073,
0.043509580194950104,
0.04303942620754242,
0.01868159882724285,
-0.014954914338886738,
-0.003802980063483119,
0.063869409263134,
0.004768322221934795,
0.0944845974445343,
-0.022465072572231293,
-0.045483771711587906,
0.0352371521294117,
-0.06457331776618958,
0.046009138226509094,
-0.013364736922085285,
0.07900958508253098,
0.07168297469615936,
0.026152612641453743,
0.0014847519341856241,
-0.027099836617708206,
0.025900237262248993,
0.09432652592658997,
0.09017260372638702,
0.006380855571478605,
-0.06480953842401505,
-0.023782536387443542,
-0.014311416074633598,
0.08921307325363159,
-0.037109654396772385,
-0.006580309942364693,
-0.0527348592877388,
0.07668957859277725,
0.03464389964938164,
0.0009296220378018916,
0.033304039388895035,
0.050774239003658295,
-0.0011003153631463647,
0.0939522460103035,
0.046742990612983704,
-0.1631937026977539,
0.06206362321972847,
-0.05756150931119919,
-0.09073595702648163,
0.01418155338615179,
0.010432541370391846,
0.0767575353384018,
-0.044360484927892685,
0.04028557613492012,
0.14664509892463684,
-0.008233658969402313,
-0.1311296969652176,
0.03099948726594448,
0.06807063519954681,
0.08808226138353348,
0.07294617593288422,
0.02752455323934555,
-0.023906780406832695,
0.018291661515831947,
-0.02780737727880478,
0.005183658096939325,
0.0550282746553421,
-0.0050600147806108,
0.04932541027665138,
0.010728247463703156,
-0.0057976567186415195,
0.019663572311401367,
-0.01023500133305788,
0.0008918028324842453,
-0.029934430494904518,
0.0038118474185466766,
-0.03913269191980362,
-0.04498012736439705,
0.09816794097423553,
-0.026667684316635132,
-0.0015378607204183936,
0.011741205118596554,
-0.04992922022938728,
0.09221900999546051,
0.006229875143617392,
-0.037090666592121124,
-0.06289107352495193,
-0.10433751344680786,
0.08581703901290894,
0.0787271186709404,
0.009863550774753094,
0.043273650109767914,
0.020973213016986847,
-0.1267901062965393,
-0.06900564581155777,
0.017671210691332817,
0.00012076943676220253,
-0.0024151871912181377,
-0.015195859596133232,
-0.09223291277885437,
0.08259724080562592,
0.012479388155043125,
0.05862222611904144,
0.06950071454048157,
-0.005184543784707785,
0.026249952614307404,
-0.09267742931842804,
0.004977635573595762,
0.019244620576500893,
-0.0025507761165499687,
-0.07141489535570145,
2.475250760482976e-33,
0.04412105679512024,
-0.00753022963181138,
0.011701729148626328,
-0.03223913535475731,
0.07981747388839722,
-0.02869018167257309,
-0.055970337241888046,
0.005443866364657879,
0.07572608441114426,
-0.01690729893743992,
-0.05186847969889641,
0.003451639087870717,
-0.07651819288730621,
0.034182462841272354,
-0.02646661549806595,
-0.052896298468112946,
-0.016523711383342743,
-0.06970902532339096,
0.041158389300107956,
-0.14305295050144196,
0.05781535431742668,
-0.07110518962144852,
0.05191655457019806,
-0.0009216678445227444,
0.019162125885486603,
-0.01104290783405304,
0.025204643607139587,
-0.003299388103187084,
-0.006060045212507248,
-0.02352273464202881,
-0.0675475224852562,
-0.03965269774198532,
0.004164866171777248,
-0.02149718441069126,
0.024654505774378777,
0.03178039938211441,
-0.04290460795164108,
-0.04070967435836792,
0.05861194431781769,
-0.019751422107219696,
-0.04448504373431206,
0.0036741632502526045,
-0.03245002031326294,
0.04867926239967346,
-0.005592531058937311,
0.0040644751861691475,
-0.04962831735610962,
0.00029068515868857503,
-0.03950086608529091,
-0.05230780690908432,
-0.06727121770381927,
-0.016148244962096214,
0.04066164791584015,
-0.0051154582761228085,
0.007757760118693113,
-0.041831620037555695,
0.03793838247656822,
0.06400952488183975,
0.02930154651403427,
0.02711234800517559,
-0.006196458823978901,
0.009350113570690155,
-0.03209872171282768,
0.04424157366156578,
0.0013988373102620244,
-0.05749452859163284,
-0.12306435406208038,
0.0733935758471489,
0.0848771184682846,
-0.08404553681612015,
-0.012361976318061352,
0.06564731895923615,
0.02858378179371357,
-0.04342731833457947,
0.019444147124886513,
0.014606377109885216,
-0.049508195370435715,
-0.07114055752754211,
-0.0082272719591856,
-0.0278733279556036,
-0.08178417384624481,
-0.00989176519215107,
0.003318868111819029,
-0.05705234035849571,
-0.028417736291885376,
-0.020812781527638435,
-0.027415424585342407,
-0.07143712788820267,
0.004921904299408197,
-0.012288804166018963,
-0.08847415447235107,
0.011235043406486511,
0.039848171174526215,
0.027507701888680458,
0.043007634580135345,
-2.0389873380103926e-33,
-0.07407139986753464,
0.1209237277507782,
-0.010619024746119976,
0.041768629103899,
-0.009363709948956966,
0.02138039469718933,
0.04941223934292793,
0.02501998096704483,
-0.029227327555418015,
0.03467611223459244,
-0.019076701253652573,
-0.014690430834889412,
-0.08253894746303558,
-0.04014744982123375,
0.018707476556301117,
0.04180007427930832,
-0.02294578030705452,
-0.06246231496334076,
0.022212930023670197,
0.01628783345222473,
-0.04859764873981476,
0.11646904796361923,
-0.05532019957900047,
0.05409613251686096,
0.03164070099592209,
0.07292310893535614,
0.011342226527631283,
0.040700215846300125,
0.027358999475836754,
0.008963079191744328,
-0.03482934087514877,
-0.011864415369927883,
-0.03326285257935524,
-0.0591207817196846,
0.031839389353990555,
-0.009209554642438889,
0.009368694387376308,
-0.05900543928146362,
-0.008469696156680584,
0.1350540667772293,
0.019864575937390327,
-0.038295675069093704,
-0.04146147146821022,
0.03603651747107506,
0.026906372979283333,
0.051424022763967514,
-0.03211361542344093,
-0.04951634630560875,
-0.05256517603993416,
0.02046395279467106,
0.0299909096211195,
0.046760398894548416,
-0.063632071018219,
0.005587174091488123,
0.00799502432346344,
-0.10169044137001038,
0.019929390400648117,
0.06788477301597595,
0.021280404180288315,
0.027707094326615334,
-0.027817213907837868,
-0.08483529090881348,
0.01725419983267784,
-0.02952663041651249,
-0.07399599999189377,
-0.025751732289791107,
0.00467473641037941,
0.0365678071975708,
0.008687637746334076,
-0.014206922613084316,
0.0315701961517334,
-0.02169341966509819,
-0.009180745109915733,
-0.057834457606077194,
0.03750338777899742,
-0.042891975492239,
0.04045531526207924,
0.005143951624631882,
-0.03275948762893677,
0.03765566274523735,
-0.0023745030630379915,
-0.011721831746399403,
0.026497546583414078,
0.050071246922016144,
0.10790453851222992,
-0.03696858137845993,
0.01551879197359085,
-0.02532833442091942,
0.048383843153715134,
0.04327290877699852,
-0.02473357506096363,
0.024148572236299515,
-0.10258617252111435,
-0.010788723826408386,
-0.018103690817952156,
-4.3732157450904197e-8,
-0.1565980315208435,
0.07826217263936996,
0.017969662323594093,
-0.010488149709999561,
0.079764723777771,
-0.03990153223276138,
0.0871317908167839,
0.07081466913223267,
0.009031162597239017,
0.059041738510131836,
0.1001223549246788,
0.021550524979829788,
-0.0296527910977602,
-0.07956048846244812,
0.0049973041750490665,
0.005554224830120802,
-0.007007020525634289,
-0.05199582129716873,
0.10067372769117355,
0.0107845114544034,
-0.01786351017653942,
-0.05895492807030678,
-0.004735099151730537,
0.06759639084339142,
0.02350599505007267,
-0.12660467624664307,
-0.0712917298078537,
0.04014968127012253,
0.039250448346138,
0.028733113780617714,
-0.05737558379769325,
0.001731278607621789,
0.09258409589529037,
0.02999534085392952,
0.05677992105484009,
0.0865304246544838,
0.004976985976099968,
0.012366552837193012,
-0.027867887169122696,
0.05427529662847519,
-0.05749804154038429,
0.0870845839381218,
0.017452919855713844,
-0.03771686926484108,
-0.026192670688033104,
-0.012782804667949677,
0.08729186654090881,
0.0040543884970247746,
-0.05016874149441719,
-0.0007062996155582368,
0.04933493584394455,
-0.04603404179215431,
-0.009940638206899166,
0.029304463416337967,
0.04046230390667915,
-0.08945590257644653,
-0.05791936442255974,
-0.03651346266269684,
-0.04308529198169708,
0.06527629494667053,
-0.01625118963420391,
0.04611724987626076,
-0.029479529708623886,
-0.0334109403192997
] | https://github.com/hormoz-lab/coed-gnn | 1 | 0 | 0 | 0 |
Accelerating Task Generalisation with Multi-Level Skill Hierarchies | https://openreview.net/forum?id=KfeRfxTemB | [
"Thomas P Cannon",
"Özgür Şimşek"
] | Poster | Developing reinforcement learning agents that can generalise effectively to new tasks is one of the main challenges in AI research. This paper introduces Fracture Cluster Options (FraCOs), a multi-level hierarchical reinforcement learning method designed to improve generalisation performance. FraCOs identifies patterns in agent behaviour and forms temporally-extended actions (options) based on the expected future usefulness of those patterns, enabling rapid adaptation to new tasks. In tabular settings, FraCOs demonstrates effective transfer and improves performance as the depth of the hierarchy increases. In several complex procedurally-generated environments, FraCOs consistently outperforms state-of-the-art deep reinforcement learning algorithms, achieving superior results in both in-distribution and out-of-distribution scenarios. | Reinforcement Learning, Generalisation, Hierarchical Reinforcement Learning | This work presents Fracture Cluster Options (FraCOs), a novel multi-level hierarchical reinforcement learning framework that improves task generalisation. | 12,221 | null | [
-0.03554587811231613,
-0.10325541347265244,
0.10002278536558151,
0.022470751777291298,
0.019948851317167282,
-0.014315110631287098,
-0.08971193432807922,
-0.0467827133834362,
-0.03702089563012123,
-0.02171851322054863,
-0.05932662636041641,
-0.10773806273937225,
-0.020581254735589027,
0.0634731575846672,
0.023418987169861794,
-0.027794428169727325,
0.05262413248419762,
-0.03875614330172539,
-0.06523827463388443,
-0.06731370091438293,
0.03505764529109001,
-0.019901994615793228,
0.08164872229099274,
-0.05510137602686882,
0.02538115903735161,
0.07794419676065445,
0.016516949981451035,
-0.04416031390428543,
0.04082917422056198,
-0.03378003463149071,
0.04137967526912689,
0.028536811470985413,
0.052680522203445435,
-0.045958008617162704,
-0.04169158264994621,
0.11405753344297409,
-0.13203264772891998,
0.04463471099734306,
-0.01808788627386093,
0.04593988507986069,
-0.00821008626371622,
-0.001277998322620988,
0.03677879646420479,
-0.013785404153168201,
0.046276554465293884,
0.017106493934988976,
0.003280098782852292,
-0.05529758334159851,
0.009763648733496666,
0.006397044286131859,
-0.041552674025297165,
0.0032007440458983183,
0.04941267520189285,
0.06001447141170502,
0.03150142356753349,
0.09633826464414597,
0.07047417759895325,
-0.031771671026945114,
0.02254391647875309,
-0.02297366037964821,
0.03854091092944145,
-0.05161655321717262,
0.01572774536907673,
-0.0914657860994339,
-0.018877362832427025,
-0.007961653172969818,
0.022304970771074295,
0.06128469109535217,
-0.014018114656209946,
0.005189936608076096,
-0.024137772619724274,
0.03626275807619095,
0.022823024541139603,
0.0033430419862270355,
0.0421275794506073,
0.048016343265771866,
-0.06043127179145813,
-0.07244302332401276,
0.021608520299196243,
0.02999234013259411,
-0.04144486039876938,
-0.020251182839274406,
-0.023230697959661484,
0.057218462228775024,
0.06011385843157768,
0.016462020576000214,
-0.06232813000679016,
0.024616053327918053,
0.059755075722932816,
-0.0342043973505497,
0.02574051544070244,
-0.007238013669848442,
0.020870443433523178,
-0.026194045320153236,
-0.034171391278505325,
0.07217275351285934,
-0.05008970573544502,
-0.09977897256612778,
-0.04896526038646698,
0.05081097036600113,
0.034174755215644836,
-0.017880653962492943,
-0.013978120870888233,
0.007694039959460497,
-0.013674749061465263,
-0.06620825082063675,
0.054881077259778976,
0.03452442213892937,
0.09178140014410019,
-0.024001672863960266,
0.030307075008749962,
0.019637636840343475,
0.06444671005010605,
0.018900947645306587,
-0.09703437983989716,
0.0037758240941911936,
0.013979652896523476,
-0.04161959886550903,
-0.09011726826429367,
0.0742654874920845,
0.05619799345731735,
-0.10709864646196365,
-0.01955990307033062,
0.03143042325973511,
0.037661999464035034,
-0.04879463091492653,
-0.0630122497677803,
2.147762809789687e-33,
0.05052928999066353,
-0.050554875284433365,
0.07126118987798691,
-0.012598346918821335,
0.09591849148273468,
-0.04571058228611946,
0.02510453760623932,
-0.020575325936079025,
-0.006921560503542423,
0.008793376386165619,
-0.02825549989938736,
0.07639046013355255,
-0.0739704966545105,
0.06671811640262604,
0.041213274002075195,
0.039503056555986404,
-0.03491482883691788,
0.08822804689407349,
-0.04828668758273125,
-0.07288197427988052,
0.027071194723248482,
0.007610611151903868,
0.024131275713443756,
-0.0549265556037426,
0.030050931498408318,
0.06468088179826736,
0.04861384257674217,
-0.02150115743279457,
0.04264674708247185,
-0.020886385813355446,
-0.056562528014183044,
0.0630921721458435,
-0.08383948355913162,
0.05799717456102371,
-0.025874873623251915,
0.05874606594443321,
-0.008655663579702377,
-0.04393838346004486,
-0.04579617455601692,
0.04340672865509987,
0.0031929274555295706,
0.02147681824862957,
0.09264243394136429,
0.002855514409020543,
-0.010630762204527855,
-0.017891477793455124,
0.004854366648942232,
0.025316691026091576,
-0.052122991532087326,
-0.015848280861973763,
0.004395671188831329,
-0.013042116537690163,
0.04394008591771126,
-0.08089029043912888,
0.05127009376883507,
0.08377152681350708,
-0.009281664155423641,
0.08093023300170898,
0.014975604601204395,
0.037799809128046036,
0.04201681911945343,
-0.056583452969789505,
-0.058413952589035034,
0.1255260854959488,
-0.05638383701443672,
0.1032879427075386,
-0.0012819120893254876,
0.08048784732818604,
0.12564311921596527,
0.02470853179693222,
0.032638948410749435,
0.0857989639043808,
-0.04514751955866814,
0.002480023307725787,
0.0004298633721191436,
-0.01887444034218788,
0.03213925287127495,
-0.06740640103816986,
0.006084976717829704,
-0.008118956349790096,
-0.032888609915971756,
-0.002695291768759489,
-0.1044759675860405,
0.0019534409511834383,
0.029400812461972237,
0.012334407307207584,
0.013847270049154758,
-0.057293280959129333,
0.01954539678990841,
-0.012728587724268436,
-0.11543870717287064,
-0.03915812447667122,
-0.03768504410982132,
-0.00683516263961792,
0.03626607730984688,
-2.3836382409096564e-33,
0.025591503828763962,
-0.04031597450375557,
-0.0936969742178917,
-0.03343793749809265,
0.08916955441236496,
0.024236902594566345,
-0.081537164747715,
-0.046367473900318146,
0.03076096437871456,
-0.013759994879364967,
-0.09770047664642334,
0.06928146630525589,
0.004380504600703716,
0.006088702939450741,
-0.05996138975024223,
-0.049359824508428574,
-0.023021647706627846,
-0.01959584467113018,
0.05396869406104088,
0.02321373112499714,
0.02992529235780239,
0.0163212139159441,
-0.0581609271466732,
0.05989168584346771,
0.064964659512043,
-0.05131497606635094,
-0.021325115114450455,
0.04991544410586357,
-0.016931872814893723,
0.07126089930534363,
-0.054615214467048645,
-0.06082853674888611,
-0.011648803018033504,
0.03185155615210533,
0.03550354391336441,
0.09240981191396713,
0.03321652114391327,
-0.03581010550260544,
-0.06119440495967865,
0.019722064957022667,
-0.024050235748291016,
-0.10286062210798264,
-0.05918699875473976,
0.049517881125211716,
-0.01750078611075878,
0.003396138548851013,
-0.046577706933021545,
-0.03775355592370033,
-0.05383114889264107,
0.0109519949182868,
-0.01965891756117344,
-0.03519943356513977,
-0.05717610567808151,
-0.03242230415344238,
-0.00734730577096343,
-0.004243479575961828,
-0.031446270644664764,
0.006368327885866165,
0.001368846744298935,
0.0024464691523462534,
-0.08102725446224213,
0.01720001921057701,
-0.02310270071029663,
0.0065592750906944275,
0.022713886573910713,
-0.0033249978441745043,
-0.039549510926008224,
0.035071343183517456,
-0.057561688125133514,
-0.011941557750105858,
-0.04282065108418465,
-0.0006799476686865091,
0.054894935339689255,
-0.00980238988995552,
0.01579027809202671,
-0.027700066566467285,
-0.03423912450671196,
-0.04601488262414932,
-0.03261208534240723,
-0.043698910623788834,
-0.007715370040386915,
-0.03306986019015312,
-0.03642816096544266,
0.07153357565402985,
0.03712945058941841,
0.05019712448120117,
0.045704495161771774,
0.17230947315692902,
0.06288642436265945,
-0.004234773572534323,
-0.008916862308979034,
-0.010931863449513912,
0.015360187739133835,
0.050211627036333084,
-0.12854425609111786,
-4.5493504075011515e-8,
-0.05949287861585617,
0.08619952946901321,
0.03204972296953201,
0.007543360814452171,
0.07451260834932327,
0.007129511330276728,
-0.0766502320766449,
0.10670751333236694,
-0.013923135586082935,
0.012469916604459286,
0.06231344863772392,
0.008029665797948837,
0.08454146236181259,
-0.03845597803592682,
0.0435267798602581,
0.03550082445144653,
0.05719146132469177,
0.011709732934832573,
-0.021731732413172722,
-0.01972629874944687,
0.021270541474223137,
0.022418664768338203,
0.0901745930314064,
-0.0324367955327034,
-0.03141281381249428,
-0.10455849766731262,
-0.10394606739282608,
0.003234486561268568,
0.024575451388955116,
0.00857527181506157,
0.002445807447656989,
-0.01714971847832203,
0.07962953299283981,
0.08463962376117706,
0.00872577354311943,
0.0880638062953949,
0.025429347530007362,
0.011726672761142254,
-0.0029627354815602303,
-0.046134766191244125,
-0.018436020240187645,
0.07718878984451294,
0.000028603646569536068,
-0.036104779690504074,
-0.042802389711141586,
0.02599993348121643,
-0.11398620903491974,
-0.0481375977396965,
0.11633005738258362,
-0.01544716116040945,
0.030123945325613022,
0.01792079024016857,
-0.034649722278118134,
0.07961449772119522,
0.09709853678941727,
0.0574546717107296,
0.03576909005641937,
-0.05077344551682472,
0.10321979969739914,
0.061885952949523926,
0.03794717416167259,
-0.05491043999791145,
-0.026583299040794373,
0.01550039928406477
] | 0 | 0 | 0 | 0 |
|
SiReRAG: Indexing Similar and Related Information for Multihop Reasoning | https://openreview.net/forum?id=yp95goUAT1 | [
"Nan Zhang",
"Prafulla Kumar Choubey",
"Alexander Fabbri",
"Gabriel Bernadett-Shapiro",
"Rui Zhang",
"Prasenjit Mitra",
"Caiming Xiong",
"Chien-Sheng Wu"
] | Poster | Indexing is an important step towards strong performance in retrieval-augmented generation (RAG) systems. However, existing methods organize data based on either semantic similarity (similarity) or related information (relatedness), but do not cover both perspectives comprehensively. Our analysis reveals that modeling only one perspective results in insufficient knowledge synthesis, leading to suboptimal performance on complex tasks requiring multihop reasoning. In this paper, we propose SiReRAG, a novel RAG indexing approach that explicitly considers both similar and related information. On the similarity side, we follow existing work and explore some variances to construct a similarity tree based on recursive summarization. On the relatedness side, SiReRAG extracts propositions and entities from texts, groups propositions via shared entities, and generates recursive summaries to construct a relatedness tree. We index and flatten both similarity and relatedness trees into a unified retrieval pool. Our experiments demonstrate that SiReRAG consistently outperforms state-of-the-art indexing methods on three multihop datasets (MuSiQue, 2WikiMultiHopQA, and HotpotQA), with an average 1.9% improvement in F1 scores. As a reasonably efficient solution, SiReRAG enhances existing reranking methods significantly, with up to 7.8% improvement in average F1 scores. Our code is available at https://github.com/SalesforceAIResearch/SiReRAG. | Retrieval-augmented generation (RAG), RAG indexing, Multi-hop question answering | We introduce an innovative RAG indexing approach that organizes data by considering both similarity and relatedness, achieving strong performance. | 12,220 | 2412.06206 | [
-0.04476034268736839,
-0.025367705151438713,
0.005491424351930618,
-0.018724236637353897,
0.0163686852902174,
0.0052431062795221806,
-0.029172666370868683,
0.005470513366162777,
0.028522146865725517,
-0.00541799096390605,
-0.025698477402329445,
-0.037410538643598557,
0.10745911300182343,
0.051337577402591705,
-0.004656257573515177,
0.04947753623127937,
0.0674143135547638,
0.030067216604948044,
-0.054382480680942535,
-0.11845553666353226,
0.027986282482743263,
0.05392662435770035,
0.03937831148505211,
-0.005122662056237459,
-0.030596498399972916,
0.00836032722145319,
0.01021565217524767,
0.042742449790239334,
0.003685315838083625,
0.0189149621874094,
0.0072629707865417,
0.1025458499789238,
-0.07263657450675964,
0.03747457265853882,
-0.08877238631248474,
0.13277146220207214,
-0.07907948642969131,
0.06911502033472061,
-0.033347804099321365,
-0.018281729891896248,
-0.048790473490953445,
0.051311757415533066,
0.003990893717855215,
0.02347276546061039,
0.05833026021718979,
0.030394231900572777,
-0.15380162000656128,
-0.022395791485905647,
0.01227495726197958,
0.08443549275398254,
-0.08735456317663193,
0.03987257555127144,
0.007545115426182747,
0.055330079048871994,
0.04379192367196083,
0.027973901480436325,
-0.03282688185572624,
0.050433382391929626,
0.005967200733721256,
-0.06241965666413307,
-0.040263108909130096,
-0.11154336482286453,
0.006921297404915094,
-0.09660810232162476,
-0.018812641501426697,
-0.0603962168097496,
0.06079484522342682,
0.02599433809518814,
-0.03948647901415825,
0.0031434388365596533,
-0.01477808691561222,
0.09658379852771759,
-0.04882192984223366,
0.10860653221607208,
-0.027331165969371796,
0.11789196729660034,
-0.045641787350177765,
-0.11339130252599716,
-0.0883798822760582,
-0.03808958828449249,
-0.05980846658349037,
0.037883564829826355,
0.026424460113048553,
0.006889179348945618,
0.11116188764572144,
0.0273249801248312,
0.03002966195344925,
-0.014530207961797714,
-0.01754879765212536,
0.002765586832538247,
-0.07367505878210068,
-0.04120282828807831,
0.09382154792547226,
0.022331077605485916,
0.07453867048025131,
0.10014280676841736,
0.013555225916206837,
-0.053939878940582275,
-0.027605077251791954,
0.030774615705013275,
0.006594483740627766,
0.1461094617843628,
-0.022677259519696236,
-0.030251266434788704,
-0.02179771289229393,
0.026981642469763756,
0.033689942210912704,
0.009340185672044754,
-0.010629403404891491,
-0.13787975907325745,
-0.010720884427428246,
0.10357757657766342,
0.003036331385374069,
0.0003433942038100213,
-0.022077718749642372,
0.007644824683666229,
0.02555084601044655,
0.030171703547239304,
0.0647454559803009,
0.023695943877100945,
0.06805580854415894,
0.034589823335409164,
0.01672040857374668,
0.055022671818733215,
0.010494614951312542,
-0.08108481019735336,
0.023364095017313957,
1.4840864134971006e-33,
0.057123009115457535,
0.05060776323080063,
0.041036464273929596,
0.10139050334692001,
0.067410908639431,
-0.0621441975235939,
-0.04645242169499397,
0.1168375015258789,
-0.056619178503751755,
-0.022311445325613022,
-0.007658054586499929,
0.07523007690906525,
-0.06506001949310303,
0.015804721042513847,
0.024583330377936363,
-0.03800168260931969,
-0.07466866075992584,
0.0325360968708992,
-0.0285202544182539,
-0.026496998965740204,
0.005474013276398182,
-0.010933727957308292,
0.014960234984755516,
0.0006738138617947698,
-0.04745770990848541,
0.05708184838294983,
0.034047167748212814,
-0.08220275491476059,
-0.0011385066900402308,
-0.030570117756724358,
-0.07227913290262222,
-0.019099244847893715,
0.02315327152609825,
-0.019486157223582268,
-0.018273895606398582,
-0.018322637304663658,
-0.0011985989985987544,
-0.07951557636260986,
0.04231230169534683,
0.061538081616163254,
-0.08570177853107452,
-0.01009923405945301,
-0.008618568070232868,
-0.09244508296251297,
-0.026586711406707764,
-0.014848449267446995,
-0.062406398355960846,
-0.01986094005405903,
0.09438326954841614,
0.0035148339811712503,
0.07144028693437576,
-0.040058210492134094,
0.06125058978796005,
-0.015511101111769676,
-0.011618787422776222,
-0.029522651806473732,
-0.001891773077659309,
0.005779835395514965,
0.053589705377817154,
0.06801403313875198,
0.025916550308465958,
0.07873756438493729,
-0.0035361009649932384,
0.06703203916549683,
-0.0027989251539111137,
0.07805510610342026,
0.050186410546302795,
0.10144598037004471,
0.12641991674900055,
0.03884405270218849,
-0.0042300657369196415,
0.0446845218539238,
0.0069866725243628025,
-0.07785753160715103,
0.025390304625034332,
0.046263959258794785,
-0.0582212470471859,
-0.0337231419980526,
-0.014727827161550522,
-0.03907158970832825,
-0.01223429199308157,
-0.013766160234808922,
-0.007891966961324215,
-0.07409267127513885,
-0.009926400147378445,
-0.034006088972091675,
0.029330380260944366,
-0.06557779759168625,
-0.046467386186122894,
-0.04679064080119133,
0.008614788763225079,
0.04684516414999962,
-0.04706859588623047,
-0.07551362365484238,
0.06922168284654617,
-1.6700354771307782e-33,
0.010359886102378368,
-0.0460013709962368,
-0.03130979835987091,
-0.006683430168777704,
0.02184157446026802,
-0.04650887846946716,
-0.018492316827178,
-0.0005955835222266614,
-0.09427057951688766,
-0.03919864073395729,
-0.07069966942071915,
-0.0663348063826561,
0.020494818687438965,
-0.0522506982088089,
-0.03642823174595833,
0.05877760797739029,
-0.011523310095071793,
0.06102879345417023,
-0.0693591833114624,
0.09463662654161453,
-0.01957288570702076,
0.07134956866502762,
-0.08203855156898499,
0.009047831408679485,
-0.010327855125069618,
0.036029599606990814,
0.024732299149036407,
0.0021983978804200888,
-0.024581780657172203,
-0.010386518202722073,
-0.009166762232780457,
-0.07080553472042084,
0.023212673142552376,
-0.12105756253004074,
-0.05534357950091362,
0.003651747712865472,
0.008149837143719196,
-0.04958438500761986,
-0.06937256455421448,
0.04598141461610794,
0.002344641136005521,
0.013679234310984612,
0.006713393609970808,
-0.01961604133248329,
-0.06578832864761353,
-0.00928400456905365,
-0.11544070392847061,
0.02522600069642067,
0.04949308931827545,
-0.006385040003806353,
0.037268608808517456,
-0.017236484214663506,
-0.02637500688433647,
-0.0280003622174263,
0.04501974210143089,
-0.06980133801698685,
-0.05695417523384094,
-0.021410347893834114,
-0.03971482068300247,
0.00687833409756422,
-0.016890428960323334,
0.05911099538207054,
-0.009563768282532692,
-0.006215385161340237,
0.039617639034986496,
-0.03907782956957817,
-0.0420047864317894,
-0.04189220815896988,
-0.06067639961838722,
0.009571666829288006,
-0.03614112734794617,
-0.015968361869454384,
0.06725288182497025,
-0.053051166236400604,
-0.01200482714921236,
-0.0023538407403975725,
0.06039024889469147,
0.020619433373212814,
-0.018169211223721504,
-0.06637255847454071,
-0.02922302484512329,
-0.030460499227046967,
0.021471280604600906,
0.052171409130096436,
-0.006704061757773161,
0.0021486796904355288,
0.03768610209226608,
0.11708274483680725,
0.027805842459201813,
0.053990498185157776,
-0.009592002257704735,
0.012002886272966862,
0.03440916910767555,
0.024523256346583366,
0.009255991317331791,
-4.310137313723317e-8,
-0.06696345657110214,
0.014418957754969597,
-0.05399139225482941,
0.04741506278514862,
-0.03747354447841644,
-0.01386262383311987,
-0.027135102078318596,
0.0682649090886116,
-0.08340612798929214,
-0.025263020768761635,
0.03937021642923355,
-0.021549364551901817,
-0.015543844550848007,
0.028965655714273453,
0.00017476369976066053,
0.023496462032198906,
0.04232332110404968,
-0.015329409390687943,
0.010789765976369381,
0.010894086211919785,
0.043832674622535706,
0.0585910938680172,
0.0008345518144778907,
0.007756729610264301,
-0.0020054136402904987,
-0.01563001051545143,
-0.03445397689938545,
0.032120347023010254,
0.11820077151060104,
-0.018937615677714348,
-0.03046281822025776,
-0.028693238273262978,
0.007013755850493908,
-0.07165374606847763,
0.034553300589323044,
0.010897867381572723,
-0.004322204273194075,
0.012551925145089626,
-0.010754446499049664,
-0.055409740656614304,
0.009362524375319481,
0.03396753966808319,
0.00443264190107584,
0.07348108291625977,
0.045441970229148865,
-0.04266652837395668,
-0.057866983115673065,
-0.02305191569030285,
0.010249026119709015,
0.01656704768538475,
-0.08871941268444061,
-0.13429255783557892,
0.03450123593211174,
-0.007645726669579744,
0.058258410543203354,
0.05450505018234253,
-0.09222032874822617,
0.044668763875961304,
0.11851668357849121,
-0.0723448172211647,
0.06714201718568802,
0.01062830165028572,
0.011980071663856506,
0.08437768369913101
] | 0 | 0 | 0 | 0 |
|
Make Haste Slowly: A Theory of Emergent Structured Mixed Selectivity in Feature Learning ReLU Networks | https://openreview.net/forum?id=27SSnLl85x | [
"Devon Jarvis",
"Richard Klein",
"Benjamin Rosman",
"Andrew M Saxe"
] | Poster | In spite of finite dimension ReLU neural networks being a consistent factor behind recent deep learning successes, a theory of feature learning in these models remains elusive. Currently, insightful theories still rely on assumptions including the linearity of the network computations, unstructured input data and architectural constraints such as infinite width or a single hidden layer. To begin to address this gap we establish an equivalence between ReLU networks and Gated Deep Linear Networks, and use their greater tractability to derive dynamics of learning. We then consider multiple variants of a core task reminiscent of multi-task learning or contextual control which requires both feature learning and nonlinearity. We make explicit that, for these tasks, the ReLU networks possess an inductive bias towards latent representations which are *not* strictly modular or disentangled but are still highly structured and reusable between contexts. This effect is amplified with the addition of more contexts and hidden layers. Thus, we take a step towards a theory of feature learning in finite ReLU networks and shed light on how structured mixed-selective latent representations can emerge due to a bias for node-reuse and learning speed. | Gated Deep Linear Networks, Feature Learning Dynamics, Structured Mixed Selectivity, ReLU Networks | We establish an equivalence between finite feature learning ReLU networks and Gated Deep Linear Networks to analyse the full learning dynamics of the ReLU network. We find a bias towards structured mixed selective representations on a set of tasks. | 12,212 | 2503.06181 | [
-0.04866139963269234,
-0.1312081664800644,
0.0621669702231884,
0.06381554156541824,
0.10083210468292236,
0.03135606274008751,
-0.03458347171545029,
-0.12747812271118164,
0.01176625955849886,
-0.06495501846075058,
-0.017265180125832558,
-0.03894799202680588,
0.0022682209964841604,
0.07224872708320618,
-0.0387553870677948,
0.022143064066767693,
0.0810655727982521,
0.09419707953929901,
-0.15590770542621613,
-0.04391276091337204,
0.026075832545757294,
-0.026283694431185722,
-0.008470266126096249,
0.00029065212584100664,
0.007630195468664169,
0.04209750145673752,
-0.007003416307270527,
-0.01826127991080284,
0.02577660232782364,
-0.034624967724084854,
0.08299495279788971,
0.04386667162179947,
-0.057810425758361816,
0.02093574032187462,
-0.07866969704627991,
0.11952666193246841,
-0.028651945292949677,
0.03085307590663433,
-0.03422495722770691,
-0.03628000617027283,
-0.010784351266920567,
0.001683396054431796,
-0.0356181338429451,
0.09969282150268555,
0.08752496540546417,
-0.03491923213005066,
-0.017843643203377724,
-0.05006972327828407,
-0.00915251299738884,
-0.023359619081020355,
-0.04028456658124924,
0.02251850627362728,
-0.032276008278131485,
0.05979006737470627,
0.053417082875967026,
0.009895642288029194,
0.03383319824934006,
0.05884832143783569,
-0.007706534583121538,
0.0007185331778600812,
0.004160908982157707,
-0.10834185034036636,
0.0033179603051394224,
-0.05937550589442253,
0.015880536288022995,
0.023106195032596588,
-0.01937876082956791,
0.0149989677593112,
-0.031231779605150223,
-0.019709452986717224,
0.06053174287080765,
0.07008372247219086,
-0.07118924707174301,
-0.03112008050084114,
0.07244785130023956,
-0.01963852345943451,
0.06375908106565475,
-0.019737739115953445,
0.04983201622962952,
-0.017747921869158745,
0.048782750964164734,
0.09686162322759628,
-0.009292553178966045,
-0.008396467193961143,
0.11165529489517212,
-0.015122774988412857,
-0.028035905212163925,
0.009221469052135944,
-0.023303197696805,
-0.025425206869840622,
-0.051623594015836716,
-0.07506106793880463,
0.0016853046836331487,
-0.07429534196853638,
0.09991955757141113,
0.007006087806075811,
0.007894894108176231,
-0.06697452813386917,
-0.03733038529753685,
0.06194163113832474,
-0.045261215418577194,
-0.03927876800298691,
0.01017649658024311,
0.05818697065114975,
0.006164532154798508,
-0.019905786961317062,
0.032950323075056076,
0.06520571559667587,
-0.0004819866444449872,
-0.08926674723625183,
0.03415645286440849,
-0.0005501064588315785,
0.046163130551576614,
0.06712722778320312,
0.006067480891942978,
-0.067184217274189,
0.04305762052536011,
-0.03571809455752373,
0.06497833132743835,
0.09065279364585876,
-0.03743581473827362,
-0.006626032758504152,
-0.02785002999007702,
-0.0021773541811853647,
0.014666391536593437,
-0.030910557135939598,
-0.0664924904704094,
2.9724808301189726e-33,
-0.0034561066422611475,
-0.02673669159412384,
0.017076008021831512,
0.03657691180706024,
0.024546435102820396,
-0.04949784278869629,
-0.03109728731215,
0.019241325557231903,
0.015033955685794353,
0.05545767769217491,
-0.007824592292308807,
0.004036167170852423,
-0.06333387643098831,
0.10525580495595932,
0.04487117379903793,
0.0029849472921341658,
-0.06453023850917816,
0.0011528951581567526,
0.03893166407942772,
-0.03781551495194435,
-0.07001899927854538,
0.004099466372281313,
-0.04495539143681526,
-0.054068125784397125,
-0.014665423892438412,
-0.054234858602285385,
0.05295582115650177,
-0.03144289180636406,
-0.02610466443002224,
0.004062522668391466,
-0.09934695810079575,
0.0561351478099823,
-0.040115367621183395,
0.10162456333637238,
-0.026570098474621773,
0.04388436675071716,
0.006218083668500185,
-0.03363070636987686,
0.05435650423169136,
-0.023275060579180717,
0.0612645298242569,
0.03181467950344086,
-0.02710549533367157,
-0.036406271159648895,
-0.0668521597981453,
-0.06750591844320297,
0.021135898306965828,
0.03345000371336937,
-0.10208367556333542,
-0.059365395456552505,
0.025794576853513718,
0.07573041319847107,
-0.009587761014699936,
-0.09392882883548737,
0.01658264547586441,
0.003722126130014658,
-0.03026367723941803,
0.06780078262090683,
0.07187893986701965,
0.08927901834249496,
-0.024829715490341187,
0.03755848854780197,
-0.05876744166016579,
0.08197153359651566,
-0.01292477734386921,
0.0755477249622345,
-0.016699595376849174,
0.08005024492740631,
0.11920876801013947,
-0.032781295478343964,
-0.039835043251514435,
0.08847222477197647,
-0.040279876440763474,
-0.11978600174188614,
0.052952416241168976,
-0.019295206293463707,
0.023216988891363144,
-0.09091117233037949,
0.007929529063403606,
0.08505801856517792,
0.01726469025015831,
0.030672157183289528,
-0.04969178885221481,
0.045602262020111084,
-0.09233833849430084,
-0.05850527808070183,
0.020054668188095093,
-0.05222077667713165,
0.01812141016125679,
-0.01828945055603981,
-0.05350786820054054,
-0.017146985977888107,
0.04879583790898323,
-0.023963216692209244,
0.04395312815904617,
-3.3897532229990604e-33,
-0.04160911962389946,
-0.020707888528704643,
-0.007321022916585207,
0.0013435997534543276,
-0.01192215271294117,
0.011248827911913395,
-0.0049860915169119835,
-0.03625262528657913,
-0.04731515422463417,
-0.052001241594552994,
0.02076636627316475,
-0.04289421811699867,
-0.02564619667828083,
-0.03744717687368393,
0.06871438026428223,
0.06447069346904755,
-0.082588329911232,
0.013426954858005047,
0.07631111145019531,
-0.03620408847928047,
0.04475926607847214,
0.021000145003199577,
-0.03604663535952568,
0.074140764772892,
0.0034395602997392416,
0.03520401939749718,
-0.10128873586654663,
0.040806882083415985,
0.03153716027736664,
0.027504025027155876,
-0.054105039685964584,
-0.04972351714968681,
-0.048408858478069305,
-0.002157260663807392,
-0.02485061250627041,
0.07514142990112305,
-0.008708828128874302,
0.040748853236436844,
-0.04574793577194214,
0.018419738858938217,
0.07547331601381302,
-0.03490322828292847,
-0.03816091641783714,
0.0648391842842102,
0.019595948979258537,
-0.05081285163760185,
0.017761994153261185,
0.011085725389420986,
0.002318748738616705,
0.03755759820342064,
-0.01346016488969326,
0.03965254873037338,
-0.10910512506961823,
-0.009930243715643883,
-0.08259888738393784,
0.003313666209578514,
0.07108496129512787,
-0.014580131508409977,
0.15440814197063446,
-0.0007360199233517051,
-0.03690304979681969,
-0.08188597857952118,
-0.02564026229083538,
0.020236097276210785,
0.042972028255462646,
-0.09693529456853867,
0.03189331293106079,
-0.05263784900307655,
-0.0024241572245955467,
0.011704746633768082,
-0.0054183038882911205,
-0.037592627108097076,
0.04882392659783363,
-0.0029251931700855494,
0.020570404827594757,
-0.07437511533498764,
-0.013715346343815327,
-0.04435311257839203,
-0.0505114421248436,
0.0015467541525140405,
0.03915731608867645,
-0.011541247367858887,
-0.0091441934928298,
0.005237191449850798,
0.07168986648321152,
0.06578287482261658,
-0.026918504387140274,
0.029994912445545197,
0.058968450874090195,
-0.014353316277265549,
0.0008217617869377136,
0.02370768040418625,
-0.06469009816646576,
0.0020255378913134336,
-0.007331816013902426,
-4.709390921675549e-8,
-0.06331643462181091,
0.022534944117069244,
0.024025555700063705,
-0.04252372309565544,
0.04735457897186279,
-0.021292323246598244,
0.045293938368558884,
0.09570835530757904,
-0.0082221869379282,
0.011284993030130863,
0.11275231093168259,
0.029315369203686714,
-0.02110723964869976,
-0.0689505934715271,
0.0056262207217514515,
0.08833145350217819,
0.029332365840673447,
-0.04350926727056503,
0.06174454465508461,
-0.013681730255484581,
0.10940441489219666,
-0.005491714458912611,
0.011432717554271221,
-0.010107997804880142,
-0.0319841131567955,
-0.11805056035518646,
-0.09050653874874115,
0.12703603506088257,
0.04669786989688873,
-0.018111450597643852,
-0.02897188998758793,
0.03911419212818146,
-0.0045384191907942295,
-0.012971160002052784,
-0.025053054094314575,
0.06520989537239075,
0.06298066675662994,
-0.02064238116145134,
-0.03243899345397949,
-0.0334915816783905,
-0.013761583715677261,
0.09796073287725449,
-0.01595477946102619,
-0.0004630698822438717,
0.009804898872971535,
-0.02636536955833435,
0.006272817496210337,
-0.08647334575653076,
0.03526541590690613,
-0.0000233327918977011,
0.04629296436905861,
-0.00854798499494791,
0.08093345910310745,
0.05870392173528671,
0.06439243257045746,
0.01675933040678501,
0.035750456154346466,
-0.047345634549856186,
0.028788452968001366,
0.047252628952264786,
-0.032175738364458084,
0.051427293568849564,
-0.07844226807355881,
-0.02736605703830719
] | 0 | 0 | 0 | 0 |
|
Eliciting Human Preferences with Language Models | https://openreview.net/forum?id=LvDwwAgMEW | [
"Belinda Z. Li",
"Alex Tamkin",
"Noah Goodman",
"Jacob Andreas"
] | Poster | Language models (LMs) can be directed to perform user- and context-dependent
tasks by using labeled examples or natural language prompts.
But selecting examples or writing prompts can be challenging---especially in tasks that require users to precisely articulate nebulous preferences or reason about complex edge cases. For such tasks, we introduce **Generative Active Task Elicitation (GATE)**, a method for using *LMs themselves* to guide the task specification process. GATE is a learning framework in which models elicit and infer human preferences through free-form, language-based interaction with users.
We identify prototypical challenges that users face when specifying preferences, and design three preference modeling tasks to study these challenges:
content recommendation, moral reasoning, and email validation.
In preregistered experiments, we show that LMs that learn to perform these tasks using GATE (by interactively querying users with open-ended questions) obtain preference specifications that are more informative than user-written prompts or examples. GATE matches existing task specification methods in the moral reasoning task, and significantly outperforms them in the content recommendation and email validation tasks. Users additionally report that interactive task elicitation requires less effort than prompting or example labeling and surfaces considerations that they did not anticipate on their own. Our findings suggest that LM-driven elicitation can be a powerful tool for aligning models to complex human preferences and values. | question asking, preference elicitation, language models, evaluation, human studies | learning personalized models by asking questions in language | 12,207 | 2310.11589 | [
-0.05989222601056099,
-0.08393222838640213,
0.01828186586499214,
0.027482660487294197,
0.05521198362112045,
-0.026989197358489037,
0.04930192977190018,
0.06962616741657257,
-0.011794081889092922,
0.005010266788303852,
-0.07013443112373352,
-0.12207898497581482,
0.05337323248386383,
0.06403975933790207,
0.12072191387414932,
0.050316229462623596,
0.11456537991762161,
-0.05256395787000656,
-0.08650944381952286,
-0.004342902451753616,
0.03957618772983551,
-0.000008886480827641208,
0.06969451904296875,
-0.01181467343121767,
-0.07537487149238586,
-0.053927384316921234,
0.03288891911506653,
-0.0017572366632521152,
0.010123117826879025,
0.026504354551434517,
0.01788489706814289,
-0.0024754535406827927,
0.04898183047771454,
-0.009484563954174519,
-0.04758419096469879,
0.006755574606359005,
-0.020957553759217262,
-0.033838722854852676,
-0.07158137857913971,
-0.10894657671451569,
-0.11736374348402023,
-0.0788230299949646,
0.015254659578204155,
0.04444611817598343,
0.02008162997663021,
-0.012784519232809544,
0.003386260475963354,
0.008166834712028503,
-0.03546777740120888,
-0.007346316706389189,
-0.04889383539557457,
-0.07866468280553818,
0.029944952577352524,
0.0016549821011722088,
-0.020832695066928864,
0.015864042565226555,
-0.006262979004532099,
0.028084399178624153,
0.03708134591579437,
-0.028457120060920715,
-0.07473444193601608,
-0.11074254661798477,
-0.053878724575042725,
0.0037059381138533354,
-0.08177056163549423,
-0.03277236223220825,
0.010597513988614082,
-0.012221287935972214,
-0.09284047782421112,
0.03879234194755554,
-0.03720986843109131,
0.05143605172634125,
-0.031106969341635704,
0.04590604081749916,
0.0028346122708171606,
0.015763696283102036,
-0.026196090504527092,
-0.07012307643890381,
-0.03342796489596367,
-0.04136444255709648,
-0.0602780282497406,
0.04364124685525894,
-0.0546102374792099,
0.05180516839027405,
0.12274840474128723,
0.02600233629345894,
0.011313862167298794,
0.07207407057285309,
0.06702791154384613,
0.07236606627702713,
-0.0009296106291003525,
-0.02633737586438656,
-0.006082253064960241,
-0.001506246393546462,
0.023340493440628052,
0.07940123975276947,
-0.009674527682363987,
-0.07759430259466171,
-0.07741031050682068,
0.05077575519680977,
0.030509483069181442,
0.09039796143770218,
-0.034724343568086624,
-0.1270260512828827,
0.0547095388174057,
0.00869836937636137,
-0.02941107749938965,
0.055718161165714264,
0.011280549690127373,
-0.04675830155611038,
0.029895922169089317,
0.046331409364938736,
0.03815058246254921,
-0.046317361295223236,
0.027426665648818016,
-0.07760787755250931,
0.070671945810318,
-0.027514532208442688,
0.05061502754688263,
0.08039151132106781,
0.01868988759815693,
0.030602319166064262,
0.05213326960802078,
0.009592171758413315,
0.031522784382104874,
-0.075368732213974,
-0.03234191983938217,
1.0212985120111893e-33,
0.0671299621462822,
0.03128425404429436,
0.024760479107499123,
0.06179167330265045,
0.040880270302295685,
-0.003666552249342203,
-0.05593336001038551,
0.09184587001800537,
-0.04952307045459747,
-0.05023287609219551,
0.06213878095149994,
0.026692628860473633,
-0.010763009078800678,
0.08695188909769058,
-0.003157127182930708,
0.003117778804153204,
-0.021951453760266304,
0.007247173693031073,
-0.05317851901054382,
-0.06724017858505249,
0.017106305807828903,
-0.002290187869220972,
0.021976331248879433,
0.019324250519275665,
-0.03399213030934334,
-0.02814836986362934,
0.03974107652902603,
-0.06077536940574646,
0.0675014853477478,
-0.005641098599880934,
-0.08108403533697128,
-0.08979914337396622,
0.048501383513212204,
0.0013616398209705949,
-0.058047741651535034,
0.04645787179470062,
-0.025858409702777863,
-0.030062908306717873,
0.08454965054988861,
-0.015180378220975399,
0.04652579873800278,
0.016252724453806877,
0.08398203551769257,
0.08814530074596405,
-0.03215562179684639,
-0.05978390574455261,
0.02572980523109436,
-0.0009828897891566157,
0.025191253051161766,
-0.00966294389218092,
0.07427413016557693,
0.012256142683327198,
0.03538350388407707,
-0.020693965256214142,
0.006950323469936848,
-0.03228326514363289,
-0.01393955759704113,
0.02383807674050331,
-0.021242521703243256,
-0.058675337582826614,
0.0033922153525054455,
0.06752616912126541,
0.036316484212875366,
-0.003962223418056965,
-0.04462115094065666,
-0.031006501987576485,
-0.0366608090698719,
0.05829205736517906,
0.010412978008389473,
-0.05876068025827408,
-0.03943827748298645,
0.03677178919315338,
-0.023335443809628487,
-0.03036060743033886,
-0.011033348739147186,
0.05930820107460022,
0.006146133411675692,
-0.07549457252025604,
0.07695933431386948,
-0.034414127469062805,
0.07767143845558167,
-0.054049406200647354,
0.05207476019859314,
-0.012890120036900043,
0.009691325016319752,
-0.0731879398226738,
0.04408787563443184,
-0.06732740998268127,
-0.0038523327093571424,
-0.07463598251342773,
0.008520130068063736,
-0.023652980104088783,
-0.0350421778857708,
0.035651132464408875,
0.02390732802450657,
-1.862521024038997e-33,
-0.032752495259046555,
-0.056075192987918854,
-0.020169757306575775,
0.027719566598534584,
-0.03900458291172981,
-0.008683443069458008,
0.012841053307056427,
-0.000867775350343436,
0.009344027377665043,
-0.0953449010848999,
-0.08395452052354813,
-0.024551767855882645,
0.08065098524093628,
0.0022299205884337425,
0.05505213513970375,
-0.03831030800938606,
-0.055441711097955704,
-0.026314804330468178,
-0.02478840760886669,
0.06849933415651321,
-0.020293401554226875,
0.12077198177576065,
-0.12890666723251343,
0.03533916920423508,
-0.06667674332857132,
0.026795364916324615,
-0.04938935860991478,
0.03255128487944603,
0.04633614420890808,
-0.014212574809789658,
-0.06736647337675095,
-0.035567428916692734,
-0.02635377086699009,
-0.08303835242986679,
-0.006447202526032925,
-0.00012177274038549513,
0.03160669282078743,
-0.0317273810505867,
-0.07311974465847015,
0.12100129574537277,
0.044783152639865875,
-0.009134707972407341,
-0.035079699009656906,
0.012010768055915833,
-0.07883874326944351,
0.04620465263724327,
-0.005865494720637798,
-0.03569575026631355,
0.012636969797313213,
-0.028808459639549255,
-0.000639832578599453,
-0.00897421408444643,
-0.034552544355392456,
-0.0008354336023330688,
-0.04061005637049675,
-0.09147929400205612,
-0.07307650148868561,
-0.027377985417842865,
0.11094304919242859,
0.011250201612710953,
-0.06530926376581192,
0.04263336583971977,
0.08080912381410599,
-0.04187660291790962,
0.0002603808243293315,
-0.06869200617074966,
-0.019247109070420265,
-0.013887244276702404,
-0.056192342191934586,
0.01418938860297203,
0.051855798810720444,
-0.03836338222026825,
-0.035867735743522644,
-0.09299453347921371,
0.06835298985242844,
0.02840694598853588,
0.047031838446855545,
-0.0639362633228302,
-0.06435723602771759,
-0.03719532489776611,
0.11470060795545578,
0.02066916413605213,
0.046807143837213516,
0.021204134449362755,
0.0036679974291473627,
0.01343004684895277,
-0.03342338278889656,
0.12053903192281723,
-0.009163680486381054,
0.0059417481534183025,
0.050861652940511703,
0.0418548546731472,
0.02830856293439865,
0.10208220034837723,
-0.04290036857128143,
-4.2112400677751793e-8,
-0.04837055876851082,
-0.010519561357796192,
0.0077474541030824184,
0.08217623829841614,
0.04419422522187233,
-0.019137121737003326,
-0.022802822291851044,
-0.00125077273696661,
0.02541511133313179,
-0.07536906003952026,
0.008684187196195126,
0.04696952924132347,
-0.03854449838399887,
-0.051858093589544296,
-0.014249501749873161,
0.09986361116170883,
0.028775213286280632,
-0.022381499409675598,
0.026634996756911278,
-0.00758513854816556,
0.05927872657775879,
0.0011196645209565759,
-0.00992735568434,
-0.007901717908680439,
0.04014009237289429,
0.01997087523341179,
-0.0601007342338562,
0.10234599560499191,
0.07251560688018799,
0.0614430196583271,
0.029921505600214005,
-0.04064856842160225,
-0.03177597373723984,
0.010214926674962044,
0.006255120504647493,
0.046667587012052536,
-0.013105549849569798,
-0.0632447600364685,
0.01868746615946293,
-0.037166353315114975,
0.13331376016139984,
0.08299000561237335,
-0.02358211949467659,
-0.004724365193396807,
0.007431317586451769,
0.024541975930333138,
-0.004964249208569527,
-0.11662069708108902,
0.016453474760055542,
-0.013652921654284,
-0.06368004530668259,
-0.11121315509080887,
0.01712135039269924,
0.06638859957456589,
0.02160518430173397,
0.008634123019874096,
-0.03202375769615173,
0.034788038581609726,
0.06922899186611176,
0.0333644263446331,
0.08094803988933563,
0.09129489958286285,
-0.023544002324342728,
0.05034381151199341
] | https://github.com/alextamkin/generative-elicitation | 126 | 0 | 0 | 0 |
SymDiff: Equivariant Diffusion via Stochastic Symmetrisation | https://openreview.net/forum?id=i1NNCrRxdM | [
"Leo Zhang",
"Kianoosh Ashouritaklimi",
"Yee Whye Teh",
"Rob Cornish"
] | Poster | We propose SymDiff, a method for constructing equivariant diffusion models using the framework of stochastic symmetrisation. SymDiff resembles a learned data augmentation that is deployed at sampling time, and is lightweight, computationally efficient, and easy to implement on top of arbitrary off-the-shelf models. In contrast to previous work, SymDiff typically does not require any neural network components that are intrinsically equivariant, avoiding the need for complex parameterisations or the use of higher-order geometric features. Instead, our method can leverage highly scalable modern architectures as drop-in replacements for these more constrained alternatives. We show that this additional flexibility yields significant empirical benefit for E(3)-equivariant molecular generation. To the best of our knowledge, this is the first application of symmetrisation to generative modelling, suggesting its potential in this domain more generally. | Equivariance, Diffusion Models, Symmetrisation, Molecular Generation, Markov Categories | We propose SymDiff, a novel method for constructing equivariant diffusion models using the recently introduced framework of stochastic symmetrisation. | 12,205 | 2410.06262 | [
-0.053422488272190094,
-0.14253850281238556,
0.08960968255996704,
-0.008471368812024593,
0.023380247876048088,
-0.028265342116355896,
-0.0066609070636332035,
-0.06195877492427826,
-0.033086102455854416,
-0.02940616011619568,
0.04582223296165466,
-0.05497157201170921,
-0.0416061207652092,
0.018792347982525826,
0.05021878704428673,
-0.008233590982854366,
-0.02511351928114891,
0.08920451253652573,
-0.06612298637628555,
-0.030376382172107697,
-0.04063595458865166,
0.020131191238760948,
0.004615808837115765,
-0.022315196692943573,
0.0980844646692276,
0.00484218867495656,
0.0889296904206276,
0.0023767584934830666,
0.005506827030330896,
-0.04262569174170494,
0.10112989693880081,
0.024769017472863197,
-0.08933252841234207,
-0.04577939584851265,
-0.028938794508576393,
0.04715847969055176,
-0.07912877947092056,
0.022871170192956924,
-0.011195961385965347,
-0.0613149031996727,
-0.014733798801898956,
-0.005430011078715324,
0.008736769668757915,
0.026893680915236473,
0.04308418184518814,
-0.025034351274371147,
0.0383216068148613,
-0.013524544425308704,
0.02801864594221115,
0.016141457483172417,
0.030264265835285187,
-0.01508853118866682,
-0.1180419772863388,
0.02712693065404892,
-0.00003282932811998762,
-0.02257394790649414,
-0.014817857183516026,
-0.06041877344250679,
-0.022630658000707626,
-0.06402792036533356,
-0.010927782393991947,
-0.06481599062681198,
-0.005131151061505079,
-0.03688305616378784,
-0.018934952095150948,
-0.01245441660284996,
0.10652091354131699,
0.05822693184018135,
0.08041831105947495,
0.051896873861551285,
-0.03286511078476906,
0.058515727519989014,
-0.1325419843196869,
0.03175416216254234,
-0.007708340883255005,
0.05990190804004669,
0.06934189796447754,
-0.06661724299192429,
0.06494057178497314,
-0.0376739576458931,
-0.023843083530664444,
0.09583902359008789,
0.10871625691652298,
-0.0018531790701672435,
0.05823059007525444,
0.01295660063624382,
-0.0062211728654801846,
0.034923896193504333,
0.02861589379608631,
-0.009544461034238338,
0.00344859273172915,
0.056155022233724594,
0.017592638731002808,
-0.06091282516717911,
0.0058184340596199036,
0.06959344446659088,
0.001203338848426938,
-0.008449437096714973,
0.002875166479498148,
0.10746251046657562,
-0.0047495425678789616,
0.07316694408655167,
-0.015838542953133583,
-0.01371791772544384,
0.0621679462492466,
-0.0845520868897438,
0.05352407321333885,
0.009187559597194195,
-0.06037738546729088,
-0.08314286917448044,
-0.03307298943400383,
0.0391213484108448,
0.00273317308165133,
-0.06874559819698334,
-0.06595415621995926,
0.003455429570749402,
-0.009188578464090824,
-0.03579917550086975,
0.06519018113613129,
0.02390507608652115,
-0.0494328998029232,
0.031103210523724556,
-0.12467706948518753,
0.09284909814596176,
0.0824788510799408,
0.018237782642245293,
-0.06173017993569374,
3.415176594413454e-33,
0.01841624081134796,
0.008205442689359188,
0.06588586419820786,
0.03090045042335987,
0.05694342032074928,
0.0504809133708477,
0.011776018887758255,
-0.04463166370987892,
-0.01961723528802395,
-0.005473448429256678,
-0.14574071764945984,
-0.036661919206380844,
-0.06417685747146606,
0.012888503260910511,
0.008049217984080315,
-0.06856151670217514,
-0.0645972415804863,
0.05385793745517731,
0.004000271204859018,
-0.034813206642866135,
0.02775879018008709,
0.013805779628455639,
-0.02814575470983982,
0.05039701610803604,
-0.08345934003591537,
-0.0008494479698128998,
0.05389391630887985,
0.010096444748342037,
0.009790156967937946,
0.028527939692139626,
-0.05826205015182495,
-0.058108050376176834,
-0.08041011542081833,
0.07300494611263275,
0.03165546432137489,
-0.04620122164487839,
-0.03474941477179527,
-0.08316995203495026,
0.091921366751194,
-0.04739858955144882,
0.025323815643787384,
0.05366341397166252,
-0.013330018147826195,
-0.0015951256500557065,
-0.15585364401340485,
0.02703903801739216,
0.05766008049249649,
-0.03301183506846428,
-0.01549313124269247,
-0.07855594158172607,
0.02409793809056282,
0.030242882668972015,
0.021693281829357147,
-0.015032864175736904,
-0.05226866155862808,
0.01207599975168705,
0.07469051331281662,
-0.01307758316397667,
0.014441319741308689,
0.042287908494472504,
-0.025878814980387688,
0.04834362491965294,
0.028423473238945007,
-0.008732317946851254,
0.05221946910023689,
0.052486877888441086,
-0.003854545531794429,
0.05957454815506935,
0.0755334347486496,
-0.03677092492580414,
-0.03568357601761818,
0.011934979818761349,
-0.043256595730781555,
-0.06861510127782822,
0.08263915777206421,
0.03290192037820816,
0.018369119614362717,
-0.07906031608581543,
0.037837132811546326,
0.02053978107869625,
-0.08454471081495285,
0.07543428242206573,
-0.06798169761896133,
-0.004282957874238491,
0.03419657051563263,
-0.0728021115064621,
-0.021404791623353958,
-0.11586958169937134,
-0.028575394302606583,
-0.09688879549503326,
-0.04668903723359108,
-0.088503897190094,
-0.042891472578048706,
-0.01228961069136858,
-0.01578557677567005,
-5.020094459954478e-33,
-0.07406019419431686,
0.019429776817560196,
0.019559256732463837,
0.04947550967335701,
-0.0180628914386034,
-0.06378896534442902,
-0.020914586260914803,
0.02557861991226673,
0.03672529384493828,
0.006320602726191282,
-0.0008217808208428323,
-0.01692448928952217,
-0.02540154568850994,
-0.09535766392946243,
0.06376739591360092,
-0.00006363438296830282,
-0.04922901466488838,
-0.03349215164780617,
-0.014555319212377071,
0.040925391018390656,
-0.06729485839605331,
0.049316175282001495,
-0.038840312510728836,
0.02494313009083271,
-0.040416546165943146,
0.023154471069574356,
-0.019704323261976242,
0.056321993470191956,
0.03734564036130905,
-0.00026514369528740644,
-0.04460061714053154,
-0.02404094487428665,
-0.016764605417847633,
0.025479180738329887,
-0.0637989342212677,
0.06505735218524933,
0.08423694968223572,
0.021133825182914734,
-0.05063639581203461,
0.06618435680866241,
0.009657451882958412,
-0.033347658812999725,
-0.014637008309364319,
0.08355055749416351,
0.0325373075902462,
0.03800307959318161,
-0.053133174777030945,
0.035446230322122574,
-0.007248906884342432,
-0.037810973823070526,
0.031720466911792755,
0.07782804220914841,
-0.034776583313941956,
0.014144353568553925,
-0.011835997924208641,
-0.01910201646387577,
-0.013298406265676022,
-0.09072788059711456,
0.06898030638694763,
0.023749327287077904,
-0.08129139989614487,
-0.09038720279932022,
0.049202222377061844,
-0.027024194598197937,
-0.06797251850366592,
-0.061168164014816284,
-0.031202634796500206,
-0.03489575535058975,
0.043081045150756836,
-0.006665206514298916,
0.03260481357574463,
-0.008073143661022186,
-0.0019657372031360865,
-0.001146097551099956,
0.021750185638666153,
-0.07828480750322342,
0.09244899451732635,
-0.03609584644436836,
0.02796158194541931,
-0.04419507831335068,
0.029771506786346436,
-0.009286043234169483,
0.052756018936634064,
0.038768939673900604,
0.06146864965558052,
-0.058199893683195114,
0.0715259313583374,
0.04407064989209175,
0.044069353491067886,
-0.013607103377580643,
-0.03389503434300423,
0.0320323146879673,
0.05287528038024902,
0.18039023876190186,
-0.025583669543266296,
-4.569229261619512e-8,
-0.03364719823002815,
0.017136290669441223,
0.06703708320856094,
-0.02320573478937149,
0.10955309867858887,
0.025242727249860764,
-0.0022226597648113966,
0.1101367399096489,
-0.006634360644966364,
-0.02450893074274063,
0.11440412700176239,
-0.026691650971770287,
-0.008972633630037308,
-0.05206897109746933,
0.019510705024003983,
0.07539988309144974,
0.005188151728361845,
0.02492651343345642,
-0.044467177242040634,
0.02424001134932041,
-0.009890514425933361,
0.054857343435287476,
-0.03242134675383568,
-0.02025403454899788,
0.14567497372627258,
-0.05453293025493622,
-0.008138788864016533,
-0.004915351513773203,
0.07224806398153305,
0.038617540150880814,
-0.018012551590800285,
-0.016055839136242867,
0.03476538509130478,
0.03193167969584465,
-0.032993536442518234,
0.08971745520830154,
0.03541414812207222,
-0.056776344776153564,
-0.06600802391767502,
-0.043016668409109116,
-0.05802805349230766,
0.011093522422015667,
-0.023949366062879562,
0.03291567787528038,
0.011955351568758488,
0.0796864777803421,
-0.0019657681696116924,
-0.0011351932771503925,
0.02346731536090374,
0.031318020075559616,
0.015797212719917297,
-0.031721677631139755,
-0.021526819095015526,
0.003479578997939825,
-0.04634916037321091,
0.005305393598973751,
-0.06110594794154167,
-0.058755066245794296,
0.07908652722835541,
-0.011493771336972713,
-0.003172890515998006,
-0.012681315653026104,
-0.003914413042366505,
-0.0020600040443241596
] | 0 | 0 | 0 | 0 |
|
AdaWM: Adaptive World Model based Planning for Autonomous Driving | https://openreview.net/forum?id=NEu8wgPctU | [
"Hang Wang",
"Xin Ye",
"Feng Tao",
"Chenbin Pan",
"Abhirup Mallik",
"Burhaneddin Yaman",
"Liu Ren",
"Junshan Zhang"
] | Poster | World model based reinforcement learning (RL) has emerged as a promising approach for autonomous driving, which learns a latent dynamics model and uses it to train a planning policy. To speed up the learning process, the pretrain-finetune paradigm is often used, where online RL is initialized by a pretrained model and a policy learned offline. However, naively performing such initialization in RL may result in dramatic performance degradation during the online interactions in the new task. To tackle this challenge, we first analyze the performance degradation and identify two primary root causes therein: the mismatch of the planning policy and the mismatch of the dynamics model, due to distribution shift. We further analyze the effects of these factors on performance degradation during finetuning, and our findings reveal that the choice of finetuning strategies plays a pivotal role in mitigating these effects. We then introduce AdaWM, an Adaptive World Model based planning method, featuring two key steps: (a) mismatch identification, which quantifies the mismatches and informs the finetuning strategy, and (b) alignment-driven finetuning, which selectively updates either the policy or the model as needed using efficient low-rank updates. Extensive experiments on the challenging CARLA driving tasks demonstrate that AdaWM significantly improves the finetuning process, resulting in more robust and efficient performance in autonomous driving systems. | World Model, Autonomous Driving, Reinforcement Learning | null | 12,202 | 2501.13072 | [
-0.00860612466931343,
-0.13626404106616974,
0.03332294151186943,
0.10481686890125275,
0.038909513503313065,
-0.008543863892555237,
-0.05933695286512375,
0.00037232463364489377,
-0.02984675019979477,
0.01692085899412632,
-0.05310770869255066,
0.02095573954284191,
-0.01713433489203453,
0.0522192120552063,
-0.012109915725886822,
0.013986732810735703,
0.04797855764627457,
0.01969355158507824,
-0.10222238302230835,
-0.002811271231621504,
0.027855081483721733,
0.05714542791247368,
0.08158998936414719,
0.06636893004179001,
-0.012859068810939789,
0.005763188935816288,
0.018464786931872368,
0.01392260193824768,
0.020749269053339958,
-0.0670221671462059,
0.04866471514105797,
-0.017447317019104958,
0.013933444395661354,
-0.06644590944051743,
-0.0848037526011467,
0.008721999824047089,
-0.1068452000617981,
-0.019095158204436302,
0.015048472210764885,
-0.02027466520667076,
0.00003205435132258572,
-0.0054702539928257465,
-0.008649013005197048,
-0.018331147730350494,
0.023279570043087006,
-0.0006855186657048762,
-0.004821046721190214,
0.010248295962810516,
0.027127524837851524,
0.0036803469993174076,
-0.05944937840104103,
0.023024417459964752,
0.004891716409474611,
-0.038133569061756134,
0.008140137419104576,
0.05993069335818291,
0.05854695290327072,
0.05455732345581055,
0.024658646434545517,
-0.04265357926487923,
-0.054332803934812546,
-0.11979622393846512,
0.037609755992889404,
0.0005647158832289279,
-0.09523121267557144,
-0.0043804701417684555,
0.011264266446232796,
0.0663006603717804,
0.023782914504408836,
0.020950181409716606,
-0.031742796301841736,
0.049748700112104416,
-0.008155045099556446,
-0.022866137325763702,
0.04390101507306099,
-0.006936687044799328,
0.024654243141412735,
0.045561324805021286,
0.05337492376565933,
0.0070387679152190685,
0.03266371414065361,
0.05102958157658577,
-0.004944810178130865,
0.03089124709367752,
0.11024192720651627,
-0.02976025827229023,
-0.03665933385491371,
0.038033775985240936,
0.10898128896951675,
-0.06306836754083633,
0.0052374377846717834,
-0.03007332608103752,
0.017329024150967598,
0.016000548377633095,
-0.009844371117651463,
0.07355491071939468,
0.0331294946372509,
-0.10851354151964188,
-0.03341357782483101,
0.06741824001073837,
0.0035524785052984953,
0.07259640842676163,
-0.013414384797215462,
0.06473003327846527,
-0.021285980939865112,
-0.039535291492938995,
0.030542345717549324,
0.08151022344827652,
0.03742220997810364,
-0.0408419668674469,
0.10435216873884201,
-0.032845549285411835,
0.07651117444038391,
0.04131516069173813,
-0.05914710834622383,
0.04224029928445816,
-0.02849714644253254,
0.015826774761080742,
-0.04608205705881119,
0.09972240030765533,
-0.053357068449258804,
-0.10122950375080109,
0.044959668070077896,
-0.008106554858386517,
0.016773449257016182,
-0.08814219385385513,
0.02996501512825489,
2.6159328223634334e-33,
0.023075710982084274,
-0.05584871023893356,
0.026088332757353783,
0.007331768982112408,
0.07629387080669403,
-0.010115494951605797,
0.01087302714586258,
0.036056045442819595,
-0.02761474810540676,
-0.02485712617635727,
0.03522269055247307,
-0.02933007851243019,
-0.045769140124320984,
0.09244277328252792,
0.024935146793723106,
-0.02058098278939724,
-0.12904705107212067,
0.06160922348499298,
-0.025971049442887306,
0.018978210166096687,
0.07484904676675797,
-0.05190036818385124,
-0.021979231387376785,
-0.10862772166728973,
0.04381978511810303,
0.10210677981376648,
0.034822117537260056,
0.06792410463094711,
0.017252640798687935,
0.003222052240744233,
-0.03271852806210518,
0.005158877000212669,
-0.16224542260169983,
0.05592937022447586,
0.013012277893722057,
0.006105436477810144,
-0.11571194231510162,
-0.06877758353948593,
-0.020839214324951172,
-0.019285619258880615,
0.0023163575679063797,
0.00661202147603035,
-0.033537957817316055,
0.01907213404774666,
-0.03318142145872116,
-0.07987100630998611,
0.07714405655860901,
0.00985740590840578,
-0.07082640379667282,
0.029207512736320496,
-0.028052357956767082,
0.021145934239029884,
-0.03477032110095024,
-0.11931141465902328,
-0.08693021535873413,
0.017473386600613594,
0.02510121837258339,
0.03245459496974945,
-0.03195315971970558,
0.03428605943918228,
0.00746992789208889,
-0.02943108230829239,
-0.044791627675294876,
0.09525245428085327,
0.04629452899098396,
0.03514767065644264,
-0.013549691066145897,
0.0551363006234169,
0.04887295141816139,
-0.03673925623297691,
0.08482769131660461,
-0.012204051949083805,
-0.004292334895581007,
0.039012446999549866,
0.08130922168493271,
-0.06366810947656631,
0.06827333569526672,
-0.034018851816654205,
-0.02870054356753826,
-0.016364963725209236,
-0.036613039672374725,
0.09052897244691849,
-0.04602324217557907,
-0.024727940559387207,
0.0005757787148468196,
-0.05803588777780533,
0.004589042626321316,
-0.03151141107082367,
0.003025825135409832,
0.03712120279669762,
-0.054226361215114594,
-0.03211746737360954,
-0.07566265016794205,
0.07558442652225494,
-0.009270243346691132,
-2.937532281043072e-33,
0.01016774121671915,
-0.07281395047903061,
0.013355372473597527,
0.02442242205142975,
-0.009551688097417355,
-0.048967305570840836,
-0.007849541492760181,
-0.009003500454127789,
0.05618228018283844,
-0.06111212447285652,
-0.04939763620495796,
-0.002209431491792202,
0.00839778408408165,
0.015533478930592537,
0.05768740922212601,
-0.0774659588932991,
0.03647613897919655,
-0.0008307173266075552,
0.04086185246706009,
0.00558787677437067,
0.03356018662452698,
0.006143157836049795,
-0.15206369757652283,
0.035771384835243225,
-0.013282712548971176,
0.01771407574415207,
-0.020158318802714348,
0.07988335192203522,
0.010166095569729805,
0.06356927007436752,
-0.0233127661049366,
0.01161184348165989,
-0.09640110284090042,
-0.006139015778899193,
-0.03822919726371765,
0.1009223684668541,
-0.026649050414562225,
-0.005033695138990879,
-0.09835873544216156,
0.07440005987882614,
0.047706905752420425,
-0.0778050497174263,
-0.047891609370708466,
-0.014624799601733685,
0.04663758724927902,
0.013291189447045326,
-0.07121692597866058,
0.022734640166163445,
-0.025480709969997406,
0.036478012800216675,
-0.0012881457805633545,
0.03270344063639641,
-0.06745921075344086,
0.007272620685398579,
-0.05415816605091095,
-0.011875750496983528,
0.013216055929660797,
-0.011423819698393345,
0.00379342888481915,
0.05195192992687225,
-0.01443629339337349,
0.04168239235877991,
0.021729296073317528,
0.023790163919329643,
-0.04675954952836037,
-0.057713378220796585,
-0.02335387095808983,
0.01805567927658558,
0.07574533671140671,
0.035929325968027115,
0.0089067742228508,
0.004772732965648174,
-0.007172385696321726,
-0.016272511333227158,
0.07414223253726959,
-0.07803581655025482,
-0.020439879968762398,
0.029256435111165047,
-0.05003857985138893,
-0.04627064988017082,
0.035069502890110016,
0.049244288355112076,
0.013015102595090866,
0.11924582719802856,
-0.06539065390825272,
0.07490487396717072,
-0.00926565658301115,
0.1115037277340889,
-0.011510484851896763,
0.01865476556122303,
0.009144812822341919,
0.001270189881324768,
-0.04088229686021805,
0.051247935742139816,
-0.12106610089540482,
-4.8360707438632744e-8,
-0.037624381482601166,
0.091512531042099,
0.07776863127946854,
-0.005716786719858646,
0.04292730987071991,
-0.03227612376213074,
0.009592287242412567,
0.01666618324816227,
-0.06125852093100548,
-0.01076897419989109,
0.07271447777748108,
0.011015212163329124,
0.0401395782828331,
-0.050851792097091675,
-0.011290349997580051,
0.06901288777589798,
0.026834459975361824,
-0.023408953100442886,
-0.0007984157418832183,
0.03379957377910614,
0.05598653852939606,
0.018876221030950546,
-0.022532083094120026,
-0.061645690351724625,
0.09896431118249893,
-0.10011646896600723,
0.0023041153326630592,
0.08338344842195511,
0.0005585690960288048,
-0.059273574501276016,
-0.010698637925088406,
0.021649496629834175,
0.08753470331430435,
0.03604402393102646,
0.01953037828207016,
0.025459563359618187,
0.04416375234723091,
-0.043410103768110275,
0.03769073635339737,
-0.04267771542072296,
-0.014993920922279358,
0.10641276091337204,
-0.013101629912853241,
-0.03410772606730461,
-0.054627757519483566,
-0.004941944964230061,
-0.08743254095315933,
-0.09055529534816742,
0.030786290764808655,
-0.10158386826515198,
-0.026026884093880653,
-0.0026559948455542326,
-0.053786568343639374,
0.047397855669260025,
0.10194537788629532,
0.01807672344148159,
0.011918175034224987,
-0.09324800223112106,
0.02459966391324997,
0.08658517897129059,
-0.027803659439086914,
-0.07345547527074814,
-0.06717783212661743,
0.022975368425250053
] | 0 | 0 | 0 | 0 |
|
A Meta-Learning Approach to Bayesian Causal Discovery | https://openreview.net/forum?id=eeJz7eDWKO | [
"Anish Dhir",
"Matthew Ashman",
"James Requeima",
"Mark van der Wilk"
] | Poster | Discovering a unique causal structure is difficult due to both inherent identifiability issues, and the consequences of finite data.
As such, uncertainty over causal structures, such as those obtained from a Bayesian posterior, are often necessary for downstream tasks.
Finding an accurate approximation to this posterior is challenging, due to the large number of possible causal graphs, as well as the difficulty in the subproblem of finding posteriors over the functional relationships of the causal edges.
Recent works have used Bayesian meta learning to view the problem of posterior estimation as a supervised learning task.
Yet, these methods are limited as they cannot reliably sample from the posterior over causal structures and fail to encode key properties of the posterior, such as correlation between edges and permutation equivariance with respect to nodes.
To address these limitations, we propose a Bayesian meta learning model that allows for sampling causal structures from the posterior and encodes these key properties.
We compare our meta-Bayesian causal discovery against existing Bayesian causal discovery methods, demonstrating the advantages of directly learning a posterior over causal structure. | neural processes, bayesian causal discovery, transformers | We propose a method that allows for sampling from an approximate posterior over causal structures using neural processes. | 12,200 | 2412.16577 | [
-0.05478589981794357,
-0.04459083452820778,
0.05301499739289284,
0.019810518249869347,
0.06257659196853638,
-0.007568931672722101,
0.011769895441830158,
-0.05153067409992218,
-0.017211057245731354,
0.003366272198036313,
-0.04609926789999008,
-0.016672922298312187,
0.054381392896175385,
0.05531834810972214,
-0.06180914491415024,
0.0056159961968660355,
0.06563442200422287,
0.08204194903373718,
0.00266471691429615,
-0.03816013038158417,
0.004982778802514076,
-0.005589578300714493,
-0.00007183281559264287,
-0.005341191776096821,
-0.006698968354612589,
0.02971974015235901,
0.031736940145492554,
0.04125296324491501,
-0.018218573182821274,
0.006016690284013748,
0.05223263055086136,
-0.006465583574026823,
-0.13730137050151825,
-0.08647606521844864,
-0.004413633141666651,
0.010192090645432472,
0.011387587524950504,
0.04245038330554962,
0.0070092030800879,
0.01494289468973875,
0.0546889491379261,
0.02414674684405327,
0.03405001014471054,
0.06082247942686081,
0.042220793664455414,
0.02469507046043873,
0.007192214950919151,
-0.012816663831472397,
0.03435182943940163,
0.016318149864673615,
-0.11680322885513306,
-0.028974901884794235,
-0.0419219508767128,
0.03713347762823105,
0.08156054466962814,
0.014297401532530785,
0.02933286689221859,
-0.10365471243858337,
-0.003944893833249807,
0.043963875621557236,
-0.03652607277035713,
-0.08910834789276123,
-0.05723319575190544,
-0.04051227122545242,
-0.0075331926345825195,
0.09233475476503372,
-0.01924268901348114,
0.07555379718542099,
-0.06559664011001587,
0.048733871430158615,
0.03710746020078659,
0.09765841066837311,
-0.10829680413007736,
0.07999961823225021,
0.03943854570388794,
0.02541244588792324,
0.0019432398257777095,
0.01241668127477169,
-0.0007547063287347555,
-0.06565801799297333,
-0.03346904367208481,
0.0779227465391159,
-0.02031162567436695,
-0.0007675616070628166,
0.11227454245090485,
-0.02291688695549965,
0.012533199042081833,
0.04226501286029816,
0.01556502003222704,
0.12566189467906952,
-0.041618283838033676,
-0.04539698362350464,
0.020061882212758064,
-0.05201440677046776,
0.001023861113935709,
0.058111030608415604,
0.005216124467551708,
-0.028383325785398483,
0.04785928130149841,
0.08307992666959763,
-0.013290571980178356,
0.054414670914411545,
0.007033094763755798,
-0.006143449805676937,
-0.06534795463085175,
0.02292851358652115,
-0.0879751518368721,
0.07561490684747696,
0.018116338178515434,
-0.08539484441280365,
-0.04797925055027008,
-0.011138630099594593,
0.014175541698932648,
-0.020072713494300842,
0.005539272911846638,
-0.024005550891160965,
0.05254718288779259,
-0.0311546940356493,
0.035501886159181595,
0.0551900751888752,
-0.0045256419107317924,
0.020288566127419472,
-0.0209028460085392,
0.05032346025109291,
0.015821406617760658,
-0.08186232298612595,
-0.08568095415830612,
1.707580766696042e-33,
0.13444438576698303,
0.016284871846437454,
0.10017634183168411,
0.0532601922750473,
-0.003865815233439207,
0.06523244827985764,
-0.05801040306687355,
-0.005838900338858366,
0.0501556470990181,
-0.002714070025831461,
-0.025632748380303383,
-0.025754626840353012,
-0.0878906324505806,
-0.07552997767925262,
0.04611297324299812,
0.016321294009685516,
-0.046140849590301514,
0.07303273677825928,
-0.012011268176138401,
-0.06934578716754913,
0.036127351224422455,
-0.07858941704034805,
0.00758154783397913,
-0.018433086574077606,
-0.0030175824649631977,
0.04717585816979408,
0.02553173340857029,
0.008690631948411465,
-0.021838944405317307,
0.02950194850564003,
-0.05660737678408623,
-0.036382775753736496,
0.05920574069023132,
0.08811477571725845,
-0.031513918191194534,
-0.05308394134044647,
-0.08861342817544937,
-0.042665544897317886,
-0.03965632990002632,
-0.030266473069787025,
0.017630072310566902,
-0.0162736177444458,
0.0668458640575409,
0.013292853720486164,
-0.08484940975904465,
-0.1109980046749115,
-0.03802934288978577,
-0.06203247234225273,
0.023961219936609268,
-0.016733800992369652,
0.062446724623441696,
-0.016128994524478912,
0.00131038052495569,
-0.10221783071756363,
-0.053518760949373245,
0.06545592844486237,
-0.0377003476023674,
0.04013223946094513,
-0.02835647203028202,
-0.023572346195578575,
0.05106765776872635,
0.032064929604530334,
-0.05873788893222809,
0.08017148822546005,
0.009298496879637241,
0.15808936953544617,
-0.04078775271773338,
-0.008693622425198555,
0.044319797307252884,
-0.01949976198375225,
-0.026467103511095047,
0.03692174702882767,
-0.008788586594164371,
-0.0427284799516201,
0.04036368429660797,
0.030213158577680588,
-0.03786398097872734,
-0.1038077250123024,
0.0068394192494452,
0.02684047445654869,
-0.09177614748477936,
-0.08027996867895126,
-0.04290354996919632,
0.08259893208742142,
-0.009737377986311913,
0.00010375748388469219,
0.027317766100168228,
-0.03419388085603714,
0.020280281081795692,
-0.043146323412656784,
-0.0028521381318569183,
0.022060304880142212,
-0.0348774679005146,
0.017973095178604126,
-0.0001338808360742405,
-2.9129361637780697e-33,
-0.07581474632024765,
0.01219523698091507,
0.1318960040807724,
-0.02440710924565792,
0.004742931108921766,
-0.05032312497496605,
-0.01782374642789364,
-0.02491636388003826,
-0.022654786705970764,
-0.05724772810935974,
-0.06956852972507477,
-0.0540972463786602,
-0.0281614288687706,
-0.03908780962228775,
-0.020865095779299736,
0.07632556557655334,
0.03125083073973656,
-0.017708750441670418,
-0.04047293961048126,
-0.008146457374095917,
-0.007553659379482269,
0.01582186296582222,
-0.004325711168348789,
0.03359236568212509,
0.005800548940896988,
-0.005053539294749498,
0.06515923142433167,
0.06220278516411781,
0.0331011600792408,
0.04922179505228996,
-0.03606267645955086,
0.016608141362667084,
-0.0709930807352066,
-0.07464955002069473,
-0.0940353125333786,
0.10054926574230194,
0.020528698340058327,
0.02760917693376541,
-0.025215093046426773,
0.012175135314464569,
0.05519985407590866,
0.01690354384481907,
-0.08534366637468338,
-0.04257062077522278,
-0.010208317078649998,
-0.014055110514163971,
-0.02677347883582115,
0.08377707749605179,
0.03284689411520958,
-0.028549799695611,
-0.05340394005179405,
0.0821545422077179,
-0.03259330987930298,
0.013841206207871437,
-0.1078680232167244,
-0.07888864725828171,
0.04164878651499748,
-0.012127146124839783,
0.12257211655378342,
0.00009429780766367912,
-0.06201577186584473,
-0.009059370495378971,
-0.013716282323002815,
0.06421992182731628,
-0.009029269218444824,
-0.027495000511407852,
0.007319389842450619,
-0.044177036732435226,
-0.007451488170772791,
0.04226860776543617,
0.030687501654028893,
0.0072176032699644566,
-0.05964040756225586,
-0.0780944973230362,
0.05524603649973869,
-0.00015786770381964743,
-0.03692977875471115,
0.04189355671405792,
-0.09355217218399048,
0.01195030938833952,
0.10981229692697525,
0.011199128814041615,
0.008264046162366867,
-0.00020228966604918242,
0.02358989790081978,
0.02049526758491993,
0.027950556948781013,
0.049141500145196915,
0.058835942298173904,
-0.019363533705472946,
-0.09196696430444717,
0.022245770320296288,
-0.08759177476167679,
0.0065989005379378796,
-0.08978894352912903,
-3.517301649935689e-8,
-0.11461309343576431,
0.022494670003652573,
0.014023044146597385,
0.05400091037154198,
0.1103668287396431,
0.020927077159285545,
-0.04525294154882431,
0.12336864322423935,
0.026183957234025,
-0.052916523069143295,
0.07790643721818924,
0.0014392725424841046,
0.01228275615721941,
-0.029160242527723312,
-0.022974159568548203,
0.03314444422721863,
0.052328921854496,
0.0011182540329173207,
-0.007496532052755356,
0.024855080991983414,
0.04177720844745636,
0.004166147205978632,
0.00890440121293068,
0.014250122010707855,
0.06808148324489594,
-0.15395201742649078,
-0.050682712346315384,
0.04305851832032204,
0.036573536694049835,
0.08148503303527832,
-0.0550541952252388,
-0.010672731325030327,
0.013290418311953545,
-0.0048267836682498455,
0.009094365872442722,
0.03607132285833359,
0.01980637013912201,
-0.06999196857213974,
-0.06136454641819,
-0.10527418553829193,
0.018765414133667946,
0.06428921967744827,
-0.05401937663555145,
0.05322163179516792,
0.08398773521184921,
0.034653592854738235,
0.01568657159805298,
0.02070792391896248,
-0.02993888221681118,
-0.001515153213404119,
0.05811939015984535,
-0.02352325990796089,
0.011658754199743271,
-0.03519371151924133,
0.001406569848768413,
0.052857428789138794,
0.05883258953690529,
-0.0402248315513134,
-0.0111469104886055,
-0.028094643726944923,
0.05176578834652901,
0.06920979917049408,
-0.025137463584542274,
0.026278238743543625
] | 0 | 0 | 0 | 0 |
|
Is Your Video Language Model a Reliable Judge? | https://openreview.net/forum?id=m8yby1JfbU | [
"Ming Liu",
"Wensheng Zhang"
] | Poster | As video language models (VLMs) gain more applications in various scenarios,
the need for robust and scalable evaluation of their performance becomes increasingly critical. The traditional human expert-based evaluation of VLMs has limitations in consistency and scalability, which sparked interest in automatic methods
such as employing VLMs to evaluate VLMs. However, the reliability of VLMs as
judges remains underexplored. Existing methods often rely on a single VLM as
the evaluator. However, this approach can be unreliable or biased because such a
model may lack the ability to fully understand the content and may have inherent
biases, ultimately compromising evaluation reliability. A remedy is to apply the
principle of collective thoughts, aggregating evaluations from multiple VLMs to
enhance reliability. This study investigates the efficacy of such approaches, particularly when the pool of judges includes both reliable and unreliable models. Our
findings reveal that incorporating collective judgments from such a mixed pool
does not necessarily improve the accuracy of the final evaluation. The inclusion of
less reliable judges can introduce noise, undermining the overall reliability of the
outcomes. To explore the factors that impact evaluation reliability, we fine-tune
an underperforming VLM judge, Video-LLaVA, and observe that improved understanding ability alone is insufficient to make VLM judges more reliable. These
findings stress the limitations of collective thought approaches and highlight the
need for more advanced methods that can account for the reliability of individual
models. Our study promotes the development of more reliable evaluation methods
for VLMs | Video Language Models, Model evaluation, Reliability | null | 12,195 | 2503.05977 | [
0.020023871213197708,
-0.05760335177183151,
-0.06391074508428574,
-0.0639515146613121,
0.06667520850896835,
0.014438405632972717,
0.03045494854450226,
0.030700596049427986,
0.05820086970925331,
-0.08323880285024643,
-0.06733614206314087,
-0.06665704399347305,
0.06479278951883316,
0.06987453997135162,
-0.04637591913342476,
-0.018760815262794495,
0.13863086700439453,
0.04738212376832962,
-0.12703490257263184,
-0.05774034187197685,
-0.019794223830103874,
0.010253469459712505,
0.0009608704131096601,
0.013062459416687489,
-0.02733362652361393,
-0.04519747570157051,
-0.10772289335727692,
-0.022415831685066223,
-0.0015819960972294211,
-0.05839025229215622,
-0.006723988335579634,
-0.014994654804468155,
0.021693553775548935,
0.01877078413963318,
-0.02801787480711937,
0.023683171719312668,
-0.06147924065589905,
-0.007494146469980478,
-0.0247474517673254,
-0.02905937097966671,
-0.0467403419315815,
-0.019060159102082253,
0.04724453389644623,
-0.041137196123600006,
0.010554293170571327,
-0.03374672308564186,
0.015636833384633064,
-0.014116361737251282,
-0.04652006924152374,
0.03295260667800903,
-0.13732708990573883,
-0.01635981723666191,
0.09654785692691803,
-0.021820195019245148,
-0.06746633350849152,
0.0351107232272625,
0.047682397067546844,
0.04169171303510666,
0.017252866178750992,
-0.039420705288648605,
0.004663391970098019,
-0.08295426517724991,
-0.06087080016732216,
0.05495501309633255,
0.036211028695106506,
-0.011565657332539558,
0.051114581525325775,
-0.03044046461582184,
-0.02100808173418045,
0.015164944343268871,
-0.05492132529616356,
0.04955592378973961,
0.05175314098596573,
0.06222619488835335,
-0.055044662207365036,
0.04003554955124855,
0.0038740935269743204,
-0.0818253755569458,
0.07315623015165329,
-0.038729287683963776,
0.026536496356129646,
-0.012273158878087997,
-0.005120907444506884,
-0.03165894374251366,
0.15242120623588562,
-0.014088493771851063,
0.032256077975034714,
0.01046664733439684,
-0.06229693442583084,
0.006721675395965576,
-0.10163313895463943,
0.035944242030382156,
0.06989872455596924,
-0.02153109759092331,
0.11467063426971436,
0.05832938477396965,
-0.008240333758294582,
-0.08862000703811646,
0.054645001888275146,
0.08007269352674484,
0.030823053792119026,
0.05881989002227783,
0.01854875683784485,
-0.09560617059469223,
-0.007059340365231037,
-0.012469256296753883,
0.020087381824851036,
0.08386439085006714,
-0.01132621057331562,
-0.06554152816534042,
-0.09097728878259659,
0.08249562233686447,
0.002118387958034873,
-0.01052241399884224,
0.0426609106361866,
0.02599073201417923,
0.005078241694718599,
0.04846024513244629,
-0.013993547298014164,
0.02030998095870018,
0.011985134333372116,
0.013123084791004658,
0.10771898180246353,
-0.033668652176856995,
0.12351110577583313,
0.016864217817783356,
0.05332862213253975,
5.337603198293255e-33,
0.01723482459783554,
0.0020912857726216316,
0.019471202045679092,
0.04889018461108208,
0.04434610903263092,
-0.02373095415532589,
-0.06509104371070862,
-0.004791403189301491,
-0.03740367665886879,
-0.013849331066012383,
0.0035609034821391106,
-0.007044869940727949,
-0.00929174479097128,
0.00017256241699215025,
0.021306224167346954,
0.12447957694530487,
-0.0838007926940918,
0.03954657167196274,
-0.06781362742185593,
0.039767056703567505,
0.10691609233617783,
-0.011719265021383762,
0.04017483443021774,
-0.005707248114049435,
-0.06350164115428925,
-0.03364795818924904,
0.08660183846950531,
-0.01859789714217186,
0.01845860667526722,
0.01563917100429535,
-0.13853143155574799,
-0.03181648999452591,
0.033520594239234924,
0.02545296773314476,
0.07651050388813019,
-0.01780141517519951,
-0.07129070907831192,
0.0119314631447196,
0.039129648357629776,
-0.00010571616439847276,
-0.027149979025125504,
0.01836407743394375,
0.009255382232367992,
-0.004960410296916962,
-0.050317373126745224,
0.055333055555820465,
0.03547393158078194,
-0.012488334439694881,
-0.12084519863128662,
0.0009177331812679768,
0.01541562844067812,
0.02070341445505619,
0.08088444918394089,
0.005504633765667677,
-0.04247364029288292,
0.04510239139199257,
0.050312452018260956,
0.03172243759036064,
-0.011464154347777367,
-0.04180574044585228,
-0.02742845192551613,
0.10297296941280365,
-0.06385190784931183,
-0.01606127992272377,
0.04332723096013069,
0.0027621272020041943,
-0.018295161426067352,
0.00043910302338190377,
0.07374287396669388,
-0.0642583891749382,
-0.020834995433688164,
0.0389506071805954,
0.008029049262404442,
-0.034097738564014435,
-0.015531015582382679,
-0.030019694939255714,
-0.047605447471141815,
-0.008770729415118694,
0.034850653260946274,
-0.027040740475058556,
0.04848760738968849,
-0.03240485489368439,
-0.0018206302775070071,
-0.05656641721725464,
0.0022642763797193766,
-0.03341158106923103,
-0.011132458224892616,
-0.04622117057442665,
0.004755796864628792,
0.018382063135504723,
-0.053742796182632446,
-0.0496089830994606,
0.07189780473709106,
-0.03292032703757286,
0.05403122678399086,
-5.2130893265243886e-33,
-0.06584453582763672,
0.02762790396809578,
0.03972117602825165,
0.11559370160102844,
0.007254173047840595,
-0.012457258068025112,
-0.008526005782186985,
-0.03477659076452255,
-0.027579667046666145,
-0.07595311850309372,
-0.08051928877830505,
-0.1029769629240036,
0.0062072803266346455,
0.02485129050910473,
-0.009137800894677639,
-0.05567765608429909,
-0.05280499905347824,
-0.05591480806469917,
0.04725692421197891,
0.04024513438344002,
0.0945667251944542,
0.04883408173918724,
-0.07531663030385971,
0.043075595051050186,
-0.0629807859659195,
0.02938620001077652,
-0.058597978204488754,
-0.039471160620450974,
0.04365719482302666,
-0.0031720888800919056,
0.03358897194266319,
-0.053090937435626984,
-0.038293320685625076,
-0.011106924153864384,
0.03928298503160477,
-0.01165184285491705,
0.12262088805437088,
-0.055508412420749664,
-0.011409623548388481,
0.10826410353183746,
-0.03662853688001633,
0.026469599455595016,
-0.1159760057926178,
-0.11106348037719727,
0.0021565500646829605,
-0.018774498254060745,
-0.07280585169792175,
-0.06271017342805862,
-0.1060909852385521,
-0.009154264815151691,
-0.08692353218793869,
0.02655024826526642,
-0.06864813715219498,
0.0038838370237499475,
-0.02930191531777382,
-0.10736334323883057,
0.02621517702937126,
-0.026868486776947975,
0.0335509330034256,
0.04488418623805046,
-0.08512203395366669,
-0.0002656347642187029,
-0.011882292106747627,
-0.0344531387090683,
0.029450327157974243,
0.010570484213531017,
0.009009672328829765,
0.025081954896450043,
0.0149155855178833,
0.039328575134277344,
0.017484210431575775,
-0.0406000092625618,
-0.016307834535837173,
0.012587231583893299,
0.10867868363857269,
0.03809479624032974,
0.03127061203122139,
-0.014743940904736519,
0.0027295364998281,
-0.07693807780742645,
-0.017024844884872437,
-0.09035400301218033,
0.07780896872282028,
-0.011333635076880455,
0.02948356792330742,
0.10710489004850388,
0.09776704013347626,
0.012996966950595379,
-0.04141787812113762,
0.048761095851659775,
-0.0036868141032755375,
0.009791048243641853,
0.004666053224354982,
0.05209021642804146,
-0.037872206419706345,
-5.194053898094353e-8,
-0.07099092751741409,
-0.00021909190400037915,
0.002546423114836216,
0.0742255449295044,
-0.02599278837442398,
-0.006348986644297838,
-0.018824806436896324,
-0.003727790666744113,
0.017062388360500336,
-0.025633955374360085,
0.025540849193930626,
-0.025329725816845894,
-0.035498395562171936,
0.005214029923081398,
-0.014358827844262123,
0.013685330748558044,
0.0016661203699186444,
0.09652185440063477,
-0.02772660180926323,
0.002144861500710249,
0.06842164695262909,
0.04084796831011772,
0.00997152365744114,
-0.0364961139857769,
0.011990508064627647,
0.0022610214073210955,
-0.015525996685028076,
0.02719756029546261,
-0.019665436819195747,
-0.01617109775543213,
0.031894370913505554,
0.09332981705665588,
-0.01979043520987034,
-0.057286497205495834,
0.037550635635852814,
0.007414383348077536,
-0.019902067258954048,
0.0050154984928667545,
0.04806068167090416,
0.04212731868028641,
0.04307636618614197,
0.11678440123796463,
0.003705717623233795,
0.04051796346902847,
0.05984169989824295,
-0.005852194037288427,
-0.045888982713222504,
-0.09102962166070938,
-0.00954684428870678,
-0.008446170017123222,
-0.029323862865567207,
-0.06412339210510254,
-0.09713482111692429,
0.11093602329492569,
0.09404729306697845,
0.008800843730568886,
0.06981971859931946,
-0.01815883256494999,
0.004315056838095188,
0.056475210934877396,
0.038881611078977585,
0.037793293595314026,
-0.0012287040008231997,
-0.0038318305741995573
] | 0 | 0 | 0 | 0 |
|
Boosting Multiple Views for pretrained-based Continual Learning | https://openreview.net/forum?id=AZR4R3lw7y | [
"Quyen Tran",
"Tung Lam Tran",
"Khanh Doan",
"Toan Tran",
"Dinh Phung",
"Khoat Than",
"Trung Le"
] | Poster | Recent research has shown that Random Projection (RP) can effectively improve the performance of pre-trained models in Continual learning (CL). The authors hypothesized that using RP to map features onto a higher-dimensional space can make them more linearly separable. In this work, we theoretically analyze the role of RP and present its benefits for improving the model’s generalization ability
in each task and facilitating CL overall. Additionally, we take this result to the next level by proposing a Multi-View Random Projection scheme for a stronger ensemble classifier. In particular, we train a set of linear experts, among which diversity is encouraged based on the principle of AdaBoost, which was initially very challenging to apply to CL. Moreover, we employ a task-based adaptive backbone
with distinct prompts dedicated to each task for better representation learning. To properly select these task-specific components and mitigate potential feature shifts caused by misprediction, we introduce a simple yet effective technique called the self-improvement process. Experimentally, our method consistently outperforms state-of-the-art baselines across a wide range of datasets. | continual learning, ViT-pretrained continual learning | null | 12,194 | null | [
-0.015180161222815514,
-0.11022886633872986,
0.016415344551205635,
0.04175148159265518,
0.08943384885787964,
0.060888588428497314,
0.0013377139111980796,
-0.09579835087060928,
-0.003658761503174901,
-0.02995019592344761,
-0.09618058055639267,
-0.01688351109623909,
0.022410087287425995,
0.04335048794746399,
-0.07543391734361649,
0.0016764283645898104,
0.10678628832101822,
0.036806076765060425,
-0.036346435546875,
0.024652624502778053,
-0.025058196857571602,
0.009537753649055958,
0.033711280673742294,
0.07666368782520294,
0.07982881367206573,
-0.021478217095136642,
0.008386255241930485,
-0.07507109642028809,
0.030610116198658943,
-0.0475575290620327,
0.05946711078286171,
-0.008980123326182365,
-0.008409039117395878,
-0.0024493481032550335,
-0.09130465239286423,
0.0002292661665705964,
-0.07743887603282928,
0.09611763805150986,
0.00318576255813241,
-0.01009522844105959,
-0.003061434021219611,
0.03312747925519943,
-0.00869558285921812,
0.03629368916153908,
0.037402406334877014,
-0.005410701967775822,
0.03289709985256195,
-0.10593307018280029,
0.08205582946538925,
0.025802871212363243,
-0.07374948263168335,
-0.03517021983861923,
-0.035517387092113495,
0.018427535891532898,
-0.018965793773531914,
0.0077629173174500465,
-0.006872252561151981,
0.03551064059138298,
-0.01913820020854473,
0.008470340631902218,
0.07734416425228119,
-0.04240201413631439,
0.011739850975573063,
0.011681489646434784,
-0.023196915164589882,
0.02682299166917801,
0.04010443389415741,
0.09685513377189636,
0.017617933452129364,
0.05513476952910423,
-0.03332418575882912,
0.10416413843631744,
0.002459808485582471,
0.011053000576794147,
0.09999492764472961,
0.05167336016893387,
0.030183952301740646,
-0.03580033779144287,
0.12107200175523758,
-0.03143743798136711,
-0.03187830373644829,
0.0615105964243412,
0.033959124237298965,
-0.03361940756440163,
0.15550047159194946,
0.009187729097902775,
0.023172883316874504,
-0.03716078773140907,
-0.007985029369592667,
-0.07807100564241409,
-0.0037654140032827854,
-0.04232435300946236,
0.03906448185443878,
-0.021862443536520004,
-0.08342011272907257,
-0.017099620774388313,
0.008119214326143265,
-0.1038585677742958,
0.019619066268205643,
0.0859488844871521,
-0.03220740705728531,
-0.014318904839456081,
-0.0073956395499408245,
-0.04545603320002556,
-0.05708867684006691,
-0.09009236097335815,
-0.0232380498200655,
0.06912187486886978,
0.04734698310494423,
-0.01950761117041111,
0.061009928584098816,
0.007882270030677319,
-0.05956970900297165,
0.07072383165359497,
0.02682177722454071,
0.01900634914636612,
-0.034945495426654816,
0.0058265491388738155,
0.01916363276541233,
0.10712878406047821,
-0.03132127597928047,
0.03354144096374512,
-0.0052267881110310555,
0.027016036212444305,
0.022755445912480354,
0.0024380588438361883,
-0.008792479522526264,
1.636018874669889e-33,
0.043306492269039154,
-0.005956859327852726,
0.053759049624204636,
0.02905523031949997,
0.09783094376325607,
-0.015572018921375275,
0.007099005859345198,
0.07771074771881104,
0.0003908138023689389,
0.011327682994306087,
0.02938503585755825,
-0.007465837523341179,
0.02017856575548649,
0.11227020621299744,
0.02347465418279171,
-0.005787791218608618,
-0.06638351082801819,
0.051053158938884735,
-0.08006730675697327,
-0.0606934055685997,
0.023903660476207733,
0.03660363331437111,
-0.017470017075538635,
-0.11881639808416367,
-0.0004842974303755909,
-0.008489539846777916,
0.0465555265545845,
0.0010134269250556827,
-0.09540366381406784,
0.04092229902744293,
-0.09870577603578568,
-0.03597898408770561,
-0.01626957394182682,
0.0489908792078495,
-0.020868800580501556,
-0.06628956645727158,
0.0011080914409831166,
-0.040515270084142685,
0.027890043333172798,
-0.021759070456027985,
0.010746382176876068,
0.010042608715593815,
-0.022840455174446106,
0.004385102540254593,
-0.01539689116179943,
-0.04645870253443718,
0.018158715218305588,
0.06817049533128738,
-0.12175240367650986,
-0.013867266476154327,
0.040425412356853485,
-0.014052322134375572,
-0.056063443422317505,
0.011471102014183998,
-0.10008635371923447,
0.001025170786306262,
-0.05429820716381073,
0.02093864232301712,
0.09062499552965164,
-0.01872927136719227,
-0.04678376764059067,
0.07009762525558472,
-0.05229513719677925,
0.11207260191440582,
-0.0003507670189719647,
0.03575397655367851,
0.0772460550069809,
0.0707995593547821,
0.043783850967884064,
0.008501030504703522,
0.003978199791163206,
-0.045342620462179184,
-0.004218554124236107,
-0.05981290712952614,
0.040444690734148026,
-0.02933366969227791,
0.059883713722229004,
-0.1002059131860733,
0.024293432012200356,
0.07846143841743469,
0.001746815163642168,
0.011535336263477802,
0.008055961690843105,
-0.08260755985975266,
-0.037923235446214676,
-0.047013621777296066,
0.06577354669570923,
-0.0914372056722641,
-0.00485928263515234,
0.053902518004179,
-0.059747256338596344,
-0.036168232560157776,
-0.025908835232257843,
-0.05345238372683525,
-0.02393876202404499,
-1.9442902477921184e-33,
0.04185641184449196,
-0.0020614848472177982,
0.009823865257203579,
-0.026089806109666824,
0.02360026352107525,
-0.03746515512466431,
-0.02566506154835224,
0.02213011123239994,
-0.08278506994247437,
-0.13448946177959442,
-0.01969335600733757,
-0.09639724344015121,
0.040835462510585785,
0.06174061447381973,
0.00920303724706173,
0.025531474500894547,
-0.040797099471092224,
0.023490941151976585,
-0.056730691343545914,
0.03127651289105415,
-0.007322970777750015,
0.05225808918476105,
-0.06332390010356903,
0.015485464595258236,
0.0069197542034089565,
0.03640271723270416,
0.004876034799963236,
0.10175644606351852,
0.03423982858657837,
-0.020802002400159836,
-0.062344592064619064,
0.006804245058447123,
-0.0426250658929348,
-0.10690318793058395,
-0.09242819994688034,
0.09700606018304825,
-0.02953936532139778,
-0.03700623661279678,
-0.03214096277952194,
0.12667250633239746,
0.020609183236956596,
-0.07988777756690979,
0.015366006642580032,
0.006879815366119146,
0.010714051313698292,
-0.0617036372423172,
-0.05114566534757614,
0.031893324106931686,
0.029209254309535027,
0.06219024211168289,
-0.05283404886722565,
0.002818977925926447,
-0.1450987458229065,
0.03695448860526085,
-0.07823912054300308,
-0.06712330877780914,
-0.07215666025876999,
-0.015656255185604095,
0.024913089349865913,
0.005758361425250769,
-0.032558053731918335,
-0.053435783833265305,
0.05650938302278519,
-0.015809282660484314,
0.03638821467757225,
-0.05725226178765297,
-0.0005136788240633905,
0.009869886562228203,
0.030611250549554825,
0.02971162088215351,
0.011327874846756458,
0.01300053484737873,
-0.04688237980008125,
-0.014170806854963303,
-0.026480427011847496,
-0.00871036946773529,
0.02149873599410057,
-0.030547378584742546,
0.012772878631949425,
0.0006288072327151895,
-0.0010228652972728014,
-0.02716744691133499,
0.0321512445807457,
0.07880812883377075,
0.08432977646589279,
0.12360431998968124,
0.015584881417453289,
0.09131962805986404,
0.07626509666442871,
-0.0022196348290890455,
0.00804123468697071,
0.004041377454996109,
-0.03598165884613991,
0.033981844782829285,
0.011662282980978489,
-4.718741308806784e-8,
-0.05543119087815285,
0.09361582249403,
0.01880703493952751,
0.04683252424001694,
0.007216454949229956,
-0.03087618574500084,
-0.03388550877571106,
0.09072893112897873,
-0.07747823745012283,
-0.05142470449209213,
0.08873219043016434,
-0.01720668561756611,
-0.0247129388153553,
-0.03493400290608406,
0.02874159626662731,
0.013895835727453232,
0.006518068723380566,
0.028157958760857582,
-0.0412147119641304,
-0.06134311109781265,
0.03650578111410141,
-0.029995132237672806,
0.053139738738536835,
-0.02760913223028183,
0.003116385079920292,
-0.0452144555747509,
0.007829507812857628,
0.0665210410952568,
0.0760197713971138,
-0.02891390770673752,
-0.020915841683745384,
0.035649556666612625,
-0.003323120065033436,
-0.01883559674024582,
0.04263724386692047,
0.06949015706777573,
0.042886439710855484,
-0.009061898104846478,
-0.007921434007585049,
-0.011315524578094482,
-0.043102163821458817,
0.026749899610877037,
0.004710735287517309,
0.0496961809694767,
-0.0278562493622303,
0.0176771841943264,
0.02589000202715397,
-0.07277975976467133,
0.007528777699917555,
-0.07801496982574463,
-0.0065771047957241535,
-0.06345803290605545,
-0.04687291383743286,
0.10933782160282135,
0.07723008841276169,
0.07314425706863403,
-0.02865980565547943,
-0.06594091653823853,
-0.021173540502786636,
0.05173668637871742,
0.04325489699840546,
0.012314530089497566,
-0.09794311970472336,
0.01302339881658554
] | 0 | 0 | 0 | 0 |
|
Matrix Product Sketching via Coordinated Sampling | https://openreview.net/forum?id=eHfq8Q3LeD | [
"Majid Daliri",
"Juliana Freire",
"Danrong Li",
"Christopher Musco"
] | Poster | We revisit the well-studied problem of approximating a matrix product, $\bv{A}^T\bv{B}$, based on small space sketches $\mathcal{S}(\bv{A})$ and $\mathcal{S}(\bv{B})$ of $\bv{A} \in \R^{n \times d}$ and $\bv{B}\in \R^{n \times m}$. We are interested in the setting where the sketches must be computed independently of each other, except for the use of a shared random seed. We prove that, when $\bv{A}$ and $\bv{B}$ are sparse, methods based on \emph{coordinated random sampling} can outperform classical linear sketching approaches, like Johnson-Lindenstrauss Projection or CountSketch. For example, to obtain Frobenius norm error $\epsilon\|\bv{A}\|_F\|\bv{B}\|_F$, coordinated sampling requires sketches of size $O(s/\epsilon^2)$ when $\bv{A}$ and $\bv{B}$ have at most $s \leq d,m$ non-zeros per row. In contrast, linear sketching leads to sketches of size $O(d/\epsilon^2)$ and $O(m/\epsilon^2)$ for $\bv{A}$ and $\bv{B}$. We empirically evaluate our approach on two applications: 1) distributed linear regression in databases, a problem motivated by tasks like dataset discovery and augmentation, and 2) approximating attention matrices in transformer-based language models. In both cases, our sampling algorithms yield an order of magnitude improvement over linear sketching. | Sketching Algorithm, Matrix Multiplication, Model Compression, Data Discovery, Efficient resource | null | 12,181 | 2501.17836 | [
-0.08180603384971619,
-0.0622260756790638,
-0.07894695550203323,
-0.07523982226848602,
0.06829136610031128,
0.037403564900159836,
-0.018461205065250397,
-0.05327459052205086,
0.03483109548687935,
0.03282136097550392,
-0.047308698296546936,
0.05257926881313324,
0.08569688349962234,
0.018566520884633064,
0.016759483143687248,
0.02627095952630043,
0.05414224788546562,
0.03566356375813484,
-0.0453895628452301,
0.061553020030260086,
-0.004480039235204458,
-0.04902026802301407,
0.004916312173008919,
-0.0656692385673523,
0.09197767823934555,
-0.07038138061761856,
0.04472833126783371,
-0.04061732441186905,
0.009135757572948933,
0.004886581562459469,
0.0408044308423996,
0.10194878280162811,
0.010235488414764404,
-0.07962843030691147,
0.06336598098278046,
0.026668062433600426,
0.002465310040861368,
0.11838926374912262,
-0.05010432004928589,
0.06756118685007095,
0.009376325644552708,
0.07669602334499359,
-0.03481302410364151,
0.03453604876995087,
0.03822728618979454,
-0.01497966144233942,
0.02894039824604988,
-0.022090133279561996,
-0.020454617217183113,
-0.0009765385766513646,
-0.07396289706230164,
0.014126108959317207,
-0.051582127809524536,
-0.061822135001420975,
-0.02845980040729046,
-0.12749676406383514,
0.041498273611068726,
-0.036371055990457535,
-0.003997631371021271,
-0.013985052704811096,
0.04294370114803314,
-0.09191504120826721,
-0.009208094328641891,
-0.022172389551997185,
0.012780810706317425,
0.014822978526353836,
0.11130491644144058,
0.08587463200092316,
0.009103267453610897,
0.033402130007743835,
-0.03936447203159332,
0.034539662301540375,
-0.056460633873939514,
0.0747581347823143,
0.062194500118494034,
0.031707171350717545,
0.03341657295823097,
-0.008917871862649918,
0.014335691928863525,
-0.0901007428765297,
-0.04565917328000069,
0.066916823387146,
-0.025244232267141342,
-0.05254453048110008,
-0.007206744514405727,
-0.013438803143799305,
-0.012438841164112091,
0.05173031985759735,
0.0994824469089508,
-0.031689368188381195,
-0.05751258134841919,
0.018683673813939095,
-0.07733434438705444,
-0.039200812578201294,
-0.024649985134601593,
-0.08040080219507217,
0.07934127748012543,
0.0006518270238302648,
0.07415712624788284,
0.06847241520881653,
0.02656424231827259,
-0.03793486952781677,
0.042800985276699066,
0.01744983345270157,
0.01808006875216961,
-0.014882095158100128,
0.026464147493243217,
0.0453953742980957,
0.08086322993040085,
-0.12420880049467087,
0.0890984907746315,
0.023638397455215454,
-0.0018130516400560737,
0.09197049587965012,
0.041006945073604584,
-0.04458553344011307,
0.09857483953237534,
-0.06273722648620605,
0.09917376190423965,
-0.008291208185255527,
-0.05669277906417847,
0.023977968841791153,
0.10539774596691132,
-0.05864603817462921,
-0.011763223446905613,
0.05435696244239807,
-0.051640477031469345,
1.4109869197587743e-33,
0.06684704124927521,
0.05622786656022072,
0.03681278973817825,
0.04267050698399544,
0.035410311073064804,
0.053087469190359116,
0.05213875323534012,
-0.04955948889255524,
-0.011594844050705433,
0.050204820930957794,
-0.047119539231061935,
0.016604876145720482,
0.03219744563102722,
0.03774092346429825,
0.03404289856553078,
0.022269045934081078,
0.0013820352032780647,
0.024254007264971733,
-0.011686353012919426,
-0.08671122789382935,
0.06015751510858536,
-0.019131679087877274,
0.012171972543001175,
-0.010972628369927406,
-0.037591833621263504,
-0.05389641597867012,
0.08729840815067291,
-0.05232689529657364,
0.010582643561065197,
0.0295474361628294,
-0.032416000962257385,
0.08680514246225357,
0.04938914626836777,
0.015124338679015636,
-0.014527224004268646,
-0.028658097609877586,
-0.013654380105435848,
-0.059282854199409485,
-0.030222348868846893,
-0.07629839330911636,
0.06610967218875885,
-0.0007536733173765242,
0.01129608042538166,
-0.05132976546883583,
-0.052591998130083084,
-0.04593701288104057,
-0.0167845468968153,
0.04608408734202385,
0.037933070212602615,
-0.07722822576761246,
0.054289814084768295,
0.025002339854836464,
0.007227173540741205,
0.011951596476137638,
0.018313318490982056,
-0.009927138686180115,
-0.008890810422599316,
-0.09642156213521957,
0.060937367379665375,
0.08680324256420135,
-0.08085379749536514,
0.006577766500413418,
-0.07113092392683029,
0.004091157577931881,
-0.008795679546892643,
0.03543274477124214,
0.03165337070822716,
-0.025308307260274887,
0.024160174652934074,
-0.014668671414256096,
-0.016739657148718834,
0.029367245733737946,
-0.04056979715824127,
-0.1233794167637825,
-0.019784074276685715,
-0.02228568308055401,
0.0680246502161026,
0.030287345871329308,
-0.06953757256269455,
-0.005318936891853809,
-0.0004586045106407255,
0.010935217142105103,
-0.05029381439089775,
-0.13438187539577484,
-0.11034492403268814,
0.004966536536812782,
0.048788052052259445,
-0.09876634180545807,
-0.08752012997865677,
-0.02437063679099083,
-0.06173252686858177,
-0.023550676181912422,
0.018892847001552582,
-0.027335578575730324,
-0.033033620566129684,
7.328350911375289e-34,
-0.060565024614334106,
-0.01896272785961628,
0.07301254570484161,
0.09870880097150803,
0.016558142378926277,
-0.09924881905317307,
0.013716381974518299,
-0.11371956020593643,
-0.024491405114531517,
-0.05882642790675163,
-0.11293736845254898,
-0.08188819885253906,
0.026184093207120895,
0.028794588521122932,
0.06445279717445374,
0.02284354157745838,
0.05684772878885269,
0.007129540201276541,
0.049199849367141724,
-0.0033144827466458082,
-0.03680217266082764,
0.03396294638514519,
0.06743364036083221,
-0.009749392978847027,
0.01558793056756258,
0.02962520532310009,
-0.005125201772898436,
0.07088974863290787,
-0.04127965867519379,
-0.04656319320201874,
0.01250532828271389,
-0.05976753681898117,
-0.08118882775306702,
-0.0009099550661630929,
-0.030252128839492798,
0.05374128371477127,
0.014812554232776165,
0.018509333953261375,
-0.015644855797290802,
0.030585328117012978,
-0.06994783878326416,
0.02819596789777279,
-0.11402364820241928,
0.0233172494918108,
0.00010435357398819178,
0.006770201493054628,
-0.04202282428741455,
-0.004321190062910318,
-0.010427745059132576,
0.040013816207647324,
0.00019131691078655422,
0.029669228941202164,
0.0025726708117872477,
0.050354212522506714,
-0.01808886229991913,
-0.04339073598384857,
-0.07957829535007477,
0.0655934140086174,
0.012828366830945015,
0.001246047206223011,
-0.0707443356513977,
-0.0844951793551445,
-0.09156357496976852,
0.05582720786333084,
0.07810147851705551,
0.029948759824037552,
0.025397086516022682,
0.019185936078429222,
0.06619467586278915,
0.040652357041835785,
-0.009588746353983879,
-0.0017027192516252398,
0.03417941927909851,
0.07297775894403458,
0.039097439497709274,
0.026088446378707886,
0.08906794339418411,
-0.03986981511116028,
-0.06872881948947906,
-0.03222469985485077,
-0.022986838594079018,
-0.03210047259926796,
0.07743634283542633,
-0.012897849082946777,
-0.0033745381515473127,
0.08255716413259506,
0.025069253519177437,
-0.0023829240817576647,
0.06156884506344795,
-0.011338836513459682,
-0.02149958908557892,
-0.020130379125475883,
0.07943414151668549,
0.06503131985664368,
0.05358681455254555,
-3.8774089716753224e-8,
-0.04378080740571022,
0.03899776190519333,
0.00403701514005661,
-0.04711863398551941,
0.08478330820798874,
-0.04059847071766853,
0.028258008882403374,
-0.005776457954198122,
-0.025926852598786354,
-0.03633646294474602,
0.024377398192882538,
-0.04936865344643593,
-0.05548131465911865,
-0.003527460154145956,
-0.022651836276054382,
0.008227789774537086,
-0.0428943932056427,
-0.0055753570050001144,
-0.003704864764586091,
0.024082982912659645,
-0.04363013058900833,
0.029414650052785873,
0.07131525129079819,
0.07039975374937057,
-0.05687658116221428,
0.002232842380180955,
-0.020978806540369987,
-0.019810188561677933,
0.15653324127197266,
0.05013749375939369,
-0.07851671427488327,
0.0036469788756221533,
0.07793603837490082,
0.0346631221473217,
0.038692522794008255,
-0.026791686192154884,
0.009718737564980984,
0.03341207653284073,
-0.030666761100292206,
-0.01940532587468624,
0.03218714892864227,
0.009780441410839558,
0.028117090463638306,
-0.01776817999780178,
0.03373950347304344,
0.0036090705543756485,
-0.03661765903234482,
-0.028977181762456894,
-0.008059101179242134,
0.0006224752287380397,
-0.020224226638674736,
0.0000459167713415809,
-0.02152966894209385,
0.08368264138698578,
0.006614378187805414,
-0.0656386986374855,
-0.04614332690834999,
-0.06791933625936508,
0.1326691061258316,
0.022463522851467133,
-0.03061389923095703,
0.009725214913487434,
-0.0779031291604042,
-0.009503824636340141
] | 0 | 0 | 0 | 0 |
|
DUET: Decentralized Bilevel Optimization without Lower-Level Strong Convexity | https://openreview.net/forum?id=jxMAPMqNr5 | [
"Zhen Qin",
"Zhuqing Liu",
"Songtao Lu",
"Yingbin Liang",
"Jia Liu"
] | Poster | Decentralized bilevel optimization (DBO) provides a powerful framework for multi-agent systems to solve local bilevel tasks in a decentralized fashion without the need for a central server.
However, most existing DBO methods rely on lower-level strong convexity (LLSC) to guarantee unique solutions and a well-defined hypergradient for stationarity measure, hindering their applicability in many practical scenarios not satisfying LLSC.
To overcome this limitation, we introduce a new single-loop DBO algorithm called diminishing quadratically-regularized bilevel decentralized optimization (DUET), which eliminates the need for LLSC by introducing a diminishing quadratic regularization to the lower-level (LL) objective.
We show that DUET achieves an iteration complexity of $O(1/T^{1-5p-\frac{11}{4}\tau})$ for approximate KKT-stationary point convergence under relaxed assumptions, where $p$ and $\tau $ are control parameters for LL learning rate and averaging, respectively.
In addition, our DUET algorithm incorporates gradient tracking to address data heterogeneity, a key challenge in DBO settings.
To the best of our knowledge, this is the first work to tackle DBO without LLSC under decentralized settings with data heterogeneity.
Numerical experiments validate the theoretical findings and demonstrate the practical effectiveness of our proposed algorithms. | Decentralized optimization, Bilevel optimization, Convex optimization, Heterogeneous data distributions, Multi agent learning | null | 12,177 | null | [
0.016620352864265442,
-0.05920254439115524,
0.0022192823234945536,
-0.026906676590442657,
-0.048493072390556335,
-0.03415519371628761,
-0.0047012255527079105,
-0.008248462341725826,
-0.054647188633680344,
-0.0016213092021644115,
-0.02465173229575157,
0.061143651604652405,
0.03384223207831383,
0.030417466536164284,
-0.005738548003137112,
0.10972867161035538,
0.12682609260082245,
0.032959964126348495,
-0.034241873770952225,
0.01755465567111969,
-0.026513371616601944,
-0.05262384191155434,
-0.006523506250232458,
0.008906476199626923,
0.061200253665447235,
-0.007109492085874081,
0.019278014078736305,
0.03487846627831459,
-0.013520250096917152,
0.01738808862864971,
-0.004894655663520098,
-0.029401255771517754,
-0.018712252378463745,
-0.033721692860126495,
-0.025157470256090164,
0.06769324094057083,
-0.03400415927171707,
0.02358979918062687,
0.03037240356206894,
0.08469629287719727,
0.07097099721431732,
0.028442377224564552,
-0.05787520855665207,
-0.10528074949979782,
0.03876049816608429,
-0.03299267590045929,
0.03389841318130493,
-0.021035999059677124,
0.024588020518422127,
-0.07646492123603821,
0.016341252252459526,
-0.06976428627967834,
-0.04982977360486984,
0.019376466050744057,
0.051601968705654144,
-0.016982302069664,
-0.014586425386369228,
-0.019786369055509567,
0.07569580525159836,
-0.04872971773147583,
-0.0067451042123138905,
-0.09125963598489761,
0.07641991972923279,
-0.031248487532138824,
0.021056191995739937,
-0.03050418384373188,
0.02529004216194153,
0.03976832330226898,
-0.024371588602662086,
-0.01061649713665247,
-0.07935024052858353,
0.01629016175866127,
-0.056159939616918564,
-0.035800885409116745,
0.02559650130569935,
-0.00453024310991168,
0.013718823902308941,
0.023426618427038193,
0.06807759404182434,
-0.046850960701704025,
0.002889784285798669,
0.05075361952185631,
-0.04050492122769356,
-0.03832981735467911,
0.01229025050997734,
-0.07798096537590027,
0.028606846928596497,
-0.05050009861588478,
0.10980258136987686,
-0.0514945350587368,
0.04493144899606705,
0.11187050491571426,
-0.0923837348818779,
-0.04911962151527405,
0.0693105012178421,
-0.06846677511930466,
0.061463408172130585,
-0.0287486519664526,
0.01761278696358204,
0.13059952855110168,
0.03751309961080551,
0.03800287842750549,
-0.02739053964614868,
-0.057456616312265396,
0.02025333233177662,
-0.09282708913087845,
0.0885826125741005,
0.06323758512735367,
0.0386664979159832,
-0.09433367103338242,
-0.012418679893016815,
0.040023624897003174,
0.04650014638900757,
0.00935111753642559,
-0.07076709717512131,
0.018438372761011124,
0.01920926943421364,
-0.0068956222385168076,
0.0054250420071184635,
0.07246487587690353,
-0.0369204580783844,
-0.060809679329395294,
0.001224132371135056,
0.09624867886304855,
0.03546838089823723,
0.08306067436933517,
-0.05621712654829025,
3.1978989532914317e-33,
-0.0031265942379832268,
0.100801981985569,
0.07304679602384567,
-0.10614889860153198,
0.0019481579074636102,
0.02372993901371956,
0.008381414227187634,
-0.02783026732504368,
-0.07444693148136139,
0.01661982387304306,
0.09862517565488815,
-0.020498337224125862,
-0.07340561598539352,
-0.00970623828470707,
0.08789694309234619,
-0.029621213674545288,
-0.007202028296887875,
-0.007618311792612076,
0.04819922521710396,
-0.07582321763038635,
0.07358696311712265,
-0.06869848072528839,
-0.04353773966431618,
-0.06971165537834167,
0.030504489317536354,
-0.008576301857829094,
0.08909721672534943,
-0.02163238637149334,
0.04183749482035637,
0.02079836279153824,
-0.023144656792283058,
-0.04450057074427605,
0.010448499582707882,
0.09931287914514542,
-0.017780983820557594,
0.04505763575434685,
-0.04457198455929756,
-0.05276232585310936,
-0.000469536695163697,
0.008809039369225502,
-0.024214595556259155,
0.036494623869657516,
-0.038628581911325455,
-0.042773254215717316,
-0.052171558141708374,
-0.05765516310930252,
-0.014519518241286278,
0.03691231831908226,
0.013689333572983742,
-0.12092003971338272,
-0.08926757425069809,
0.025910265743732452,
-0.05388738214969635,
0.013369400054216385,
0.011443762108683586,
-0.022253818809986115,
0.04481291025876999,
0.04329307749867439,
0.03487127274274826,
0.10173742473125458,
0.06797599047422409,
-0.0296311117708683,
-0.05043365806341171,
-0.02780020236968994,
0.004952359478920698,
0.02905307337641716,
-0.03158053383231163,
0.051402293145656586,
0.05556027591228485,
-0.06215381994843483,
0.016633912920951843,
0.0005094668595120311,
0.06412464380264282,
-0.06931187212467194,
0.035139016807079315,
-0.032532066106796265,
0.11335109174251556,
-0.05997024103999138,
-0.004039925057440996,
-0.06752265989780426,
-0.04676010459661484,
0.0850224643945694,
-0.05626635625958443,
-0.008742048405110836,
-0.05819527432322502,
-0.08814255893230438,
-0.013527894392609596,
-0.06624096632003784,
-0.05748565122485161,
-0.006182236596941948,
-0.09888283908367157,
0.01394121628254652,
-0.03038470633327961,
0.01554565317928791,
0.00198104209266603,
-2.617507250109566e-33,
0.01683320850133896,
-0.0397041030228138,
-0.018851350992918015,
0.06599651277065277,
0.060063187032938004,
-0.057841185480356216,
-0.09513644874095917,
-0.10060586780309677,
-0.04676727578043938,
-0.012539342977106571,
-0.025814048945903778,
-0.013417190872132778,
0.019236331805586815,
0.013206887058913708,
0.02580343373119831,
0.0013811163371428847,
-0.016312317922711372,
0.10427366197109222,
0.015504341572523117,
0.013794348575174809,
-0.046522822231054306,
0.07240406423807144,
0.02886979468166828,
0.06747448444366455,
0.09771476686000824,
0.06307339668273926,
-0.0711074247956276,
0.04109807685017586,
0.04935624822974205,
0.03993470221757889,
-0.04554842412471771,
-0.004102600738406181,
-0.021466486155986786,
0.003911519423127174,
-0.07569969445466995,
0.10633637756109238,
-0.06585295498371124,
0.06728105992078781,
-0.014004099182784557,
0.1276768147945404,
0.05725352466106415,
-0.07773493975400925,
-0.017144912853837013,
-0.04398619011044502,
-0.004253658931702375,
0.0005011740722693503,
-0.0078058610670268536,
0.009263993240892887,
0.08015526086091995,
-0.0049291797913610935,
0.01893199421465397,
-0.01641540229320526,
-0.06171872466802597,
0.1015072837471962,
-0.005427314899861813,
0.015765702351927757,
0.012067600153386593,
0.049312833696603775,
0.0006782498676329851,
-0.12637543678283691,
-0.04415496811270714,
-0.04227394610643387,
-0.021430442109704018,
0.10757815092802048,
0.0017950375331565738,
0.005058941897004843,
0.04342271760106087,
-0.014369387179613113,
0.015272191725671291,
0.0007711253128945827,
-0.09182999283075333,
0.030624110251665115,
0.014020782895386219,
0.01733488030731678,
0.022095542401075363,
0.09501062333583832,
0.02429104782640934,
0.01718292571604252,
-0.0335964672267437,
-0.030154025182127953,
0.025041047483682632,
0.037811923772096634,
-0.009849452413618565,
-0.03843862935900688,
0.001865670899860561,
-0.07200927287340164,
0.04441855475306511,
0.09124142676591873,
-0.02554834634065628,
-0.0640600174665451,
-0.0031437366269528866,
-0.03730194643139839,
-0.03072216548025608,
-0.04018009826540947,
-0.07384204119443893,
-4.883822057877296e-8,
-0.04676320031285286,
-0.0566699355840683,
-0.03086232580244541,
0.022857535630464554,
0.057780783623456955,
-0.05501933395862579,
0.009503464214503765,
-0.05409984663128853,
0.03479815647006035,
-0.013945524580776691,
0.0053483666852116585,
-0.02504216507077217,
0.005644777789711952,
-0.11384524405002594,
0.011543688364326954,
0.07212988287210464,
-0.02447955496609211,
0.031610798090696335,
0.004348983056843281,
-0.020648863166570663,
-0.09599555283784866,
0.021097388118505478,
-0.02017507515847683,
-0.01672285795211792,
0.02463560178875923,
-0.03766823932528496,
0.03387243300676346,
0.10190782696008682,
0.11365930736064911,
0.008871600031852722,
0.02886105701327324,
-0.0024055170360952616,
-0.01935032568871975,
0.01556811761111021,
-0.01063102949410677,
0.08387654274702072,
-0.09659601747989655,
0.08480455726385117,
-0.02704424224793911,
0.038585931062698364,
0.01806105673313141,
0.06082616001367569,
-0.042419884353876114,
-0.026728259399533272,
0.062417030334472656,
0.023169009014964104,
-0.01721889339387417,
0.013776514679193497,
0.06391698867082596,
-0.06423059850931168,
-0.032661620527505875,
0.06749028712511063,
0.04688281938433647,
0.02149040810763836,
0.05464375764131546,
0.05135411024093628,
0.025833705440163612,
-0.0740085169672966,
0.011903049424290657,
0.05015876889228821,
-0.002886915346607566,
-0.018731888383626938,
-0.04462684690952301,
-0.015297992154955864
] | 0 | 0 | 0 | 0 |
|
Preserving Deep Representations in One-Shot Pruning: A Hessian-Free Second-Order Optimization Framework | https://openreview.net/forum?id=eNQp79A5Oz | [
"Ryan Lucas",
"Rahul Mazumder"
] | Poster | We present SNOWS, a one-shot post-training pruning framework aimed at reducing the cost of vision network inference without retraining. Current leading one-shot pruning methods minimize layer-wise least squares reconstruction error which does not take into account deeper network representations. We propose to optimize a more global reconstruction objective. This objective accounts for nonlinear activations deep in the network to obtain a better proxy for the network loss. This nonlinear objective leads to a more challenging optimization problem---we demonstrate it can be solved efficiently using a specialized second-order optimization framework. A key innovation of our framework is the use of Hessian-free optimization to compute exact Newton descent steps without needing to compute or store the full Hessian matrix. A distinct advantage of SNOWS is that it can be readily applied on top of any sparse mask derived from prior methods, readjusting their weights to preserve deep feature representations. SNOWS obtains state-of-the-art results on various one-shot pruning benchmarks including residual networks and Vision Transformers (ViT/B-16 and ViT/L-16, 86m and 304m parameters respectively). Our open-source implementation is available at https://github.com/mazumder-lab/SNOWS. | Neural Network Pruning, Structured Pruning, Optimization, Hessian-free Optimization | An improved one-shot pruning framework using Hessian-free optimization | 12,176 | 2411.18376 | [
-0.08002004772424698,
-0.06951364874839783,
0.0051998598501086235,
0.039510101079940796,
0.08859392255544662,
0.045412369072437286,
-0.06715789437294006,
-0.07636596262454987,
0.01580217108130455,
-0.048094961792230606,
-0.018326956778764725,
-0.011832124553620815,
0.009949810802936554,
0.05244942381978035,
-0.05853169783949852,
0.01782064512372017,
0.046297162771224976,
0.12345840781927109,
-0.11976346373558044,
-0.04849143698811531,
-0.05501477047801018,
-0.027687575668096542,
-0.0014090666081756353,
0.01357671432197094,
0.06665050983428955,
-0.017929142341017723,
-0.018571708351373672,
-0.021051404997706413,
-0.031099291518330574,
-0.01585567183792591,
0.03760872781276703,
-0.08171261847019196,
-0.02922884002327919,
-0.015916362404823303,
-0.054115790873765945,
0.09561291337013245,
-0.020771164447069168,
0.012439129874110222,
-0.05385758727788925,
-0.03409542143344879,
-0.010756084695458412,
0.040904700756073,
-0.020791348069906235,
0.024779941886663437,
0.05720025300979614,
0.03885314241051674,
0.06670402735471725,
-0.04483048990368843,
0.09363691508769989,
0.008952303789556026,
-0.039316557347774506,
0.020521702244877815,
-0.05562744662165642,
0.06245270371437073,
0.02826221100986004,
0.04957308992743492,
0.004160976968705654,
-0.03535430505871773,
-0.07186421006917953,
0.003636706853285432,
0.020493805408477783,
-0.10604992508888245,
-0.013548674061894417,
0.04565323144197464,
-0.01098644733428955,
-0.01958070695400238,
0.05355897173285484,
-0.06595079600811005,
0.03563273325562477,
0.045087359845638275,
0.008706294000148773,
0.09814517945051193,
0.02014564536511898,
0.0032535274513065815,
0.037119269371032715,
0.0343763493001461,
0.12415731698274612,
-0.038781505078077316,
0.04939502477645874,
-0.008405431173741817,
0.041853487491607666,
0.044091083109378815,
0.02986316941678524,
0.05749950185418129,
0.09096301347017288,
-0.047911182045936584,
0.005367383360862732,
0.024743810296058655,
0.020263023674488068,
0.001092315185815096,
-0.11364812403917313,
-0.032854363322257996,
-0.029690483585000038,
0.005546350497752428,
-0.031721074134111404,
-0.038292113691568375,
0.015983596444129944,
-0.01156233623623848,
-0.09501999616622925,
0.08441311120986938,
-0.04815633222460747,
-0.0335603766143322,
0.05404899641871452,
-0.07654600590467453,
0.04668496549129486,
-0.005822022911161184,
0.026442425325512886,
0.053697142750024796,
0.041610900312662125,
-0.0494685098528862,
0.07411210238933563,
0.031530387699604034,
-0.0014932142803445458,
0.035636626183986664,
0.08722028881311417,
0.03857021406292915,
-0.022115344181656837,
-0.05427448824048042,
0.00876716710627079,
0.07222960889339447,
-0.0916406586766243,
0.02464057132601738,
-0.008804477751255035,
0.026264933869242668,
0.03947493061423302,
0.024893689900636673,
-0.010853220708668232,
1.5987366019656216e-33,
-0.026841148734092712,
0.10318546742200851,
0.09788352996110916,
-0.05861368030309677,
0.08593037724494934,
-0.017738373950123787,
0.06692098081111908,
0.04328995943069458,
-0.03650829568505287,
0.019467076286673546,
-0.0713508278131485,
-0.05553818121552467,
-0.06918378174304962,
0.09614347666501999,
-0.013516707345843315,
-0.034226689487695694,
-0.014829450286924839,
-0.00964382290840149,
0.05094599351286888,
0.02564595453441143,
-0.008081736974418163,
-0.019154919311404228,
-0.03350112587213516,
-0.0415017306804657,
0.0026991977356374264,
-0.028881533071398735,
0.10346455127000809,
-0.012780442833900452,
-0.05321017652750015,
0.01414494588971138,
-0.06888578832149506,
0.03523368015885353,
-0.025345580652356148,
0.015101255849003792,
0.06457824259996414,
-0.05151720717549324,
0.04470586031675339,
0.007809551432728767,
0.05928446725010872,
0.018290668725967407,
-0.018985560163855553,
0.09036725759506226,
-0.047429513186216354,
-0.08756808936595917,
-0.05911934748291969,
-0.1150771751999855,
0.023301901295781136,
0.0944724753499031,
-0.00867137499153614,
-0.019244829192757607,
0.05454642325639725,
0.026094239205121994,
0.024975981563329697,
-0.024168070405721664,
-0.05671250447630882,
-0.047164659947156906,
0.07865633070468903,
0.006285297684371471,
0.048613011837005615,
0.026172000914812088,
0.047483835369348526,
0.01813613809645176,
-0.04800157621502876,
0.0957280844449997,
-0.04821557551622391,
0.012152462266385555,
0.009643171913921833,
0.055429305881261826,
-0.02411317639052868,
-0.010139011777937412,
-0.042878516018390656,
0.020812280476093292,
-0.018971970304846764,
-0.10009399801492691,
0.0818018764257431,
0.03540096431970596,
0.050309617072343826,
-0.08058899641036987,
0.0585978627204895,
0.05733229219913483,
-0.0707641988992691,
0.064566470682621,
-0.009265510365366936,
0.009672618471086025,
-0.07417786121368408,
0.002815493382513523,
0.05211104080080986,
-0.026319311931729317,
0.013590214774012566,
0.012803221121430397,
-0.07708695530891418,
-0.051979221403598785,
-0.032482508569955826,
-0.06405619531869888,
0.03478134796023369,
-3.1878835118336854e-34,
0.0017182973679155111,
0.054470937699079514,
-0.07434526830911636,
0.03393951803445816,
-0.019225668162107468,
-0.02276904322206974,
-0.019186940044164658,
0.026803018525242805,
-0.03318767249584198,
-0.04314042255282402,
0.02636677213013172,
-0.026373572647571564,
-0.0210353285074234,
-0.02389977127313614,
0.03323403373360634,
0.004177024122327566,
-0.10812380164861679,
-0.01553342118859291,
-0.025903914123773575,
-0.06294362992048264,
0.039615511894226074,
0.11056467145681381,
-0.030330859124660492,
0.01788409985601902,
-0.07499147951602936,
0.028587235137820244,
-0.009075005538761616,
0.09415429085493088,
0.0062776170670986176,
-0.030702395364642143,
-0.032208774238824844,
-0.00602547125890851,
-0.05116083472967148,
-0.020842110738158226,
-0.046208880841732025,
0.09023898839950562,
-0.01102789118885994,
-0.015458151698112488,
-0.015562182292342186,
0.13950233161449432,
0.036594148725271225,
-0.04695560410618782,
-0.013924598693847656,
0.06953977048397064,
0.037899091839790344,
-0.014195876196026802,
-0.01599905453622341,
-0.06913446635007858,
0.0002293673751410097,
0.07653240859508514,
-0.03352586552500725,
0.04162268340587616,
-0.11169233173131943,
0.07397359609603882,
-0.053392428904771805,
-0.05555389076471329,
-0.06269521266222,
0.007094088010489941,
0.08432887494564056,
-0.025332029908895493,
-0.018540633842349052,
-0.06097494065761566,
-0.1360059380531311,
0.029835905879735947,
-0.01909482851624489,
-0.06939510256052017,
0.007458074018359184,
0.04580136016011238,
0.037958934903144836,
0.08900969475507736,
0.02752954699099064,
0.025968851521611214,
0.018793268129229546,
0.03495362028479576,
-0.050056107342243195,
-0.03516010567545891,
0.04264511913061142,
-0.0007618588279001415,
0.01816338114440441,
0.01632199063897133,
0.09943483024835587,
-0.031220845878124237,
0.04989892989397049,
0.019768042489886284,
0.1151546984910965,
0.08181430399417877,
0.06792427599430084,
0.008182198740541935,
0.008463791571557522,
-0.01596156507730484,
-0.08410874754190445,
0.011972625739872456,
-0.006416823714971542,
-0.023653052747249603,
-0.0522528775036335,
-3.8376551714236484e-8,
-0.10112423449754715,
0.09507694840431213,
-0.006595305632799864,
0.008292310871183872,
-0.013711989857256413,
-0.07347750663757324,
0.016049455851316452,
0.09738033264875412,
-0.00944842305034399,
-0.02135850116610527,
0.07082513719797134,
0.032139312475919724,
-0.007291723042726517,
-0.02245481126010418,
-0.10575400292873383,
0.022221330553293228,
0.004026062786579132,
-0.014286923222243786,
0.016626965254545212,
-0.0002112960210070014,
-0.0550568662583828,
0.025107432156801224,
-0.03657257556915283,
0.017265578731894493,
0.03858150914311409,
-0.058981753885746,
0.0038678585551679134,
0.05134715512394905,
0.11811602860689163,
0.013758650049567223,
-0.10431269556283951,
0.018430698662996292,
0.0233936607837677,
-0.039863258600234985,
-0.030756521970033646,
0.051108963787555695,
0.01502787135541439,
0.06499453634023666,
-0.023627184331417084,
-0.08499187976121902,
-0.05607219412922859,
0.022886550053954124,
0.011369193904101849,
-0.010806534439325333,
-0.027877366170287132,
0.0586041696369648,
-0.02017681859433651,
-0.045723069459199905,
0.025290556252002716,
-0.010648513212800026,
0.04095352813601494,
0.0013797805877402425,
-0.012085785157978535,
0.11008510738611221,
0.012155461125075817,
0.003606269834563136,
0.0034326654858887196,
-0.05725866183638573,
-0.014039603993296623,
0.06140479817986488,
-0.010114644654095173,
0.005002019926905632,
-0.10338424891233444,
-0.04691615328192711
] | 0 | 0 | 0 | 0 |
|
Chemistry-Inspired Diffusion with Non-Differentiable Guidance | https://openreview.net/forum?id=4dAgG8ma3B | [
"Yuchen Shen",
"Chenhao Zhang",
"Sijie Fu",
"Chenghui Zhou",
"Newell Washburn",
"Barnabas Poczos"
] | Poster | Recent advances in diffusion models have shown remarkable potential in the conditional generation of novel molecules. These models can be guided in two ways: (i) explicitly, through additional features representing the condition, or (ii) implicitly, using a property predictor. However, training property predictors or conditional diffusion models requires an abundance of labeled data and is inherently challenging in real-world applications. We propose a novel approach that attenuates the limitations of acquiring large labeled datasets by leveraging domain knowledge from quantum chemistry as a non-differentiable oracle to guide an unconditional diffusion model. Instead of relying on neural networks, the oracle provides accurate guidance in the form of estimated gradients, allowing the diffusion process to sample from a conditional distribution specified by quantum chemistry. We show that this results in more precise conditional generation of novel and stable molecular structures. Our experiments demonstrate that our method: (1) significantly reduces atomic forces, enhancing the validity of generated molecules when used for stability optimization; (2) is compatible with both explicit and implicit guidance in diffusion models, enabling joint optimization of molecular properties and stability; and (3) generalizes effectively to molecular optimization tasks beyond stability optimization. Our implementation is available at https://github.com/A-Chicharito-S/ChemGuide. | guided diffusion, ai4science, molecule generation | null | 12,166 | 2410.06502 | [
-0.08832688629627228,
-0.07187901437282562,
-0.038001663982868195,
0.07778909802436829,
0.03605286777019501,
0.0020689500961452723,
-0.02105877362191677,
-0.019389571622014046,
-0.012848027050495148,
-0.06059708446264267,
0.003207210684195161,
-0.06749963760375977,
-0.011565647087991238,
0.07930734753608704,
-0.050300560891628265,
0.011119350790977478,
0.04965827241539955,
0.04485110193490982,
-0.09442138671875,
-0.013246634043753147,
-0.029349103569984436,
0.07567667961120605,
0.050810202956199646,
0.012382068671286106,
0.026311278343200684,
0.1073237732052803,
0.07213277369737625,
-0.004638712387531996,
0.007864661514759064,
-0.016954954713582993,
0.1117308959364891,
-0.021330561488866806,
-0.11343148350715637,
-0.009838690049946308,
0.015375697053968906,
-0.00899985246360302,
-0.12719100713729858,
0.02249956876039505,
0.006246526259928942,
-0.073475182056427,
-0.021488439291715622,
0.008692270144820213,
-0.040609750896692276,
0.05672987550497055,
0.07452879846096039,
-0.024132732301950455,
0.04351573437452316,
-0.008846484124660492,
0.008732402697205544,
-0.01403732504695654,
0.02167435735464096,
-0.03486528620123863,
-0.005623891018331051,
-0.010596572421491146,
-0.00977503601461649,
0.057447317987680435,
0.0432840920984745,
0.026682790368795395,
-0.038711030036211014,
-0.043006692081689835,
-0.024164676666259766,
-0.12954941391944885,
-0.004431163892149925,
-0.010448990389704704,
-0.0320124551653862,
0.045735202729701996,
-0.017697107046842575,
0.08019997179508209,
0.011818191036581993,
0.06159045919775963,
-0.04874838888645172,
0.08552948385477066,
-0.06372524797916412,
0.022888848558068275,
0.024588562548160553,
0.015155088156461716,
0.11490727216005325,
0.032801851630210876,
0.029744978994131088,
-0.08917362987995148,
-0.011224132962524891,
-0.03680653125047684,
0.0673837885260582,
0.030206678435206413,
0.06917953491210938,
0.05744826793670654,
0.04530741646885872,
0.03297477215528488,
0.021110795438289642,
-0.032702937722206116,
-0.022156044840812683,
0.033631451427936554,
-0.01134360022842884,
-0.06341765820980072,
-0.06786543875932693,
0.059587594121694565,
0.04983977973461151,
-0.07481987774372101,
0.0034332459326833487,
0.0999947190284729,
0.010098966769874096,
0.07127972692251205,
-0.01771976798772812,
-0.07588658481836319,
0.05134496092796326,
-0.057001542299985886,
0.07006882131099701,
0.07649026811122894,
-0.011400321498513222,
-0.005535263102501631,
0.038941141217947006,
0.09507397562265396,
0.03015628643333912,
-0.053249556571245193,
0.01650838367640972,
-0.05897698178887367,
-0.0020772104617208242,
-0.00305798277258873,
0.007311181630939245,
0.03279159963130951,
-0.09409631043672562,
-0.014385693706572056,
-0.10239729285240173,
0.03295448049902916,
-0.008403178304433823,
-0.025556357577443123,
-0.12902668118476868,
6.9227805647576375e-34,
0.00938829593360424,
-0.05772598460316658,
0.0619451105594635,
0.042246606200933456,
0.051470182836055756,
-0.04614511877298355,
0.015041839331388474,
-0.02943280339241028,
-0.0664285346865654,
-0.009818303398787975,
0.016515536233782768,
-0.042999740689992905,
-0.12034232169389725,
0.019730374217033386,
0.03423352539539337,
0.003995060920715332,
0.008898886851966381,
0.034593772143125534,
0.028858602046966553,
-0.025886140763759613,
-0.03206222131848335,
-0.05079612880945206,
-0.052004437893629074,
-0.029129397124052048,
-0.0438089556992054,
0.039987627416849136,
0.048700567334890366,
0.004409237764775753,
0.01378372497856617,
-0.01851472444832325,
-0.0883355438709259,
0.028729375451803207,
-0.07192211598157883,
-0.010915129445493221,
0.10024833679199219,
0.022700324654579163,
-0.07382423430681229,
-0.03116622567176819,
0.02708028070628643,
-0.025529984384775162,
0.024551816284656525,
0.029483146965503693,
0.03874519094824791,
0.06616774201393127,
-0.06944908946752548,
0.05418930575251579,
-0.019414082169532776,
-0.004685422871261835,
-0.058538906276226044,
-0.07043137401342392,
-0.015858113765716553,
-0.003120576264336705,
0.009306693449616432,
-0.03181331604719162,
-0.015772240236401558,
-0.04136577248573303,
0.02378656342625618,
0.006818311754614115,
-0.032648760825395584,
-0.020992549136281013,
0.028067007660865784,
0.07099603116512299,
-0.06694146245718002,
0.05268780142068863,
-0.036404091864824295,
0.07605696469545364,
-0.12083867937326431,
0.05695255845785141,
0.031198253855109215,
0.043656278401613235,
-0.008019584231078625,
0.06769672781229019,
0.03695804253220558,
-0.0829203650355339,
0.11703728884458542,
0.03493925556540489,
0.0015426769386976957,
-0.09056754410266876,
0.06471529603004456,
0.005463621113449335,
0.05272573232650757,
-0.008809245191514492,
-0.013816077262163162,
0.012406060472130775,
-0.07619605958461761,
-0.10016525536775589,
-0.013667961582541466,
-0.09042282402515411,
-0.08939962089061737,
0.006938815116882324,
-0.021473504602909088,
-0.0915081799030304,
0.053081609308719635,
0.0031983719673007727,
-0.0006003105663694441,
-2.5324251526680412e-33,
-0.013916843570768833,
0.08027410507202148,
0.0653301253914833,
-0.028011273592710495,
-0.03329348564147949,
0.02998492680490017,
-0.07423681765794754,
-0.01993049494922161,
0.08398569375276566,
-0.03317075967788696,
0.05360136553645134,
-0.005240340251475573,
-0.010232970118522644,
-0.02110232040286064,
-0.03284401446580887,
0.0409574881196022,
-0.021854843944311142,
0.015648899599909782,
0.0872982069849968,
0.02593212202191353,
-0.07468830794095993,
0.09522240608930588,
-0.059623364359140396,
0.06320635229349136,
-0.07702997326850891,
-0.016255035996437073,
-0.013772744685411453,
0.06340828537940979,
-0.014415749348700047,
0.017024321481585503,
-0.0673292949795723,
-0.017262842506170273,
-0.05315856635570526,
-0.04160693660378456,
-0.11091159284114838,
0.049339957535266876,
0.008025421760976315,
-0.019741207361221313,
-0.03215692937374115,
0.011632795445621014,
-0.04485795646905899,
-0.042310964316129684,
-0.08359596878290176,
0.031452521681785583,
0.04892405867576599,
0.02852597087621689,
0.00018533883849158883,
0.023866696283221245,
-0.00762311602011323,
0.020461760461330414,
0.04455864056944847,
0.06483249366283417,
-0.019154969602823257,
-0.021167460829019547,
-0.03345189243555069,
-0.030981028452515602,
-0.09279894083738327,
-0.06281906366348267,
0.1598241627216339,
-0.0024816447403281927,
-0.06687247008085251,
-0.051294341683387756,
0.02401113137602806,
0.009129523299634457,
-0.11204013973474503,
-0.0645512267947197,
-0.0722501128911972,
0.05200916528701782,
0.07095359265804291,
-0.01761842891573906,
-0.02345048077404499,
-0.019096743315458298,
0.10170003771781921,
0.056874655187129974,
0.034747809171676636,
-0.10276616364717484,
0.049217432737350464,
-0.1188792958855629,
-0.031426914036273956,
-0.010633026249706745,
0.043670155107975006,
0.05020580068230629,
0.007224637549370527,
0.01758493110537529,
0.06552086770534515,
0.0464455783367157,
0.04685375466942787,
0.00019739262643270195,
0.03684575483202934,
0.01973080076277256,
0.008935901336371899,
0.028497792780399323,
0.020658234134316444,
0.053146544843912125,
-0.06967917084693909,
-4.221691085604107e-8,
0.023326944559812546,
0.0023805759847164154,
0.04547436907887459,
-0.009199628606438637,
0.05430194362998009,
0.06267841160297394,
0.10673345625400543,
0.025005467236042023,
-0.007669093087315559,
-0.016451528295874596,
0.020691992715001106,
0.062420591711997986,
-0.06353706866502762,
-0.04614574834704399,
0.0007549191941507161,
0.09243842959403992,
0.0038297418504953384,
-0.015309182927012444,
-0.011639873497188091,
-0.0011350789573043585,
-0.002285504713654518,
0.0436282716691494,
0.012857749126851559,
-0.007740940432995558,
0.10560835897922516,
-0.08756355941295624,
0.006227595265954733,
-0.030382560566067696,
0.040963996201753616,
0.012996170669794083,
-0.0045431265607476234,
-0.006377932149916887,
0.020487524569034576,
0.08949791640043259,
0.027083462104201317,
0.09206350892782211,
0.004217774141579866,
-0.051350757479667664,
-0.054822228848934174,
-0.007945307530462742,
-0.0519062876701355,
0.08576465398073196,
-0.06570689380168915,
0.03652196750044823,
-0.020467368885874748,
0.016250066459178925,
0.06783753633499146,
-0.007128565106540918,
0.05986255034804344,
0.00484502175822854,
0.0260336734354496,
-0.021419860422611237,
0.021633030846714973,
-0.04695133492350578,
0.023478738963603973,
0.0037666524294763803,
-0.07077278196811676,
0.0018603039206936955,
-0.01162752229720354,
0.006639903876930475,
0.03683414310216904,
-0.006405489519238472,
-0.018979957327246666,
-0.01468503475189209
] | 0 | 0 | 0 | 0 |
|
Efficient Diffusion Transformer Policies with Mixture of Expert Denoisers for Multitask Learning | https://openreview.net/forum?id=nDmwloEl3N | [
"Moritz Reuss",
"Jyothish Pari",
"Pulkit Agrawal",
"Rudolf Lioutikov"
] | Poster | Diffusion Policies have become widely used in Imitation Learning, offering several appealing properties, such as generating multimodal and discontinuous behavior.
As models are becoming larger to capture more complex capabilities, their computational demands increase, as shown by recent scaling laws.
Therefore, continuing with the current architectures will present a computational roadblock.
To address this gap, we propose Mixture-of-Denoising Experts (MoDE) as a novel policy for Imitation Learning.
MoDE surpasses current state-of-the-art Transformer-based Diffusion Policies while enabling parameter-efficient scaling through sparse experts and noise-conditioned routing, reducing both active parameters by 40\% and inference costs by 90\% via expert caching.
Our architecture combines this efficient scaling with noise-conditioned self-attention mechanism, enabling more effective denoising across different noise levels.
MoDE achieves state-of-the-art performance on 134 tasks in four established imitation learning benchmarks (CALVIN and LIBERO).
Notably, by pretraining MoDE on diverse robotics data, we achieve 4.01 on CALVIN ABC and 0.95 on LIBERO-90.
It surpasses both CNN-based and Transformer Diffusion Policies by an average of $57\%$ across 4 benchmarks, while using 90\% fewer FLOPs and fewer active parameters compared to default Diffusion Transformer architectures. Furthermore, we conduct comprehensive ablations on MoDE's components, providing insights for designing efficient and scalable Transformer architectures for Diffusion Policies.
Code and demonstrations are available at https://mbreuss.github.io/MoDE_Diffusion_Policy. | Robotics, Imitation Learning | Mixture-of-Denoising Experts (MoDE), a novel Diffusion Policy that leverages a noise-conditioned routing strategy to achieve more efficient denoising and improved performance compared to prior approaches. | 12,163 | 2412.12953 | [
-0.039785418659448624,
-0.14161153137683868,
0.03461853414773941,
0.010863972827792168,
-0.017520194873213768,
-0.05345052108168602,
0.007430872414261103,
-0.05743923410773277,
-0.02167508751153946,
0.005794482305645943,
-0.02673344686627388,
-0.003146709408611059,
0.008732343092560768,
0.04398667812347412,
-0.029847402125597,
0.07022944837808609,
0.06986573338508606,
0.0887853354215622,
-0.1319645196199417,
-0.055687472224235535,
0.014789464883506298,
-0.04916651174426079,
0.005242649000138044,
0.01488307025283575,
0.05385438725352287,
0.06069174408912659,
0.029209954664111137,
-0.0938902497291565,
-0.02246527187526226,
-0.03025410696864128,
0.030981145799160004,
-0.05522529035806656,
-0.009550807997584343,
-0.038564737886190414,
-0.03624541684985161,
0.03800000622868538,
-0.12132206559181213,
-0.013463091105222702,
0.03260892629623413,
-0.034203436225652695,
0.0009680493385531008,
-0.004025006201118231,
-0.0417296402156353,
0.04787874594330788,
0.06090908870100975,
0.06936414539813995,
0.037862665951251984,
-0.022789746522903442,
0.13935275375843048,
-0.004375695250928402,
0.011620478704571724,
-0.0426153764128685,
-0.0004503693198785186,
0.058165427297353745,
0.005890101660043001,
0.059515323489904404,
0.05197908356785774,
0.05076536536216736,
0.02250698022544384,
-0.01699146255850792,
-0.05698705092072487,
-0.09422727674245834,
0.017161229625344276,
-0.014300988055765629,
-0.04744495451450348,
-0.007173237390816212,
0.032274018973112106,
-0.012164785526692867,
-0.0034797019325196743,
0.03798174113035202,
-0.0353851318359375,
0.04120025411248207,
0.020139571279287338,
0.025902971625328064,
0.04354453086853027,
-0.06793469935655594,
0.04765453562140465,
-0.05287721008062363,
0.015366545878350735,
-0.06838333606719971,
0.030058182775974274,
-0.04203208535909653,
-0.0006120685720816255,
-0.005383974872529507,
0.1361846774816513,
-0.0787932351231575,
-0.020394375547766685,
0.01870090700685978,
0.083969347178936,
-0.02693456970155239,
-0.07997170090675354,
0.0018798242090269923,
-0.05833374708890915,
-0.05314488708972931,
-0.02143300697207451,
0.038307204842567444,
0.00254146964289248,
-0.007426716852933168,
-0.06954216957092285,
0.09068060666322708,
-0.01916286163032055,
-0.04508408159017563,
-0.004185122437775135,
0.023762376978993416,
-0.006883017718791962,
-0.08697748929262161,
0.05698752775788307,
0.06413748115301132,
0.04507962986826897,
0.02813929133117199,
0.016591215506196022,
0.021004291251301765,
0.07236544042825699,
0.01376365590840578,
0.05871860310435295,
0.01025956030935049,
0.03464929759502411,
-0.0013775735860690475,
-0.00716642988845706,
0.09439104795455933,
-0.010475574992597103,
-0.06002221256494522,
-0.028451908379793167,
0.0031183643732219934,
0.05634360387921333,
0.02455408126115799,
-0.03317905589938164,
1.3620582531027428e-33,
0.004257838241755962,
0.030266569927334785,
0.040431052446365356,
-0.01313052512705326,
0.03227878734469414,
-0.02191142551600933,
0.04642269387841225,
-0.003542862832546234,
-0.03942118212580681,
-0.07404175400733948,
-0.0030659798067063093,
0.03788650035858154,
-0.10059843957424164,
0.12543047964572906,
-0.017582932487130165,
-0.03079984337091446,
-0.05490094795823097,
0.027259141206741333,
0.02375025860965252,
-0.05999040603637695,
0.08170445263385773,
0.0025142631493508816,
-0.0690898671746254,
-0.038609571754932404,
-0.046609509736299515,
-0.059038590639829636,
0.05280481278896332,
-0.07546991109848022,
0.031196950003504753,
0.010338155552744865,
-0.14968937635421753,
0.009297535754740238,
-0.09019582718610764,
0.0105479396879673,
0.0521559938788414,
-0.00028984868549741805,
0.002508722012862563,
-0.016174158081412315,
-0.026167377829551697,
-0.0025454473216086626,
-0.0024868131149560213,
0.010671748779714108,
-0.014379525557160378,
-0.03062542900443077,
-0.1369774490594864,
-0.048164721578359604,
0.02053338661789894,
0.026328209787607193,
-0.0157600250095129,
-0.023258062079548836,
0.05763014033436775,
-0.056035641580820084,
-0.03310149163007736,
-0.10033117979764938,
0.018894240260124207,
-0.04784107580780983,
0.06400549411773682,
0.0033419292885810137,
0.03693752363324165,
-0.049465980380773544,
0.038699787110090256,
-0.013543887995183468,
-0.009456881321966648,
0.06083613634109497,
0.060271743685007095,
0.04104609414935112,
-0.01882370375096798,
0.1300283968448639,
0.06245513632893562,
-0.042630575597286224,
-0.030228665098547935,
0.044486355036497116,
0.017433052882552147,
-0.07773777097463608,
0.06711751967668533,
0.012586498633027077,
0.04438118636608124,
-0.028339417651295662,
-0.009334909729659557,
0.0267679113894701,
-0.04572683945298195,
-0.03483074530959129,
-0.07463394105434418,
-0.07148883491754532,
-0.0020287511870265007,
0.033912915736436844,
0.02382245846092701,
-0.0782092809677124,
-0.031617484986782074,
-0.04930515214800835,
-0.08810275793075562,
-0.044112227857112885,
-0.026872316375374794,
0.028238162398338318,
-0.03903396427631378,
-2.6342863299416003e-33,
0.02750585786998272,
0.05139836668968201,
0.04320167750120163,
0.10691075772047043,
-0.008118231780827045,
-0.006116935517638922,
-0.013029449619352818,
0.024611612781882286,
-0.0086093470454216,
-0.06946435570716858,
-0.04379098117351532,
-0.00870431587100029,
0.07444983720779419,
-0.034175675362348557,
0.021922120824456215,
-0.027291473001241684,
-0.07059595733880997,
0.046571213752031326,
-0.028582757338881493,
-0.008230073377490044,
-0.02064918912947178,
0.047594573348760605,
-0.06130697950720787,
0.0859895795583725,
-0.10572831332683563,
0.0061313435435295105,
-0.11094506829977036,
0.12609025835990906,
0.062035199254751205,
0.007682787720113993,
-0.0438164621591568,
-0.025030994787812233,
0.03232612833380699,
-0.04466724023222923,
-0.04927600547671318,
0.15414337813854218,
-0.03357577696442604,
0.06902367621660233,
-0.024305669590830803,
0.07135346531867981,
0.055417295545339584,
-0.0651480033993721,
-0.010785537771880627,
0.006633032113313675,
0.015877768397331238,
-0.09036614745855331,
-0.02986205369234085,
-0.03374350070953369,
-0.04206521809101105,
0.032599590718746185,
0.04031676799058914,
0.06479962915182114,
-0.11270316690206528,
0.00817656610161066,
-0.044953852891922,
0.0036253901198506355,
-0.0534491240978241,
0.014263720251619816,
0.060620855540037155,
-0.009901676326990128,
-0.001956187654286623,
0.0017914988566190004,
-0.0255192369222641,
-0.032578349113464355,
-0.07550099492073059,
0.011898176744580269,
-0.019236404448747635,
0.05039519816637039,
0.03284406661987305,
0.04906362295150757,
-0.036139924079179764,
-0.014392487704753876,
0.049051664769649506,
-0.037701744586229324,
0.040246959775686264,
-0.0904177725315094,
0.05869743227958679,
-0.014645915478467941,
-0.009326044470071793,
0.009576072916388512,
0.008470402099192142,
0.010135090909898281,
0.025413336232304573,
0.018142959102988243,
0.028536129742860794,
0.08618629723787308,
-0.01497455034404993,
0.025479771196842194,
0.03545597195625305,
0.01765206828713417,
0.035642750561237335,
0.01729009672999382,
0.03178028762340546,
0.020846856757998466,
-0.06267163157463074,
-4.6475882697905035e-8,
-0.06698761135339737,
0.09794922173023224,
0.08188342303037643,
0.004729425068944693,
0.09507586807012558,
-0.06422222405672073,
0.029605485498905182,
0.02611183002591133,
-0.041124895215034485,
-0.02536783181130886,
0.14254306256771088,
-0.014792821370065212,
-0.046923596411943436,
-0.0734310895204544,
0.01575215347111225,
0.051703885197639465,
0.0690818503499031,
0.0036961755249649286,
-0.004510377999395132,
-0.02670031040906906,
0.0016971987206488848,
-0.025353841483592987,
0.005664643365889788,
-0.030237888917326927,
0.05434762313961983,
-0.00013572494208347052,
-0.0487767793238163,
0.024137213826179504,
0.043741777539253235,
0.02676480822265148,
-0.056369882076978683,
0.01149546168744564,
0.0671767070889473,
-0.03371879830956459,
-0.009310070425271988,
0.06528914719820023,
0.07308518886566162,
-0.03594339266419411,
-0.02735540084540844,
-0.08645299077033997,
0.028101256117224693,
-0.0216693002730608,
0.01235869899392128,
0.024366792291402817,
-0.026751229539513588,
0.030669456347823143,
-0.035161975771188736,
-0.10252492874860764,
0.06096638739109039,
0.02861258015036583,
0.09668611735105515,
-0.029281096532940865,
-0.0003832704678643495,
0.101154625415802,
0.10594644397497177,
0.019115880131721497,
0.02086396887898445,
-0.02404131181538105,
0.01244961004704237,
0.08906614780426025,
0.07777431607246399,
-0.000007854765499359928,
-0.07116305828094482,
0.03680572658777237
] | 0 | 0 | 0 | 0 |
|
No Location Left Behind: Measuring and Improving the Fairness of Implicit Representations for Earth Data | https://openreview.net/forum?id=hSZaCIznB2 | [
"Daniel Cai",
"Randall Balestriero"
] | Poster | Implicit neural representations (INRs) exhibit growing promise in addressing Earth representation challenges, ranging from emissions monitoring to climate modeling. However, existing methods disproportionately prioritize global average performance, whereas practitioners require fine-grained insights to understand biases and variations in these models. To bridge this gap, we introduce FAIR-Earth: a first-of-its-kind dataset explicitly crafted to challenge and examine inequities in Earth representations. FAIR-Earth comprises various high-resolution Earth signals, and uniquely aggregates extensive metadata along stratifications like landmass size and population density to assess the fairness of models. Evaluating state-of-the-art INRs across the various modalities of FAIR-Earth, we uncover striking performance disparities. Certain subgroups, especially those associated with high-frequency signals (e.g., islands, coastlines), are consistently poorly modeled by existing methods. In response, we propose spherical wavelet encodings, building on previous spatial encoding research for INRs. Leveraging the multi-resolution analysis capabilities of wavelets, our encodings yield more consistent performance over various scales and locations, offering more accurate and robust representations of the biased subgroups. These open-source contributions represent a crucial step towards facilitating the equitable assessment and deployment of implicit Earth representations. | implicit neural representations, dataset, fairness in AI, representation learning, geospatial modeling, Earth representation, wavelet, location encoding | null | 12,161 | 2502.06831 | [
-0.0164328683167696,
-0.1226254478096962,
0.1281110644340515,
0.025460347533226013,
0.11113614588975906,
0.007782391272485256,
-0.04044913500547409,
-0.06866363435983658,
0.056312669068574905,
-0.014410410076379776,
-0.08669582009315491,
-0.10504031926393509,
-0.013888775371015072,
0.038995303213596344,
0.01450601126998663,
-0.009129655547440052,
0.0669407993555069,
-0.010803673416376114,
-0.10326467454433441,
-0.061647649854421616,
0.043596476316452026,
0.061118051409721375,
-0.0000495602835144382,
0.042359858751297,
0.06451107561588287,
-0.07331398129463196,
0.0249053705483675,
0.03831341117620468,
-0.04939465969800949,
0.045430704951286316,
0.08393501490354538,
0.04506836459040642,
-0.0004906678223051131,
0.03206367790699005,
-0.05993166193366051,
0.04346243664622307,
-0.029173988848924637,
-0.018609125167131424,
-0.029566366225481033,
-0.026422180235385895,
0.045782238245010376,
-0.023188281804323196,
0.020881401374936104,
0.005568197928369045,
0.01063396967947483,
0.06442739814519882,
0.055230334401130676,
-0.04424194246530533,
-0.03235573694109917,
-0.008571849204599857,
0.07999115437269211,
0.012581824325025082,
-0.03858839347958565,
0.016355479136109352,
-0.05450933426618576,
0.007958278059959412,
-0.026749124750494957,
-0.012080728076398373,
0.00902759563177824,
-0.003999954089522362,
-0.06576920300722122,
-0.0804787427186966,
-0.047007109969854355,
0.022195955738425255,
0.05018587037920952,
0.03756677359342575,
0.030747754499316216,
0.028605390340089798,
-0.04123445972800255,
-0.008193599060177803,
0.018056198954582214,
0.0017624489264562726,
-0.014689294621348381,
-0.018007181584835052,
0.05108293518424034,
0.0017243914771825075,
-0.04871184751391411,
0.02388354204595089,
0.05449672415852547,
-0.08574773371219635,
0.04300199821591377,
0.019448749721050262,
0.01722121424973011,
0.03590986132621765,
0.12845037877559662,
-0.0351853147149086,
-0.0023239520378410816,
0.0822460949420929,
0.016482627019286156,
0.006317225284874439,
-0.04609694704413414,
0.000514269049745053,
0.08079651743173599,
0.09036768227815628,
-0.029913559556007385,
0.055151693522930145,
0.06409291177988052,
-0.06300994753837585,
0.005389149766415358,
0.08344601839780807,
0.02638370729982853,
0.018573367968201637,
-0.02562412992119789,
-0.012619348242878914,
0.06961214542388916,
-0.041636668145656586,
0.020460454747080803,
0.06225084140896797,
0.004850872326642275,
-0.006983364466577768,
-0.006452939007431269,
-0.042220067232847214,
0.011762307956814766,
0.04398803412914276,
-0.012405927293002605,
0.0018137517618015409,
0.028459586203098297,
-0.013057020492851734,
0.0317339152097702,
-0.033280208706855774,
-0.0551239438354969,
-0.050585970282554626,
0.05025073140859604,
0.09163547307252884,
0.016903189942240715,
0.02271874062716961,
-0.09071603417396545,
1.576575962058729e-33,
0.004705843050032854,
0.0940585508942604,
0.04734204337000847,
-0.025303203612565994,
-0.007118418347090483,
0.008521431125700474,
-0.02156667225062847,
0.07059144973754883,
0.008316447027027607,
0.07551924139261246,
-0.06696263700723648,
0.05148028954863548,
-0.06767293065786362,
0.08694963902235031,
0.018259719014167786,
0.03729860112071037,
-0.0689619928598404,
0.06274628639221191,
-0.09727965295314789,
0.044864945113658905,
-0.027219761162996292,
-0.03543207049369812,
0.02027270384132862,
-0.016470640897750854,
-0.015100088901817799,
0.0002970992645714432,
0.007067759986966848,
-0.02557111158967018,
-0.02872438356280327,
-0.006790153216570616,
-0.019784461706876755,
-0.002647666959092021,
0.01957038976252079,
0.031696002930402756,
0.010276367887854576,
-0.05400829762220383,
0.02008718252182007,
0.01799270138144493,
-0.04458639770746231,
0.01363606657832861,
0.03866523504257202,
0.028427109122276306,
-0.008526724763214588,
-0.045442309230566025,
0.017101548612117767,
-0.07860646396875381,
0.07890869677066803,
-0.003588691120967269,
-0.01643609069287777,
0.05701012164354324,
0.02729327231645584,
0.013583227060735226,
0.004709878470748663,
-0.10447672754526138,
0.04170762002468109,
0.03206721320748329,
0.028781546279788017,
0.027661243453621864,
0.03987812250852585,
0.03578914329409599,
-0.08634084463119507,
-0.04879174381494522,
0.03668326511979103,
-0.007182418368756771,
0.0292096808552742,
0.013978014700114727,
0.059950634837150574,
0.05290825664997101,
-0.07641042023897171,
-0.006073363590985537,
0.05574198439717293,
0.09368956089019775,
-0.00129323813598603,
-0.04418731853365898,
0.003016907023265958,
-0.0034077614545822144,
0.04667036607861519,
0.02484169229865074,
-0.03187743201851845,
0.05426958575844765,
0.014782195910811424,
-0.0041241697035729885,
-0.01298332680016756,
-0.051704518496990204,
-0.038492344319820404,
-0.02566632814705372,
0.030410751700401306,
-0.053403157740831375,
0.0505235493183136,
-0.03893700987100601,
-0.06191644445061684,
-0.06699114292860031,
-0.026564806699752808,
-0.03901813179254532,
-0.07985491305589676,
-9.32494321910381e-34,
-0.07329031080007553,
0.012079991400241852,
-0.06449531763792038,
-0.0027309099677950144,
-0.05562075227499008,
-0.09664107859134674,
0.00002867802686523646,
0.021103903651237488,
0.03958648070693016,
0.0033913992810994387,
-0.011709313839673996,
-0.024220144376158714,
0.12346016615629196,
-0.01949373260140419,
0.012363739311695099,
-0.035274527966976166,
0.008228970691561699,
-0.011322733014822006,
-0.01882019266486168,
-0.02155442163348198,
0.04452400654554367,
-0.016319138929247856,
-0.034315213561058044,
0.045960988849401474,
-0.08784155547618866,
0.05348978564143181,
-0.009496433660387993,
-0.00361484382301569,
0.00045793852768838406,
0.014691449701786041,
-0.08369679003953934,
0.025571409612894058,
-0.025707637891173363,
-0.124392569065094,
-0.05707816407084465,
0.020409276708960533,
0.06721068173646927,
-0.026690978556871414,
-0.09503009915351868,
-0.0012924154289066792,
-0.003659564536064863,
0.00599714508280158,
0.0003320196701679379,
0.007318727672100067,
-0.0442751981317997,
-0.029078343883156776,
-0.009606197476387024,
0.058539293706417084,
-0.049958497285842896,
0.04033942148089409,
0.016247648745775223,
0.023129064589738846,
-0.06707736849784851,
0.1097414419054985,
-0.004908557049930096,
-0.06327833980321884,
0.0154463155195117,
-0.036619555205106735,
0.07551950216293335,
-0.035217009484767914,
-0.07602283358573914,
-0.07098861038684845,
-0.09899703413248062,
-0.00020475310157053173,
-0.017679225653409958,
-0.0764584019780159,
-0.07866034656763077,
-0.034936193376779556,
0.03148828074336052,
0.1420462429523468,
-0.00893460214138031,
-0.1082630529999733,
-0.07023392617702484,
-0.08515939116477966,
-0.015128357335925102,
0.04200118035078049,
0.053623393177986145,
-0.04598918929696083,
-0.02584257163107395,
0.008424208499491215,
0.0044470252469182014,
-0.02118903398513794,
0.0347011499106884,
0.08779248595237732,
0.09096184372901917,
0.04370126873254776,
-0.00778144970536232,
-0.04307117685675621,
-0.02696162462234497,
0.030647382140159607,
-0.02133149653673172,
0.04574017599225044,
-0.06366518139839172,
0.06788402050733566,
-0.07961518317461014,
-4.43458354482118e-8,
-0.008262078277766705,
0.10350990295410156,
0.026591280475258827,
0.005466398783028126,
-0.02678925171494484,
-0.054175566881895065,
0.029835229739546776,
0.07963844388723373,
-0.050455570220947266,
-0.02345765009522438,
0.0864090770483017,
-0.1065611019730568,
-0.01821564882993698,
-0.051108524203300476,
0.11934538185596466,
-0.0017247041687369347,
0.0030451470520347357,
0.07358618825674057,
-0.03453386202454567,
0.08118514716625214,
0.0329403281211853,
0.14714306592941284,
-0.022947225719690323,
-0.12067612260580063,
0.1137620285153389,
-0.012668726965785027,
0.04020414873957634,
0.08360249549150467,
0.010404269210994244,
-0.04419953376054764,
-0.04358731582760811,
0.012387718074023724,
0.03558644652366638,
-0.06679320335388184,
-0.060218170285224915,
0.0018024552846327424,
0.006631226744502783,
0.04107009619474411,
-0.0720733180642128,
0.009656098671257496,
-0.010235434398055077,
0.08111933618783951,
-0.07947119325399399,
-0.0026451165322214365,
0.031120646744966507,
0.027808567509055138,
0.06597121804952621,
-0.02242281287908554,
0.053226273506879807,
0.029545927420258522,
-0.06371680647134781,
-0.04507995396852493,
-0.019015230238437653,
0.03258490934967995,
0.057703133672475815,
0.04134771227836609,
-0.007312208879739046,
-0.05809938535094261,
0.01838338002562523,
0.10836886614561081,
0.04761061444878578,
0.00722493277862668,
-0.14569465816020966,
-0.013049825094640255
] | 0 | 0 | 0 | 0 |
|
InvestESG: A multi-agent reinforcement learning benchmark for studying climate investment as a social dilemma | https://openreview.net/forum?id=2TasVD7FXp | [
"Xiaoxuan Hou",
"Jiayi Yuan",
"Joel Z Leibo",
"Natasha Jaques"
] | Poster | **InvestESG** is a novel multi-agent reinforcement learning (MARL) benchmark designed to study the impact of Environmental, Social, and Governance (ESG) disclosure mandates on corporate climate investments. The benchmark models an intertemporal social dilemma where companies balance short-term profit losses from climate mitigation efforts and long-term benefits from reducing climate risk, while ESG-conscious investors attempt to influence corporate behavior through their investment decisions. Companies allocate capital across mitigation, greenwashing, and resilience, with varying strategies influencing climate outcomes and investor preferences. We are releasing open-source versions of InvestESG in both PyTorch and JAX, which enable scalable and hardware-accelerated simulations for investigating competing incentives in mitigate climate change. Our experiments show that without ESG-conscious investors with sufficient capital, corporate mitigation efforts remain limited under the disclosure mandate. However, when a critical mass of investors prioritizes ESG, corporate cooperation increases, which in turn reduces climate risks and enhances long-term financial stability. Additionally, providing more information about global climate risks encourages companies to invest more in mitigation, even without investor involvement. Our findings align with empirical research using real-world data, highlighting MARL's potential to inform policy by providing insights into large-scale socio-economic challenges through efficient testing of alternative policy and market designs. | multi-agent reinforcement learning, climate change, ai for climate | We introduced, InvestESG, a novel multi-agent reinforcement learning (MARL) benchmark designed to study the impact of Environmental, Social, and Governance (ESG) disclosure mandates on corporate climate investments. | 12,159 | 2411.09856 | [
-0.019860416650772095,
-0.0004431376000866294,
0.00345737230964005,
0.06301070749759674,
0.044608283787965775,
-0.010574057698249817,
-0.015940526500344276,
-0.062016744166612625,
0.06165807321667671,
0.046133238822221756,
-0.044682763516902924,
-0.048806868493556976,
-0.004397875629365444,
0.005255195777863264,
0.02749517560005188,
-0.018093783408403397,
-0.0427238829433918,
0.04736523702740669,
-0.07063222676515579,
-0.06489866226911545,
-0.0003056999121326953,
-0.07068053632974625,
0.0064732227474451065,
0.015267369337379932,
0.042065903544425964,
-0.045596685260534286,
-0.015252288430929184,
0.064490407705307,
-0.035224661231040955,
0.06653329730033875,
0.03946460410952568,
-0.03451581671833992,
-0.014566615223884583,
-0.027342820540070534,
-0.008979463949799538,
0.09472332894802094,
-0.08933117985725403,
-0.035390455275774,
-0.020772811025381088,
0.001150984549894929,
-0.01937023364007473,
0.011697310023009777,
0.02773713506758213,
-0.06520729511976242,
0.033083368092775345,
-0.03908015042543411,
0.049825649708509445,
-0.05451507866382599,
0.004817023407667875,
-0.0411057248711586,
-0.016401780769228935,
-0.05532420799136162,
-0.00450734281912446,
-0.039432886987924576,
-0.04537560045719147,
0.0003608058614190668,
0.008995824493467808,
0.004961702041327953,
-0.023273024708032608,
-0.07493120431900024,
0.06271248310804367,
-0.14287127554416656,
-0.01038832776248455,
0.039040278643369675,
0.012582749128341675,
0.0528901182115078,
-0.026003357023000717,
0.13847553730010986,
-0.05337221920490265,
-0.03392557054758072,
0.04420933499932289,
-0.0110021336004138,
-0.03723730891942978,
-0.07147862017154694,
-0.05479349568486214,
0.023767732083797455,
-0.013732432387769222,
0.07662118971347809,
0.11316774040460587,
-0.0473809577524662,
0.029694730415940285,
-0.02829093672335148,
-0.013488314114511013,
0.007087414618581533,
0.0008985454333014786,
-0.02570480667054653,
0.09084147959947586,
0.06503655016422272,
0.12391689419746399,
0.07418357580900192,
0.0060083805583417416,
0.010480986908078194,
0.07004012912511826,
0.015528975054621696,
-0.09789504110813141,
0.07022102177143097,
-0.0055135274305939674,
-0.07335885614156723,
-0.1388188600540161,
0.08488239347934723,
0.030758177861571312,
0.039038196206092834,
-0.022845834493637085,
-0.027641598135232925,
0.011740980669856071,
-0.048114702105522156,
-0.0021758300717920065,
-0.05109988898038864,
0.06520295888185501,
-0.038980208337306976,
-0.06485878676176071,
0.06165669113397598,
0.022208472713828087,
0.00714767025783658,
0.03610733896493912,
0.056612517684698105,
-0.0338987372815609,
0.008169637061655521,
0.08975423872470856,
0.014658487401902676,
0.01938208006322384,
-0.03197968378663063,
0.012590749189257622,
0.013100452721118927,
0.05046139657497406,
0.029891731217503548,
-0.05426116660237312,
7.76131624567435e-34,
0.033146053552627563,
0.03947325795888901,
0.0326823852956295,
0.0673767700791359,
0.07253451645374298,
0.018151793628931046,
0.05515461415052414,
0.0074628796428442,
-0.11262478679418564,
-0.014327053911983967,
0.05924418568611145,
0.012667960487306118,
-0.04947338625788689,
0.09500636160373688,
0.015807894989848137,
-0.04766945168375969,
-0.09305756539106369,
0.05074865370988846,
0.059034667909145355,
-0.06993375718593597,
0.03631008043885231,
-0.04180777445435524,
0.025459440425038338,
0.007364536635577679,
0.04744797572493553,
-0.028211066499352455,
0.027882281690835953,
0.0018554871203377843,
0.04163280874490738,
0.031776778399944305,
0.05216103419661522,
0.08979078382253647,
-0.05467412620782852,
0.03263844922184944,
0.015214689075946808,
0.018059279769659042,
-0.09624451398849487,
-0.02973858453333378,
0.05088873207569122,
0.05823542922735214,
-0.05546560510993004,
-0.0036863600835204124,
0.020571118220686913,
-0.044980183243751526,
-0.015922004356980324,
-0.06499812006950378,
0.003701435402035713,
-0.027783310040831566,
-0.04852086305618286,
-0.06884010136127472,
-0.01905653066933155,
0.05714661255478859,
0.03467267006635666,
-0.08421595394611359,
-0.007643408607691526,
-0.002915993332862854,
0.07325253635644913,
-0.03687162324786186,
-0.0991443321108818,
-0.04144541919231415,
-0.020510951057076454,
0.011023757979273796,
-0.00044847140088677406,
0.07812642306089401,
-0.059499770402908325,
0.1661437451839447,
-0.028403790667653084,
0.10213816165924072,
-0.031673364341259,
0.005468868650496006,
0.018404025584459305,
-0.002491750055924058,
-0.009654584340751171,
-0.11093772202730179,
0.007798970676958561,
-0.0528099350631237,
0.03711892291903496,
-0.01663426123559475,
0.02802300825715065,
0.03008222207427025,
0.028465352952480316,
-0.013036269694566727,
0.033590104430913925,
0.010464617051184177,
-0.08394207805395126,
-0.07915253192186356,
0.020980341359972954,
0.03110748529434204,
0.07510939240455627,
0.010994602926075459,
-0.049425166100263596,
-0.012468507513403893,
0.06687997281551361,
0.045179206877946854,
-0.029691962525248528,
-1.4926912158161121e-33,
-0.061955470591783524,
-0.05235330015420914,
-0.01936539076268673,
-0.031767696142196655,
-0.014073568396270275,
-0.009525753557682037,
-0.05170195549726486,
-0.1281765252351761,
0.01993023417890072,
-0.04492173343896866,
-0.03348187729716301,
0.04098633676767349,
0.017314989119768143,
-0.015043462626636028,
0.004349558148533106,
-0.0912344753742218,
0.021262818947434425,
-0.04800129681825638,
0.036024682223796844,
-0.0821627750992775,
0.04541052505373955,
-0.047915320843458176,
0.0065008909441530704,
0.11157704889774323,
0.019785013049840927,
-0.003232179209589958,
-0.011436646804213524,
0.04273576661944389,
0.07147318869829178,
0.024830861017107964,
-0.07715785503387451,
0.03181241825222969,
-0.0335279181599617,
0.016326431185007095,
-0.039442695677280426,
0.07112424075603485,
0.019494900479912758,
-0.07274481654167175,
-0.04406699910759926,
0.04397299140691757,
0.04418575018644333,
-0.11845643818378448,
-0.0070601156912744045,
-0.013800399377942085,
0.011492623947560787,
0.006314375437796116,
-0.014395038597285748,
-0.04507918655872345,
0.04479427635669708,
0.0655750185251236,
0.011002683080732822,
0.012699926272034645,
-0.07936390489339828,
0.029493695124983788,
-0.07283356040716171,
-0.03593312203884125,
-0.00755825312808156,
-0.0025126533582806587,
0.02462957613170147,
0.04307687655091286,
-0.08156184107065201,
0.03164948895573616,
0.011682313866913319,
0.10402737557888031,
-0.08278300613164902,
-0.018176386132836342,
-0.050361622124910355,
0.026084428653120995,
0.03665876016020775,
-0.04420201852917671,
0.0855756476521492,
0.006176982074975967,
-0.043360721319913864,
-0.0840916782617569,
0.06726976484060287,
0.08725812286138535,
-0.040942464023828506,
-0.01584065519273281,
-0.08665569871664047,
0.009028861299157143,
0.052098046988248825,
0.09463649243116379,
0.0569850355386734,
0.03276219964027405,
0.041738804429769516,
-0.0027105137705802917,
0.027216320857405663,
0.05444934964179993,
-0.02148015983402729,
0.03386117145419121,
-0.058052077889442444,
0.011459428817033768,
0.04295647144317627,
0.03144918382167816,
-0.06593864411115646,
-4.196385461341379e-8,
-0.03602967411279678,
-0.0013014146825298667,
0.09140053391456604,
-0.04585832357406616,
0.06135732680559158,
-0.0655733272433281,
0.02754727564752102,
-0.010313285514712334,
0.000766182376537472,
-0.003362624906003475,
0.051314014941453934,
0.04520776867866516,
-0.0158036220818758,
0.021533915773034096,
-0.08727895468473434,
-0.029680300503969193,
-0.023862477391958237,
-0.018124686554074287,
-0.02622145228087902,
-0.008406473323702812,
0.05486282706260681,
-0.00005726444942411035,
0.004058282822370529,
-0.04940183833241463,
0.0571940541267395,
-0.032280538231134415,
-0.025797925889492035,
0.08170291781425476,
0.09746335446834564,
0.07691959291696548,
-0.011477718129754066,
-0.05584117770195007,
0.06573542952537537,
0.0007741393055766821,
-0.012984765693545341,
0.07780475169420242,
0.04753224924206734,
0.006508915685117245,
0.031424183398485184,
0.016773678362369537,
-0.07003182917833328,
0.07625073194503784,
-0.0019029647810384631,
-0.01892232708632946,
-0.010511462576687336,
-0.006207113154232502,
-0.12224496155977249,
-0.020031841471791267,
0.08008664846420288,
-0.08572706580162048,
0.026557695120573044,
-0.05344095081090927,
-0.034395020455121994,
0.08906757086515427,
0.03488970175385475,
-0.017947746440768242,
-0.02356714941561222,
0.01496014092117548,
0.04365202784538269,
0.03362443298101425,
-0.02107490971684456,
-0.08998095989227295,
-0.12612508237361908,
0.025116339325904846
] | https://github.com/yuanjiayiy/InvestESG | 6 | 0 | 0 | 0 |
Discrete GCBF Proximal Policy Optimization for Multi-agent Safe Optimal Control | https://openreview.net/forum?id=1X1R7P6yzt | [
"Songyuan Zhang",
"Oswin So",
"Mitchell Black",
"Chuchu Fan"
] | Poster | Control policies that can achieve high task performance and satisfy safety constraints are desirable for any system, including multi-agent systems (MAS). One promising technique for ensuring the safety of MAS is distributed control barrier functions (CBF). However, it is difficult to design distributed CBF-based policies for MAS that can tackle unknown discrete-time dynamics, partial observability, changing neighborhoods, and input constraints, especially when a distributed high-performance nominal policy that can achieve the task is unavailable. To tackle these challenges, we propose **DGPPO**, a new framework that *simultaneously* learns both a *discrete* graph CBF which handles neighborhood changes and input constraints, and a distributed high-performance safe policy for MAS with unknown discrete-time dynamics.
We empirically validate our claims on a suite of multi-agent tasks spanning three different simulation engines. The results suggest that, compared with existing methods, our DGPPO framework obtains policies that achieve high task performance (matching baselines that ignore the safety constraints), and high safety rates (matching the most conservative baselines), with a *constant* set of hyperparameters across all environments. | control barrier functions, multi-agent systems, black-box systems, partial observability, reinforcement learning | We propose DGPPO for solving multi-agent safe optimal control problem with unknown discrete-time dynamics, partial observability, changing neighborhoods, and input constraints, without a known performant nominal policy. | 12,158 | 2502.03640 | [
-0.014218403957784176,
-0.04431362822651863,
-0.07497365772724152,
-0.038225989788770676,
0.04916703328490257,
0.0001723900786601007,
-0.07508126646280289,
-0.06397229433059692,
-0.012791109271347523,
0.05024449899792671,
-0.08203403651714325,
-0.05593300983309746,
0.028134994208812714,
-0.013106332160532475,
0.0054357051849365234,
0.03517061099410057,
0.15038712322711945,
0.014418023638427258,
-0.10463808476924896,
-0.014173966832458973,
0.004601836204528809,
0.032280877232551575,
0.03699229285120964,
-0.04911438748240471,
-0.050256919115781784,
-0.040166616439819336,
-0.09113026410341263,
-0.015481658279895782,
-0.01193501427769661,
-0.02515660785138607,
0.013348647393286228,
-0.10193414986133575,
-0.05864395946264267,
-0.06472833454608917,
0.02095678262412548,
0.06865552067756653,
-0.11170490831136703,
0.0071129086427390575,
0.016160914674401283,
0.005773521959781647,
-0.012823696248233318,
0.043520912528038025,
-0.044115643948316574,
-0.00568918464705348,
0.03486575186252594,
0.04713578149676323,
0.013243255205452442,
-0.008539094589650631,
0.010593111626803875,
-0.04238804429769516,
-0.011992431245744228,
-0.03147570416331291,
0.05417827144265175,
0.03134559094905853,
0.06804630905389786,
-0.03980490565299988,
0.06074214354157448,
-0.020157279446721077,
0.009097039699554443,
-0.001365107949823141,
-0.08924601972103119,
-0.09418331831693649,
-0.02312885969877243,
-0.014775795862078667,
-0.027025751769542694,
0.040290430188179016,
-0.02877551130950451,
0.04566441848874092,
0.002400228288024664,
-0.005760735832154751,
-0.07161851227283478,
0.011083850637078285,
-0.00022151901794131845,
-0.0180952325463295,
0.051986370235681534,
0.02827046997845173,
-0.013725167140364647,
0.03032633662223816,
0.015119360759854317,
-0.04426293447613716,
0.09583848714828491,
0.07161898910999298,
-0.00816736277192831,
0.03416934236884117,
0.045934274792671204,
-0.05290786921977997,
-0.029827918857336044,
0.03632066026329994,
0.1297607719898224,
0.0052791014313697815,
0.017508499324321747,
0.044474873691797256,
0.05594966560602188,
0.096530020236969,
0.03174394369125366,
-0.08210642635822296,
-0.015723388642072678,
-0.06950123608112335,
-0.02525009773671627,
0.06303327530622482,
-0.017541896551847458,
-0.017026588320732117,
0.05919262766838074,
0.03583838790655136,
0.0525595024228096,
0.0018513232935220003,
-0.0038697519339621067,
0.04795973002910614,
0.0189457219094038,
-0.05610766261816025,
0.05882495269179344,
0.0856487974524498,
0.0248770359903574,
0.005375008098781109,
-0.047085970640182495,
-0.01744762621819973,
0.014534810557961464,
-0.033368971198797226,
0.03456633538007736,
0.004964808467775583,
-0.038646452128887177,
-0.05262744426727295,
0.09506825357675552,
-0.019691379740834236,
0.05259008705615997,
0.008596847765147686,
0.012641317211091518,
-8.295620672451313e-34,
0.031261146068573,
0.0025386407505720854,
0.00950078759342432,
-0.06571721285581589,
0.0564693920314312,
-0.005516359116882086,
-0.03940041735768318,
0.006429264787584543,
-0.037907324731349945,
0.001609256025403738,
0.012828315608203411,
-0.05172478407621384,
-0.06705252081155777,
-0.0040519661270082,
0.07459519803524017,
0.008311513811349869,
-0.019550830125808716,
0.05091402679681778,
0.03440278396010399,
-0.058108195662498474,
0.0919041708111763,
-0.05155959725379944,
-0.048137959092855453,
-0.030700836330652237,
0.034499675035476685,
0.05106275528669357,
0.015927638858556747,
-0.04460886865854263,
0.08387939631938934,
0.04476906359195709,
-0.12408750504255295,
0.039322901517152786,
-0.004724303726106882,
0.12728284299373627,
-0.055992767214775085,
0.08494739979505539,
-0.1292385309934616,
-0.06596503406763077,
-0.051604971289634705,
-0.0010360567830502987,
-0.06985297799110413,
-0.06382890790700912,
0.07009805738925934,
-0.027999762445688248,
-0.050004418939352036,
-0.11554165184497833,
-0.012857655994594097,
0.008589695207774639,
-0.08010726422071457,
-0.0552511103451252,
-0.03579534590244293,
0.01423775963485241,
0.03949308395385742,
-0.08867809921503067,
-0.06329083442687988,
0.008145592175424099,
0.05153951421380043,
0.05278107151389122,
-0.024656983092427254,
0.12392587214708328,
-0.01715339720249176,
-0.006373487412929535,
-0.03948681801557541,
0.058406662195920944,
0.0056842598132789135,
0.05233023688197136,
-0.12481290102005005,
0.03930278867483139,
0.10633207112550735,
-0.0011902488768100739,
-0.03115268237888813,
0.03659462928771973,
0.05126022547483444,
-0.04436429589986801,
0.05378201976418495,
-0.033109139651060104,
0.10193212330341339,
-0.03633881360292435,
-0.012732254341244698,
-0.031752925366163254,
0.0010162636172026396,
0.0533013753592968,
-0.059075042605400085,
0.0116861118003726,
0.010273566469550133,
-0.09764326363801956,
-0.059920016676187515,
0.03920843452215195,
-0.0655921921133995,
-0.022158687934279442,
-0.11570311337709427,
-0.02834389917552471,
0.015599404461681843,
0.06603988260030746,
0.014924153685569763,
-3.454597669905493e-34,
-0.0018630549311637878,
0.011147656477987766,
-0.00819386262446642,
0.054168879985809326,
0.0010624571004882455,
-0.04004399850964546,
-0.05667577683925629,
-0.04968215152621269,
0.010031287558376789,
-0.03198038786649704,
-0.10177206248044968,
-0.00758537370711565,
0.029493030160665512,
-0.009700640104711056,
-0.012361472472548485,
-0.0484234020113945,
0.008433829993009567,
-0.007595567498356104,
0.0006370199844241142,
-0.013242595829069614,
0.04087214544415474,
0.023638609796762466,
-0.1071745902299881,
0.065037801861763,
0.0587322823703289,
0.0449170283973217,
-0.02180028147995472,
0.033368051052093506,
0.05223208665847778,
0.06364034861326218,
-0.05575397610664368,
-0.0036179234739392996,
-0.035788122564554214,
0.02491285465657711,
0.00800473615527153,
0.011675536632537842,
0.03635171428322792,
0.055550165474414825,
-0.06336136907339096,
0.07280107587575912,
0.01768975704908371,
-0.050063204020261765,
-0.05227520689368248,
0.009339040145277977,
0.04253574460744858,
0.0935884490609169,
-0.029560549184679985,
0.03346707299351692,
0.01733098179101944,
0.03188072517514229,
-0.07285497337579727,
-0.007240730803459883,
-0.03234182298183441,
-0.00023402186343446374,
-0.04733993113040924,
-0.09046082198619843,
0.07953576743602753,
0.04656849801540375,
0.02714022621512413,
-0.011176664382219315,
-0.024144433438777924,
0.01901049166917801,
-0.03036990948021412,
0.10447632521390915,
0.017614351585507393,
0.054377567023038864,
-0.01956554502248764,
-0.030208002775907516,
0.048226963728666306,
0.0037996736355125904,
0.03163944184780121,
0.0627308040857315,
-0.06336735188961029,
-0.03534683585166931,
0.05965530872344971,
0.005493843927979469,
-0.04035083204507828,
-0.05008760094642639,
-0.02801884151995182,
0.05866168066859245,
0.04577861353754997,
0.058536726981401443,
0.06499698758125305,
0.10607995837926865,
0.012610038742423058,
0.06252633035182953,
0.0040549589321017265,
0.05874292179942131,
0.0679120197892189,
-0.028065672144293785,
-0.06797336041927338,
0.0487789511680603,
-0.05991525202989578,
0.03819459304213524,
-0.10190799087285995,
-4.2632315455648495e-8,
-0.020592045038938522,
0.08585310727357864,
-0.0060251387767493725,
-0.04203914478421211,
0.018363934010267258,
0.06585698574781418,
0.01162433996796608,
-0.06969072669744492,
0.017341498285531998,
0.049017611891031265,
0.09142374247312546,
0.059671007096767426,
0.049156226217746735,
-0.09823144227266312,
-0.09327930212020874,
0.012360290624201298,
-0.017760345712304115,
-0.0755700096487999,
0.015314696356654167,
0.018833959475159645,
0.016820525750517845,
0.016589760780334473,
-0.020475298166275024,
-0.02989080362021923,
0.005853246431797743,
-0.07509588450193405,
-0.010865161195397377,
-0.0049833268858492374,
0.06250949203968048,
0.08204621076583862,
-0.012873713858425617,
-0.0015638440381735563,
0.043970320373773575,
0.03358246013522148,
-0.05243559181690216,
0.062339216470718384,
0.025068409740924835,
-0.010441414080560207,
0.05921724811196327,
-0.03962677717208862,
0.0004708954074885696,
0.03153017535805702,
-0.04678284749388695,
-0.019024111330509186,
-0.002201769035309553,
0.008190862834453583,
-0.05358659848570824,
-0.008293856866657734,
0.07858414202928543,
-0.005177249666303396,
0.004559143912047148,
0.013319514691829681,
-0.022128988057374954,
0.1159360259771347,
0.08972843736410141,
-0.021809915080666542,
-0.005315424408763647,
-0.11015975475311279,
0.1292209029197693,
0.05616071820259094,
-0.018845906481146812,
-0.005022143479436636,
-0.041997648775577545,
-0.0042912596836686134
] | https://github.com/MIT-REALM/dgppo | 14 | 0 | 0 | 0 |
Aligning Language Models with Demonstrated Feedback | https://openreview.net/forum?id=1qGkuxI9UX | [
"Omar Shaikh",
"Michelle S. Lam",
"Joey Hejna",
"Yijia Shao",
"Hyundong Justin Cho",
"Michael S. Bernstein",
"Diyi Yang"
] | Poster | Language models are aligned to emulate the collective voice of many, resulting in outputs that align with no one in particular. Steering LLMs away from generic output is possible through supervised finetuning or RLHF, but requires prohibitively large datasets for new ad-hoc tasks. We argue that it is instead possible to align an LLM to a specific setting by leveraging a very small number ($<10$) of demonstrations as feedback. Our method, Demonstration ITerated Task Optimization (DITTO), directly aligns language model outputs to a user's demonstrated behaviors. Derived using ideas from online imitation learning, DITTO cheaply generates online comparison data by treating users' demonstrations as preferred over output from the LLM and its intermediate checkpoints. We evaluate DITTO's ability to learn fine-grained style and task alignment across domains such as news articles, emails, and blog posts. Additionally, we conduct a user study soliciting a range of demonstrations from participants ($N=16$). Across our benchmarks and user study, we find that win-rates for DITTO outperform few-shot prompting, supervised fine-tuning, and other self-play methods by an average of 19\% points. By using demonstrations as feedback directly, DITTO offers a novel method for effective customization of LLMs. | personalization, few-shot learning, human computer interaction, alignment | We highlight the effectiveness of using a very small number of demonstrations (<10) for task or user-specific alignment; and contribute a method that iteratively aligns an LLM to a user’s demonstrations by treating default outputs as dispreferred. | 12,157 | 2406.00888 | [
-0.06383488327264786,
-0.12353011965751648,
0.08144573867321014,
-0.038544680923223495,
0.05487925186753273,
-0.0368536040186882,
0.0074034184217453,
0.0057609472423791885,
0.036953069269657135,
-0.008111993782222271,
-0.05229723080992699,
-0.07454530149698257,
0.09112553298473358,
0.001011700602248311,
-0.0019077390898019075,
-0.047117218375205994,
0.17402604222297668,
0.06102490425109863,
-0.05496476963162422,
-0.050082385540008545,
-0.035368360579013824,
0.004218533635139465,
0.0803176686167717,
0.03395378962159157,
-0.010673786513507366,
0.0033930728677660227,
-0.041278988122940063,
-0.01015466172248125,
0.026367496699094772,
-0.050956908613443375,
-0.004679196048527956,
-0.00850069522857666,
-0.006001257803291082,
-0.006976868491619825,
-0.06554225087165833,
0.06831903755664825,
-0.09043385833501816,
-0.003797322977334261,
0.019883090630173683,
-0.07929331809282303,
-0.0626198798418045,
-0.03287787362933159,
0.026325447484850883,
0.02077106386423111,
0.02423456870019436,
-0.0036935184616595507,
-0.03754077106714249,
0.025737302377820015,
-0.02252083644270897,
0.07171379774808884,
-0.10079028457403183,
-0.035090889781713486,
0.07150889188051224,
-0.0009573331917636096,
-0.05623146519064903,
0.03990664333105087,
0.05987001210451126,
0.06897667050361633,
0.043671149760484695,
-0.0717090368270874,
0.012213898822665215,
-0.13551747798919678,
-0.029991842806339264,
-0.002282902365550399,
-0.029529685154557228,
-0.07386128604412079,
-0.005198449362069368,
0.0532509870827198,
-0.01641318015754223,
0.08743658661842346,
0.014444815926253796,
-0.0033129581715911627,
0.04758760333061218,
0.07720938324928284,
0.042424071580171585,
0.05276506394147873,
-0.006778824608772993,
-0.05713019147515297,
-0.04755852371454239,
-0.023794030770659447,
0.026093216612935066,
-0.02603110484778881,
0.006775478832423687,
-0.004713357891887426,
0.06725408881902695,
-0.02750399149954319,
-0.016141874715685844,
0.026071414351463318,
0.061425335705280304,
-0.026247132569551468,
-0.10108431428670883,
-0.024103112518787384,
-0.011212782934308052,
-0.04186791181564331,
-0.002149502979591489,
0.023568052798509598,
0.05472534894943237,
-0.04805109277367592,
-0.020955195650458336,
0.10626766830682755,
0.013016066513955593,
0.1384829729795456,
0.015987353399395943,
-0.0991557240486145,
-0.028930483385920525,
-0.043997302651405334,
-0.03235078975558281,
0.05300431326031685,
0.05480427294969559,
-0.07707514613866806,
0.037167515605688095,
0.06405190378427505,
0.009175974875688553,
0.08051977306604385,
0.10641700774431229,
-0.07651635259389877,
0.010077386163175106,
-0.05407320335507393,
-0.028327802196145058,
0.07761713117361069,
0.0534641407430172,
0.010157963261008263,
0.022509740665555,
0.0010325820185244083,
0.032484397292137146,
-0.017637483775615692,
0.017510637640953064,
2.3429121370994643e-33,
0.16342122852802277,
0.06844368577003479,
0.06990814954042435,
0.05518055334687233,
0.0988604798913002,
-0.00702511053532362,
0.03651072829961777,
0.0033160531893372536,
-0.056452758610248566,
-0.04508746787905693,
0.037066277116537094,
0.014815513975918293,
0.003305998397991061,
0.07829466462135315,
0.000766070035751909,
-0.023205794394016266,
-0.06715673953294754,
0.022800400853157043,
-0.06594949960708618,
0.008614880032837391,
0.05272320657968521,
-0.009876332245767117,
-0.035822439938783646,
-0.061928387731313705,
0.03962618112564087,
-0.044647425413131714,
-0.005777783691883087,
-0.026854515075683594,
-0.04878566414117813,
-0.00016001974290702492,
-0.09204801172018051,
-0.03613044321537018,
-0.03683768957853317,
0.06629475206136703,
0.009945358149707317,
-0.015572371892631054,
0.01716630533337593,
-0.13828028738498688,
0.004917826037853956,
-0.01047272514551878,
-0.009164704941213131,
0.008380846120417118,
0.020586155354976654,
0.03939620405435562,
-0.08155201375484467,
-0.05804619938135147,
0.01995030790567398,
0.011522204615175724,
-0.0413784384727478,
-0.03434671461582184,
0.049025729298591614,
0.01798415556550026,
0.020757928490638733,
-0.0660981833934784,
-0.024868737906217575,
-0.007622412405908108,
0.0016270108753815293,
-0.026286257430911064,
0.021086305379867554,
0.01107145193964243,
0.02296178601682186,
0.02715204656124115,
-0.033243514597415924,
0.03455950319766998,
-0.001553037203848362,
0.07191266119480133,
-0.005867323372513056,
0.035494156181812286,
0.08259139209985733,
0.04704493656754494,
-0.018109509721398354,
-0.03412275388836861,
-0.0792318731546402,
-0.03436960279941559,
0.06835301220417023,
0.026698511093854904,
0.04447874426841736,
-0.11128171533346176,
0.03815952315926552,
0.04047210514545441,
0.06527507305145264,
-0.016283899545669556,
-0.0352453775703907,
-0.04134627804160118,
-0.03624351695179939,
-0.11284621804952621,
0.04935751482844353,
-0.08595461398363113,
-0.026230016723275185,
-0.0035661282017827034,
-0.05428085848689079,
-0.018391642719507217,
0.00012791565677616745,
0.005411188583821058,
0.03654675930738449,
-1.843320058502284e-33,
-0.005720718298107386,
-0.022081216797232628,
0.028563285246491432,
0.10940408706665039,
-0.06517539173364639,
0.028322963044047356,
0.053303539752960205,
0.0031489022076129913,
-0.06532282382249832,
-0.0729118064045906,
-0.08265780657529831,
-0.006103950086981058,
0.017995372414588928,
-0.018321141600608826,
-0.03389911726117134,
-0.023860212415456772,
0.06606513261795044,
0.05158890038728714,
0.041522037237882614,
0.008383155800402164,
0.00970806460827589,
0.08834412693977356,
-0.04366244003176689,
0.02713748812675476,
-0.03889159485697746,
-0.0035209173802286386,
0.023757733404636383,
0.052315689623355865,
0.0066129546612501144,
-0.015246608294546604,
0.02176554873585701,
-0.023431159555912018,
-0.004705032333731651,
-0.013348426669836044,
-0.053625669330358505,
0.08037715405225754,
0.010717695578932762,
0.033434320241212845,
-0.005460310727357864,
0.056167494505643845,
-0.006337442900985479,
-0.04722096398472786,
-0.022869188338518143,
0.012892414815723896,
-0.04377169534564018,
-0.03323792293667793,
-0.10431379824876785,
-0.006952269002795219,
-0.047751761972904205,
0.010578697547316551,
0.010403577238321304,
0.0516926608979702,
-0.06814871728420258,
-0.025463001802563667,
-0.05619001388549805,
-0.08985251933336258,
-0.013016859069466591,
-0.007968549616634846,
0.02449144423007965,
-0.007006603758782148,
-0.0757061094045639,
0.052707552909851074,
0.010950693860650063,
-0.012513002380728722,
0.034036144614219666,
-0.08445905148983002,
-0.01557895913720131,
0.051160652190446854,
-0.009562816470861435,
-0.011704192496836185,
0.05281006917357445,
-0.0057183546014130116,
-0.012420524843037128,
-0.02401346154510975,
0.048843804746866226,
-0.03068128041923046,
0.040444593876600266,
0.026186427101492882,
-0.018927108496427536,
-0.04318831488490105,
0.02630751021206379,
0.014168351888656616,
0.05747350677847862,
0.019364729523658752,
0.04388507083058357,
0.09539958834648132,
0.014221972785890102,
0.13876640796661377,
-0.021648632362484932,
0.08386630564928055,
0.022989004850387573,
0.027206022292375565,
0.026539240032434464,
0.04828447476029396,
0.02096964232623577,
-4.603600345376435e-8,
-0.09038262069225311,
0.04292232170701027,
0.07396488636732101,
0.09211480617523193,
-0.03246458247303963,
-0.019973820075392723,
0.01139671541750431,
-0.020771753042936325,
-0.01676221564412117,
-0.052444275468587875,
0.04566524922847748,
0.023809950798749924,
-0.0076910327188670635,
-0.018834426999092102,
-0.001683161361142993,
0.0345633290708065,
-0.029769431799650192,
-0.015508058480918407,
-0.046011507511138916,
0.05150294303894043,
0.0461101233959198,
0.0494302473962307,
-0.016351893544197083,
-0.051431916654109955,
0.037091560661792755,
-0.015017842873930931,
-0.02838524803519249,
0.07765533775091171,
0.048169154673814774,
-0.06797128170728683,
-0.030889064073562622,
-0.041586339473724365,
-0.0286647267639637,
-0.060013413429260254,
0.019745642319321632,
0.013175073079764843,
0.02746530808508396,
-0.030552860349416733,
0.015394136309623718,
-0.06073402985930443,
-0.039961013942956924,
0.05272020772099495,
0.018635110929608345,
0.02278616465628147,
0.01662636734545231,
-0.004025150556117296,
-0.08391635864973068,
-0.18543533980846405,
0.0567447803914547,
-0.0782652273774147,
-0.015484148636460304,
-0.08482581377029419,
0.022954970598220825,
0.0571044385433197,
0.11127108335494995,
0.06313575804233551,
-0.059172142297029495,
-0.0019884128123521805,
0.04147839918732643,
0.148407444357872,
0.050858475267887115,
0.07151937484741211,
-0.05999495089054108,
-0.0015891881193965673
] | https://github.com/SALT-NLP/demonstrated-feedback | 120 | 0 | 0 | 0 |
See It from My Perspective: How Language Affects Cultural Bias in Image Understanding | https://openreview.net/forum?id=Xbl6t6zxZs | [
"Amith Ananthram",
"Elias Stengel-Eskin",
"Mohit Bansal",
"Kathleen McKeown"
] | Poster | Vision-language models (VLMs) can respond to queries about images in many languages. However, beyond language, culture affects how we see things. For example, individuals from Western cultures focus more on the central figure in an image while individuals from East Asian cultures attend more to scene context (Nisbett 2001). In this work, we characterize the Western bias of VLMs in image understanding and investigate the role that language plays in this disparity. We evaluate VLMs across subjective and objective visual tasks with culturally diverse images and annotations. We find that VLMs perform better on the Western split than on the East Asian split of each task. Through controlled experimentation, we trace one source of this bias in image understanding to the lack of diversity in language model construction. While inference in a language nearer to a culture can lead to reductions in bias, we show it is much more effective when that language was well-represented during text-only pre-training. Interestingly, this yields bias reductions even when prompting in English. Our work highlights the importance of richer representation of all languages in building equitable VLMs. | vision-language models, multilinguality, cultural bias, vqa, emotion classification, art | null | 12,153 | 2406.11665 | [
0.029282692819833755,
0.017516890540719032,
0.012973283417522907,
-0.020888308063149452,
0.10502398014068604,
-0.0009762378176674247,
0.024438487365841866,
-0.023637624457478523,
0.09041576087474823,
-0.04280881956219673,
-0.02962007001042366,
-0.1283564418554306,
0.037020716816186905,
0.07525857537984848,
0.04066798835992813,
-0.032626453787088394,
0.06765743345022202,
0.029773734509944916,
-0.13832198083400726,
-0.06804699450731277,
-0.02509583719074726,
-0.05103921890258789,
0.11405559629201889,
-0.003387462580576539,
0.017517320811748505,
-0.013618133962154388,
0.027403799816966057,
-0.04525589942932129,
0.02279006317257881,
-0.045458875596523285,
-0.0215121079236269,
0.05125874653458595,
0.04237818345427513,
0.07592152059078217,
-0.05309951677918434,
0.08056089282035828,
-0.0696183517575264,
0.016274617984890938,
-0.005997698754072189,
-0.054954931139945984,
-0.07825013250112534,
0.03237522765994072,
0.08531215786933899,
-0.02753531187772751,
0.08272599428892136,
0.005890038795769215,
-0.010364647954702377,
-0.012362678535282612,
-0.044266488403081894,
0.036502111703157425,
-0.10930772125720978,
-0.02163448929786682,
0.007986827753484249,
-0.0070298113860189915,
-0.05235612019896507,
-0.00477339094504714,
0.05892257019877434,
0.042848456650972366,
0.008171025663614273,
0.02846134826540947,
-0.06458381563425064,
-0.07457146048545837,
-0.01897812820971012,
0.06695378571748734,
-0.03671925142407417,
-0.04318786412477493,
0.03571295365691185,
-0.021469874307513237,
-0.058131273835897446,
-0.07914785295724869,
0.009955370798707008,
0.045158226042985916,
0.020059561356902122,
0.027740823104977608,
0.03647333383560181,
0.05808183550834656,
0.01872246339917183,
-0.10692686587572098,
-0.008941907435655594,
-0.15464691817760468,
0.09960538893938065,
0.046327054500579834,
0.05216704681515694,
-0.03837445750832558,
0.1248798668384552,
-0.012138880789279938,
-0.03600421920418739,
-0.0030027031898498535,
0.000468147627543658,
-0.002185015706345439,
-0.07523245364427567,
-0.07528753578662872,
0.02402879111468792,
0.03981855511665344,
0.11118855327367783,
-0.0033659038599580526,
-0.010177914053201675,
-0.02814599685370922,
-0.014698702841997147,
0.08204242587089539,
0.0708610936999321,
0.028317824006080627,
0.022067129611968994,
0.00685263704508543,
0.002263270318508148,
-0.03480589762330055,
0.04065779596567154,
0.013879320584237576,
-0.05854576453566551,
-0.08602605015039444,
0.01117781363427639,
0.040242161601781845,
-0.051148466765880585,
0.04599018767476082,
-0.003814948257058859,
-0.08889409154653549,
0.06212650611996651,
-0.02706197090446949,
0.00937999039888382,
-0.014819550327956676,
-0.052710726857185364,
0.007315755356103182,
-0.005421079229563475,
-0.04133483022451401,
0.01851126179099083,
-0.049636177718639374,
-0.004403345286846161,
2.4945637653250015e-33,
0.08518051356077194,
0.0057607353664934635,
0.07913417369127274,
0.010034959763288498,
0.0587453618645668,
-0.05050759017467499,
-0.013161805458366871,
0.01177778746932745,
-0.03814452141523361,
-0.0633077621459961,
0.025536099448800087,
0.014130067080259323,
-0.037574443966150284,
0.09062999486923218,
0.008345428854227066,
0.02937743440270424,
-0.05335386097431183,
0.0531478226184845,
-0.04881763085722923,
0.06292223930358887,
0.0281577929854393,
0.030290404334664345,
0.022963859140872955,
-0.05804207921028137,
-0.052421119064092636,
-0.001985171576961875,
0.056765004992485046,
-0.0874667763710022,
-0.05646839365363121,
0.01214647851884365,
-0.12641626596450806,
0.017432181164622307,
0.03338184207677841,
0.06437793374061584,
0.021938828751444817,
-0.03425385057926178,
0.014530066400766373,
-0.006567270494997501,
0.02501482143998146,
0.013916468247771263,
0.039766594767570496,
0.056708503514528275,
0.002007117262110114,
-0.022291280329227448,
-0.03994748741388321,
-0.010345354676246643,
0.02945070154964924,
0.021821394562721252,
-0.10405148565769196,
0.05444875359535217,
-0.006558781489729881,
0.01771569810807705,
0.05338497459888458,
-0.032461151480674744,
0.05849037319421768,
0.0215460192412138,
0.07856443524360657,
-0.024419335648417473,
0.017856415361166,
-0.07836835831403732,
-0.020328087732195854,
0.03575282543897629,
-0.003446227638050914,
0.09334595501422882,
0.040264759212732315,
0.029477788135409355,
0.006134556140750647,
-0.019278327003121376,
-0.018197737634181976,
-0.002601797226816416,
-0.027316447347402573,
-0.009601214900612831,
-0.03133511170744896,
-0.021460779011249542,
0.03435627371072769,
0.037014223635196686,
0.04237139970064163,
-0.074517622590065,
0.03634010627865791,
0.02535432018339634,
0.005322540644556284,
-0.03894603252410889,
0.011991934850811958,
-0.05476488173007965,
-0.07130257785320282,
-0.0027628911193460226,
0.05582602694630623,
-0.11881528794765472,
0.0953499898314476,
-0.025032415986061096,
-0.01110453624278307,
-0.05996806547045708,
-0.03109893575310707,
-0.05900143086910248,
0.03032872825860977,
-2.918699024832976e-33,
-0.013597847893834114,
0.012581721879541874,
-0.03776760399341583,
-0.006640183739364147,
-0.04953531548380852,
-0.06520488113164902,
0.11056246608495712,
0.01492222584784031,
-0.012922294437885284,
-0.059034138917922974,
-0.01061776652932167,
-0.026087023317813873,
0.022587252780795097,
0.030503954738378525,
-0.021432165056467056,
-0.1082618460059166,
0.09071258455514908,
0.07332353293895721,
0.0058774384669959545,
0.05113162472844124,
0.04077324643731117,
0.10009478777647018,
-0.0879998654127121,
-0.005887884180992842,
-0.03162671998143196,
0.09551504254341125,
-0.0684090405702591,
0.01660945639014244,
-0.05403216555714607,
-0.00256648613139987,
-0.05945894867181778,
-0.011210608296096325,
0.0017742199124768376,
-0.040152937173843384,
-0.0162129458039999,
0.023052535951137543,
0.000971889472566545,
-0.03246085345745087,
-0.07013652473688126,
0.12268824130296707,
-0.02037854492664337,
-0.02965250052511692,
-0.06841679662466049,
0.021558355540037155,
-0.079022616147995,
-0.05051765218377113,
-0.029489317908883095,
-0.031562164425849915,
-0.054262660443782806,
-0.008380668237805367,
-0.06421518325805664,
0.0013831196120008826,
-0.014767223037779331,
0.003031095489859581,
-0.05999957397580147,
-0.11356920003890991,
0.0038379349280148745,
-0.038504134863615036,
0.034016575664281845,
0.005420091561973095,
-0.054028723388910294,
0.013875335454940796,
-0.017527643591165543,
-0.04610063135623932,
-0.03761773928999901,
-0.050551339983940125,
-0.01297481544315815,
-0.03945313021540642,
0.08504848927259445,
-0.04628653824329376,
0.04801284149289131,
-0.08318468928337097,
-0.040117405354976654,
0.031821541488170624,
0.008760913275182247,
0.02401212602853775,
0.00514898169785738,
-0.004497851245105267,
0.021048013120889664,
0.01036494318395853,
0.019944116473197937,
-0.08067075163125992,
0.014229943044483662,
-0.0009118966409005225,
0.062335480004549026,
0.1212654858827591,
-0.09884882718324661,
0.019968675449490547,
0.0569850318133831,
-0.003671033540740609,
-0.024145042523741722,
-0.030362727120518684,
-0.032063551247119904,
0.05205334350466728,
-0.06600707769393921,
-4.3951459360869194e-8,
-0.10833502560853958,
0.040424641221761703,
0.008020440116524696,
0.08790759742259979,
-0.021625585854053497,
-0.02153472974896431,
-0.0512060821056366,
0.09404484182596207,
-0.05314522981643677,
0.019251668825745583,
0.013416810892522335,
0.044536422938108444,
-0.005359056405723095,
0.0034328706096857786,
0.017475204542279243,
0.04817286506295204,
0.04794624075293541,
0.07955014705657959,
0.016225382685661316,
0.02754325047135353,
0.08038578182458878,
0.0072517492808401585,
0.0053952885791659355,
-0.08736217767000198,
-0.01787455379962921,
0.00017062353435903788,
-0.08921635150909424,
0.061381805688142776,
0.02225792035460472,
-0.06564505398273468,
0.04489517956972122,
0.040217962116003036,
-0.04996303468942642,
-0.028739534318447113,
0.029805444180965424,
0.018776951357722282,
-0.02035176008939743,
-0.005071943160146475,
0.0015762913972139359,
0.013128935359418392,
0.08492818474769592,
0.04722566157579422,
-0.012113054282963276,
0.05024943873286247,
0.04453660547733307,
0.018560132011771202,
-0.00817763153463602,
-0.03159581124782562,
-0.042353514581918716,
-0.015870779752731323,
-0.04285169020295143,
-0.050849005579948425,
-0.02627897635102272,
0.11316431313753128,
0.05456919968128204,
-0.047848183661699295,
0.06738976389169693,
-0.01701987534761429,
0.03841707482933998,
0.14549775421619415,
0.03589608520269394,
0.08026743680238724,
-0.07012496143579483,
-0.0686226561665535
] | https://github.com/amith-ananthram/see-it-from-my-perspective | 2 | 0 | 0 | 0 |
Do Mice Grok? Glimpses of Hidden Progress in Sensory Cortex | https://openreview.net/forum?id=oYemKnlIrO | [
"Tanishq Kumar",
"Blake Bordelon",
"Cengiz Pehlevan",
"Venkatesh N Murthy",
"Samuel J. Gershman"
] | Poster | Does learning of task-relevant representations stop when behavior stops changing? Motivated by recent work in machine learning and the intuitive observation that human experts continue to learn after mastery, we hypothesize that task-specific representation learning in cortex can continue, even when behavior saturates. In a novel reanalysis of recently published neural data, we find evidence for such learning in posterior piriform cortex of mice following continued training on a task, long after behavior saturates at near-ceiling performance ("overtraining"). We demonstrate that class representations in cortex continue to separate during overtraining, so that examples that were incorrectly classified at the beginning of overtraining can abruptly be correctly classified later on, despite no changes in behavior during that time. We hypothesize this hidden learning takes the form of approximate margin maximization; we validate this and other predictions in the neural data, as well as build and interpret a simple synthetic model that recapitulates these phenomena. We conclude by demonstrating how this model of late-time feature learning implies an explanation for the empirical puzzle of overtraining reversal in animal learning, where task-specific representations are more robust to particular task changes because the learned features can be reused. | neuroscience; representation learning; grokking; overtraining; cortex | We find evidence for rich feature learning in mouse piriform cortex during overtraining and propose it is driven by approximate margin-maximization, a known cause of grokking in deep learning. | 12,146 | 2411.03541 | [
-0.059132467955350876,
-0.11539038270711899,
0.0975903570652008,
0.05311053991317749,
0.08880780637264252,
0.010139496996998787,
0.000900422630365938,
-0.061615634709596634,
-0.012901860289275646,
0.01593516208231449,
-0.10115791857242584,
0.015031258575618267,
-0.05215409770607948,
0.04736355319619179,
-0.06495605409145355,
-0.05237269029021263,
0.07597371935844421,
0.05932627618312836,
-0.04240253567695618,
-0.03437843918800354,
0.0404033437371254,
-0.03530005365610123,
0.04699769243597984,
0.09965068101882935,
-0.061225004494190216,
0.05990869924426079,
0.016497382894158363,
-0.04719783365726471,
-0.03219671547412872,
-0.07789958268404007,
0.002856839681044221,
-0.015314220450818539,
-0.006182419136166573,
-0.008847047574818134,
-0.07583563029766083,
0.07205093652009964,
-0.040276024490594864,
0.0562366358935833,
0.041398756206035614,
-0.009743951261043549,
0.032119233161211014,
-0.044027309864759445,
-0.012762237340211868,
0.05144626274704933,
0.03205084428191185,
-0.00568566145375371,
0.013612747192382812,
-0.13039317727088928,
-0.025044508278369904,
-0.012053770013153553,
-0.05287647619843483,
-0.000612430099863559,
-0.021653953939676285,
0.049026861786842346,
0.008190530352294445,
0.09861891716718674,
0.07562702894210815,
0.083853580057621,
-0.04608115181326866,
0.02449854463338852,
-0.051725856959819794,
-0.0412803553044796,
-0.026258857920765877,
-0.07017981261014938,
0.005077426787465811,
0.009399503469467163,
0.0014952607452869415,
0.03615918383002281,
-0.056003037840127945,
0.06166630610823631,
0.03346778079867363,
0.05511941760778427,
-0.06310488283634186,
-0.015379740856587887,
0.06636586040258408,
0.04626791924238205,
-0.015620416961610317,
0.07250097393989563,
0.06932579725980759,
-0.11502941697835922,
0.0020542670972645283,
0.03140417858958244,
0.00696131307631731,
-0.01538771204650402,
0.13634242117404938,
-0.010834609158337116,
-0.010427997447550297,
0.04301845654845238,
0.008652481250464916,
-0.02024870738387108,
-0.04074503481388092,
-0.07139035314321518,
-0.10031721740961075,
0.0005070343031547964,
0.03871879354119301,
0.03516620025038719,
-0.03561916574835777,
-0.01032544206827879,
0.03188202157616615,
0.007707805838435888,
-0.06937334686517715,
0.03874528408050537,
0.0022181225940585136,
-0.01648118533194065,
0.08770601451396942,
-0.0396992452442646,
0.009410103783011436,
0.05044545978307724,
0.06755136698484421,
-0.005385707598179579,
-0.009619674645364285,
0.0013650504406541586,
-0.03514518588781357,
0.028384394943714142,
0.05156007781624794,
0.047687023878097534,
0.015349546447396278,
0.010008834302425385,
-0.03768010437488556,
0.09504613280296326,
-0.01177788432687521,
0.01659420318901539,
-0.028074635192751884,
-0.05870262533426285,
-0.013083860278129578,
-0.057735294103622437,
-0.09081131964921951,
2.810933927182412e-33,
-0.01337418146431446,
-0.08151505142450333,
-0.01590079255402088,
-0.014812864363193512,
0.02512682043015957,
-0.09459441155195236,
-0.016503281891345978,
-0.009890662506222725,
0.032438136637210846,
0.035647545009851456,
-0.006895169615745544,
0.03142579272389412,
-0.08606041222810745,
0.10514773428440094,
-0.004932532552629709,
0.03467149659991264,
-0.07364168763160706,
0.022288072854280472,
0.0013776051346212626,
-0.07164721190929413,
-0.0064157163724303246,
0.008815256878733635,
-0.048698436468839645,
-0.0758778527379036,
-0.0013557211495935917,
-0.0032823672518134117,
-0.053508881479501724,
-0.03829745575785637,
-0.08980488777160645,
0.034230198711156845,
-0.07363130152225494,
0.009578200057148933,
-0.002257026033475995,
0.02667294442653656,
-0.031510286033153534,
-0.0586991161108017,
0.06595971435308456,
-0.09302962571382523,
0.11214819550514221,
0.005363911390304565,
0.010165348649024963,
-0.009612570516765118,
0.023275449872016907,
-0.011719102039933205,
-0.03652337193489075,
-0.07068881392478943,
0.052374836057424545,
-0.059533268213272095,
-0.07298927754163742,
-0.01249171793460846,
0.030875958502292633,
0.025100694969296455,
0.013525277376174927,
-0.08403851091861725,
-0.09225781261920929,
0.022198187187314034,
0.0009571156115271151,
0.009600944817066193,
-0.054728977382183075,
-0.026249032467603683,
0.07819150388240814,
0.05379388853907585,
-0.0018579914467409253,
0.037083473056554794,
-0.017527559772133827,
-0.01277442928403616,
-0.053716808557510376,
0.04069031402468681,
-0.01592528074979782,
0.04819487780332565,
-0.025988271459937096,
0.023784613236784935,
-0.06357462704181671,
-0.11219441890716553,
0.0883302241563797,
-0.029262719675898552,
0.06048284471035004,
-0.007938929833471775,
-0.05244967341423035,
-0.0041075944900512695,
0.01881813444197178,
-0.0529279001057148,
-0.04468347132205963,
-0.0015250637661665678,
-0.05298217386007309,
0.000656970776617527,
0.09589572995901108,
-0.07232727110385895,
-0.041892386972904205,
0.003677270608022809,
-0.02838679403066635,
-0.07348517328500748,
0.022044261917471886,
-0.013449251651763916,
-0.0341208353638649,
-3.827457728077355e-33,
-0.07585839182138443,
0.03711536526679993,
-0.018002498894929886,
-0.0041463798843324184,
-0.06263213604688644,
-0.015498856082558632,
-0.0030314335599541664,
-0.009472204372286797,
-0.1121692880988121,
-0.034120865166187286,
-0.0067223720252513885,
-0.0615340881049633,
0.009211437776684761,
0.03036246821284294,
-0.030860427767038345,
-0.035612571984529495,
0.007124052848666906,
0.06017974019050598,
0.07437673956155777,
-0.08327097445726395,
0.03901560604572296,
0.06551095843315125,
-0.016628816723823547,
0.044697970151901245,
-0.029915716499090195,
0.042134057730436325,
0.011723151430487633,
0.1278674155473709,
0.05569355562329292,
-0.026270315051078796,
-0.04320400953292847,
-0.022458689287304878,
0.011105290614068508,
0.0063228970393538475,
-0.033554431051015854,
0.0458356998860836,
-0.001701599801890552,
-0.031061694025993347,
-0.08660390973091125,
0.020646801218390465,
0.08860139548778534,
-0.06406712532043457,
-0.023379778489470482,
0.07767186313867569,
0.013319307938218117,
-0.09633355587720871,
0.003974023275077343,
0.01668102853000164,
0.003956381697207689,
0.04883377626538277,
0.0702328309416771,
0.04799746721982956,
-0.0399564765393734,
-0.021496042609214783,
-0.04727695882320404,
-0.0280538909137249,
0.020626747980713844,
-0.04910564422607422,
0.05957813933491707,
0.007355425972491503,
-0.06759331375360489,
0.021458307281136513,
0.01769174262881279,
0.015641096979379654,
0.06338745355606079,
0.025808507576584816,
-0.04690605774521828,
0.006374582648277283,
0.07931046932935715,
0.053009282797575,
0.05885154381394386,
0.034557949751615524,
-0.01062858011573553,
-0.005107648205012083,
0.04984604939818382,
0.07092831283807755,
-0.00724164629355073,
0.020728284493088722,
-0.015105700120329857,
-0.0062975469045341015,
-0.020972274243831635,
-0.07151944935321808,
-0.016524432227015495,
0.0666390210390091,
0.04564288631081581,
0.03360604867339134,
-0.0320623554289341,
0.08222650736570358,
0.03476434201002121,
-0.0547175295650959,
0.023508964106440544,
-0.03088749572634697,
-0.10016218572854996,
-0.0007469970732927322,
-0.03699544444680214,
-4.537481146371647e-8,
-0.06530626118183136,
0.033961065113544464,
0.02336854115128517,
-0.0033717367332428694,
0.1313333958387375,
0.00948029849678278,
0.0031547017861157656,
0.07827278226613998,
-0.051692765206098557,
0.0034943828359246254,
0.166671022772789,
-0.02171703241765499,
0.10885617136955261,
-0.022476984187960625,
0.051276832818984985,
0.10776017606258392,
0.03437617048621178,
0.04403376206755638,
0.014844797551631927,
0.025522615760564804,
-0.017289800569415092,
-0.014206426218152046,
0.003442013170570135,
-0.08321680873632431,
0.018530676141381264,
-0.08529860526323318,
-0.0911029502749443,
0.16663505136966705,
0.0016883441712707281,
0.006701660808175802,
0.03884406015276909,
0.035553187131881714,
0.004330073483288288,
0.029118478298187256,
0.04047528654336929,
0.05800285190343857,
0.07836734503507614,
-0.015754394233226776,
-0.02283496782183647,
0.04257257282733917,
-0.053047556430101395,
0.06305993348360062,
-0.0041892980225384235,
0.042419370263814926,
0.03837350383400917,
-0.028358465060591698,
0.019730038940906525,
-0.048878394067287445,
0.031606949865818024,
-0.020070381462574005,
-0.04152344539761543,
0.07219213247299194,
0.04109429568052292,
0.04230140522122383,
0.014174815267324448,
-0.011451958678662777,
-0.07190323621034622,
-0.043430909514427185,
-0.025427239015698433,
0.023608457297086716,
0.021504590287804604,
0.06006062403321266,
-0.06913085281848907,
0.08524321764707565
] | 0 | 0 | 0 | 0 |
|
SymmCD: Symmetry-Preserving Crystal Generation with Diffusion Models | https://openreview.net/forum?id=xnssGv9rpW | [
"Daniel Levy",
"Siba Smarak Panigrahi",
"Sékou-Oumar Kaba",
"Qiang Zhu",
"Kin Long Kelvin Lee",
"Mikhail Galkin",
"Santiago Miret",
"Siamak Ravanbakhsh"
] | Poster | Generating novel crystalline materials has potential to lead to advancements in fields such as electronics, energy storage, and catalysis. The defining characteristic of crystals is their symmetry, which plays a central role in determining their physical properties. However, existing crystal generation methods either fail to generate materials that display the symmetries of real-world crystals, or simply replicate the symmetry information from examples in a database. To address this limitation, we propose SymmCD, a novel diffusion-based generative model that explicitly incorporates crystallographic symmetry into the generative process. We decompose crystals into two components and learn their joint distribution through diffusion: 1) the asymmetric unit, the smallest subset of the crystal which can generate the whole crystal through symmetry transformations, and; 2) the symmetry transformations needed to be applied to each atom in the asymmetric unit. We also use a novel and interpretable representation for these transformations, enabling generalization across different crystallographic symmetry groups. We showcase the competitive performance of SymmCD on a subset of the Materials Project, obtaining diverse and valid crystals with realistic symmetries and predicted properties. | Crystals, Symmetry, Materials, Diffusion, Generative Models, Equivariance | We generate symmetric crystal structures by generating an asymmetric unit along with symmetry transformations | 12,119 | 2502.03638 | [
0.00014130774070508778,
-0.07416762411594391,
0.012263518758118153,
0.024196481332182884,
0.026007981970906258,
-0.05239412561058998,
-0.03578808158636093,
-0.06724043935537338,
-0.034530460834503174,
-0.015250484459102154,
0.0353248193860054,
-0.027979956939816475,
-0.015127072110772133,
0.030618716031312943,
0.031242096796631813,
-0.04192683845758438,
-0.0015904841711744666,
0.03238910064101219,
-0.02037137560546398,
0.014194768853485584,
0.04071974754333496,
-0.06481647491455078,
0.009120437316596508,
-0.04418906196951866,
0.07703810185194016,
0.08129820227622986,
0.087388776242733,
-0.04729928821325302,
-0.0061971270479261875,
-0.1236460879445076,
0.08998046070337296,
0.0063619548454880714,
-0.08239085227251053,
-0.040648870170116425,
0.02938874624669552,
0.045082248747348785,
-0.016031181439757347,
-0.0007047280087135732,
0.05351671203970909,
-0.0977301374077797,
-0.008399931713938713,
0.017233235761523247,
0.014590926468372345,
0.07238548249006271,
-0.012987034395337105,
-0.00033837099908851087,
0.024734681472182274,
-0.028046179562807083,
-0.016492580994963646,
-0.039622168987989426,
0.031625743955373764,
0.07694482803344727,
-0.056024178862571716,
0.06811759620904922,
-0.005217375233769417,
0.06884291023015976,
0.04173912853002548,
-0.03057970106601715,
0.012492797337472439,
-0.05995924398303032,
-0.057943135499954224,
-0.05987594276666641,
-0.007961473427712917,
-0.07806909829378128,
-0.024831872433423996,
-0.0010950719006359577,
0.04328116402029991,
0.01905728131532669,
0.02328362502157688,
0.016334252431988716,
-0.03583529591560364,
0.012265880592167377,
-0.061775315552949905,
0.04101702570915222,
0.023883264511823654,
0.050592128187417984,
0.09133123606443405,
-0.030804747715592384,
-0.01190692838281393,
-0.016221510246396065,
-0.0738147646188736,
0.030857091769576073,
0.0636587142944336,
0.002319979015737772,
0.058020979166030884,
0.0009758883970789611,
0.0014789416454732418,
0.025265024974942207,
0.03636602312326431,
-0.03370283544063568,
-0.030258124694228172,
0.11231136322021484,
-0.016840631142258644,
-0.05655871331691742,
0.003495913464576006,
0.079877570271492,
0.07690224796533585,
0.03749145194888115,
0.027559159323573112,
0.0639072060585022,
0.06909850239753723,
0.006303691770881414,
0.04060729220509529,
-0.031923938542604446,
0.057676952332258224,
-0.05901361629366875,
0.016013547778129578,
0.03191234543919563,
-0.030037958174943924,
0.002669036854058504,
0.02853996679186821,
0.03927937522530556,
0.02448786422610283,
0.040914103388786316,
-0.003748582676053047,
0.02614806592464447,
-0.05915534496307373,
-0.009421419352293015,
0.02822733484208584,
-0.010764030739665031,
0.010342346504330635,
0.051319669932127,
-0.09690073132514954,
0.07577619701623917,
0.005594763904809952,
-0.011476131156086922,
-0.08004029095172882,
-1.1776340260472725e-33,
-0.002014022320508957,
0.059604134410619736,
0.0450916513800621,
0.09548799693584442,
0.060252346098423004,
0.01140123512595892,
-0.04924667254090309,
-0.04511890560388565,
-0.053990282118320465,
0.020183110609650612,
-0.1213882714509964,
-0.03312431275844574,
-0.03420913219451904,
-0.0004673675575759262,
-0.031430814415216446,
-0.019372697919607162,
-0.03671808913350105,
0.05317310988903046,
-0.002013219054788351,
-0.018877312541007996,
0.008089962415397167,
0.027573078870773315,
-0.0495305210351944,
0.061670128256082535,
-0.04411205276846886,
0.05260123312473297,
0.07478698343038559,
-0.032008711248636246,
0.0048373728059232235,
-0.007289520464837551,
-0.02216091938316822,
0.040016889572143555,
-0.04518576338887215,
-0.01210884191095829,
0.08509377390146255,
-0.04156900942325592,
-0.05751073360443115,
-0.019334428012371063,
0.056680601090192795,
-0.07953739166259766,
-0.002240957459434867,
0.06115172430872917,
-0.055630795657634735,
0.07560504227876663,
-0.06591059267520905,
0.05009016767144203,
0.07931018620729446,
0.021227627992630005,
-0.02180020697414875,
-0.07293565571308136,
-0.02235696092247963,
-0.011732549406588078,
0.01612791419029236,
0.016021400690078735,
-0.020247522741556168,
0.016374357044696808,
0.028830403462052345,
-0.06726798415184021,
0.06036006659269333,
0.005661708302795887,
0.09769472479820251,
0.03956996649503708,
-0.046183567494153976,
-0.011459521017968655,
-0.03779388591647148,
0.03660133108496666,
0.01491056103259325,
-0.004642651882022619,
0.042595118284225464,
-0.002847563475370407,
-0.004570388235151768,
-0.0041447291150689125,
-0.02273515611886978,
0.0002031574258580804,
0.10263395309448242,
-0.02782200090587139,
0.07138015329837799,
-0.07986173778772354,
0.009725067764520645,
-0.05250754952430725,
-0.097672238945961,
0.06195569783449173,
-0.11287886649370193,
-0.10859668254852295,
-0.05715744197368622,
-0.04600225016474724,
-0.06793427467346191,
-0.11509913206100464,
-0.05276251584291458,
-0.095355324447155,
-0.004552221857011318,
-0.1392425298690796,
-0.04992867633700371,
0.04121909290552139,
-0.07892033457756042,
-1.2092951398672054e-33,
-0.0440458320081234,
-0.035182226449251175,
0.04277375712990761,
0.012331463396549225,
-0.029089365154504776,
-0.038446880877017975,
-0.014331587590277195,
0.046046193689107895,
0.03063897229731083,
-0.03631081432104111,
0.0370023176074028,
-0.00788163486868143,
-0.08512997627258301,
-0.07939787954092026,
-0.0014077817322686315,
0.028438696637749672,
0.1107330247759819,
-0.06770338863134384,
0.021278033033013344,
0.023678062483668327,
-0.04718681052327156,
0.13461539149284363,
-0.08396299183368683,
0.023180941119790077,
0.0012517701834440231,
0.05770766735076904,
0.030715713277459145,
0.0441034659743309,
0.0855838879942894,
0.052876487374305725,
0.008540824055671692,
0.025495091453194618,
-0.001395783619955182,
-0.016388140618801117,
-0.13882368803024292,
0.037694357335567474,
-0.011797500774264336,
0.009099907241761684,
-0.02837553061544895,
0.0955946072936058,
-0.01968744955956936,
-0.05357528105378151,
-0.06825421750545502,
0.08675607293844223,
0.01375583279877901,
-0.04261239245533943,
0.011148467659950256,
0.02358907274901867,
-0.016477826982736588,
0.016972355544567108,
0.057248640805482864,
-0.03144456818699837,
0.006669142283499241,
-0.044543586671352386,
-0.022199571132659912,
0.03823871165513992,
-0.10672077536582947,
-0.02327481284737587,
0.05623074248433113,
-0.005178628955036402,
0.024083392694592476,
0.0034756213426589966,
0.10910841077566147,
-0.032948024570941925,
-0.12947151064872742,
-0.01256672665476799,
-0.011726303026080132,
-0.05078212544322014,
0.022725604474544525,
0.07792874425649643,
-0.046057332307100296,
0.03382740914821625,
-0.011305411346256733,
0.02912820316851139,
0.08991701155900955,
-0.06766742467880249,
0.017610639333724976,
0.0004457634931895882,
-0.03694648668169975,
-0.016470525413751602,
-0.03819190710783005,
0.07982474565505981,
0.054978206753730774,
-0.008717766962945461,
0.010571043007075787,
-0.11948218196630478,
0.09237054735422134,
0.08582903444766998,
-0.03286903724074364,
0.05198580026626587,
-0.011864901520311832,
-0.041601452976465225,
0.03123101033270359,
0.1381758153438568,
-0.023593489080667496,
-4.080742499468215e-8,
0.009948467835783958,
0.03050895966589451,
0.045157622545957565,
-0.0313541479408741,
-0.03042459487915039,
0.04197138547897339,
0.028617829084396362,
0.0010860678739845753,
0.018751684576272964,
-0.036995064467191696,
0.014851926825940609,
-0.008292406797409058,
-0.06733900308609009,
-0.016882814466953278,
-0.024072762578725815,
0.12880846858024597,
0.016903690993785858,
0.055572595447301865,
-0.04821986332535744,
-0.010249488055706024,
0.04539015144109726,
0.04430564120411873,
0.011202444322407246,
-0.008110417053103447,
0.08290193229913712,
0.014970957301557064,
-0.11520986258983612,
-0.04664952680468559,
-0.01318846084177494,
0.02198401838541031,
0.007275086361914873,
-0.02566000632941723,
0.07644932717084885,
-0.011109585873782635,
-0.004207699093967676,
0.07813074439764023,
0.0035140032414346933,
-0.07278067618608475,
0.021118860691785812,
-0.12733450531959534,
0.010340140201151371,
0.01444875169545412,
0.007178418803960085,
0.06422511488199234,
-0.02289297990500927,
0.006675323937088251,
-0.009508516639471054,
0.014018472284078598,
0.04790453240275383,
0.09554906189441681,
0.039669740945100784,
-0.013923782855272293,
-0.0797971561551094,
-0.04162957891821861,
-0.05660786107182503,
0.021242650225758553,
-0.08259937167167664,
-0.014934703707695007,
0.0588083490729332,
-0.0579608753323555,
0.02449328638613224,
-0.045876819640398026,
-0.04452025145292282,
-0.014176979660987854
] | https://github.com/sibasmarak/SymmCD | 14 | 0 | 0 | 0 |
Benchmarking LLMs' Judgments with No Gold Standard | https://openreview.net/forum?id=uE84MGbKD7 | [
"Shengwei Xu",
"Yuxuan Lu",
"Grant Schoenebeck",
"Yuqing Kong"
] | Poster | We introduce the GEM (Generative Estimator for Mutual Information), an evaluation metric for assessing language generation by large language models (LLMs), particularly in generating informative judgments, without the need for a gold standard reference. GEM broadens the scenarios where we can benchmark LLM generation performance-from traditional ones, like machine translation and summarization, where gold standard references are readily available, to subjective tasks without clear gold standards, such as academic peer review.
GEM uses a generative model to estimate mutual information between candidate and reference responses, without requiring the reference to be a gold standard. In experiments on two human-annotated datasets, GEM demonstrates competitive correlations with human scores compared to the state-of-the-art GPT-4o Examiner, and outperforms all other baselines. Additionally, GEM is more robust against strategic manipulation, such as rephrasing or elongation, which can artificially inflate scores under a GPT-4o Examiner.
We also present GRE-bench (Generating Review Evaluation Benchmark) which evaluates LLMs based on how well they can generate high-quality peer reviews for academic research papers. Because GRE-bench is based upon GEM, it inherits its robustness properties. Additionally, GRE-bench circumvents data contamination problems (or data leakage) by using the continuous influx of new open-access research papers and peer reviews each year. We show GRE-bench results of various popular LLMs on their peer review capabilities using the ICLR2023 dataset. | Benchmarking, Peer Review, Mutual Information, Data Contamination, Large Language Models | null | 12,113 | 2411.07127 | [
-0.04274845868349075,
-0.023665478453040123,
-0.054705068469047546,
0.05350641533732414,
0.024345537647604942,
0.011166442185640335,
0.012259050272405148,
0.039267465472221375,
0.04405713453888893,
0.01107296533882618,
-0.09769705682992935,
0.005953345447778702,
0.06094173714518547,
-0.0009877386037260294,
-0.04426278918981552,
0.023268595337867737,
0.130897656083107,
0.05047763139009476,
-0.0037495307624340057,
-0.06269708275794983,
0.06540019065141678,
0.05485420674085617,
0.08830393850803375,
-0.009159388020634651,
0.036430127918720245,
-0.05672033503651619,
-0.061200227588415146,
0.009955462999641895,
-0.021071137860417366,
-0.04616944119334221,
-0.027944184839725494,
0.051553864032030106,
-0.006372828036546707,
0.07076346129179001,
-0.1365315467119217,
0.06259998679161072,
-0.018031582236289978,
0.06473683565855026,
0.014043725095689297,
-0.04545065015554428,
-0.043040309101343155,
0.013342918828129768,
-0.008923876099288464,
-0.010278526693582535,
0.08931854367256165,
-0.027573224157094955,
-0.021345553919672966,
-0.02265308052301407,
-0.07658122479915619,
0.09482124447822571,
-0.11291220784187317,
0.019013475626707077,
0.03247223421931267,
-0.05073032155632973,
-0.0336630716919899,
0.015469836071133614,
0.04187687113881111,
-0.023910285905003548,
0.002887431066483259,
-0.07537879049777985,
-0.08482985943555832,
-0.14195191860198975,
-0.047250811010599136,
-0.019364681094884872,
0.02023439295589924,
-0.06122864410281181,
-0.008848088793456554,
0.07079696655273438,
-0.07962174713611603,
-0.004308247473090887,
0.011039409786462784,
0.029910102486610413,
0.0011829206487163901,
0.052193690091371536,
-0.025572175160050392,
0.07509857416152954,
0.03239291161298752,
-0.009158465079963207,
0.046865902841091156,
-0.030072007328271866,
0.010795233771204948,
-0.007288314867764711,
0.05723322555422783,
0.03477494418621063,
0.053715161979198456,
-0.0031102995853871107,
0.08515578508377075,
0.06046099215745926,
-0.015334283001720905,
0.03271101415157318,
-0.009676878340542316,
-0.044493887573480606,
0.06638950109481812,
0.011343671940267086,
0.05238120257854462,
-0.0024855025112628937,
-0.0025728370528668165,
-0.09777618199586868,
0.020150795578956604,
0.07920848578214645,
0.06378497183322906,
0.15134502947330475,
0.01946602389216423,
-0.09932457655668259,
0.003141955239698291,
-0.007277568336576223,
0.023160777986049652,
0.04855698347091675,
0.01169165875762701,
-0.058529727160930634,
0.03007446601986885,
0.04916639253497124,
0.030280701816082,
0.01880296692252159,
0.09103083610534668,
-0.02107727713882923,
0.015086169354617596,
0.01736452616751194,
0.06314055621623993,
0.04035737365484238,
-0.027397286146879196,
0.06244836375117302,
0.028014827519655228,
-0.026434503495693207,
-0.008252177387475967,
0.019576625898480415,
-0.015534976497292519,
2.5405544307879466e-33,
0.06990443170070648,
0.06716600805521011,
-0.05008833855390549,
0.05502510815858841,
-0.04706971347332001,
0.004839589819312096,
-0.04805388301610947,
0.030470293015241623,
-0.06959321349859238,
-0.07234927266836166,
-0.01668640784919262,
0.0453171581029892,
-0.030248265713453293,
0.03336154296994209,
0.007008523680269718,
0.07579196244478226,
-0.039669219404459,
0.048539869487285614,
-0.02678743191063404,
0.005099080037325621,
0.00610702857375145,
-0.05154801905155182,
0.01703162118792534,
-0.04596894234418869,
0.029722871258854866,
0.02569897286593914,
0.039335910230875015,
-0.09785453975200653,
-0.02830313891172409,
0.000011938504940189887,
-0.06690848618745804,
-0.02589593641459942,
0.07378240674734116,
-0.014601480215787888,
0.07180428504943848,
0.006101314444094896,
-0.06409107893705368,
0.005329127889126539,
0.012490427121520042,
-0.033115580677986145,
-0.007332489360123873,
-0.005305252969264984,
0.0031296969391405582,
0.01232138555496931,
-0.05046309158205986,
-0.02617769129574299,
-0.010057521983981133,
0.05504610762000084,
0.021191490814089775,
-0.03204904496669769,
0.04671692103147507,
0.050242308527231216,
0.01585991121828556,
0.09852664917707443,
-0.014142618514597416,
0.04680655896663666,
0.005095485132187605,
-0.009919440373778343,
-0.0022595112677663565,
-0.013082410208880901,
0.03485472500324249,
0.1030169427394867,
-0.026809189468622208,
0.0362561009824276,
0.009719387628138065,
0.04107600450515747,
-0.10113845765590668,
0.07710309326648712,
0.08552680164575577,
0.009454055689275265,
-0.004574846010655165,
-0.018976911902427673,
-0.02793145924806595,
0.008465442806482315,
0.014781058765947819,
-0.052837222814559937,
0.03159470856189728,
-0.05400240048766136,
0.020879577845335007,
0.023745501413941383,
0.00006419749115593731,
0.030725425109267235,
-0.008562258444726467,
-0.11558382213115692,
-0.04314139857888222,
-0.0668976902961731,
0.048197273164987564,
-0.04460577666759491,
0.010941374115645885,
-0.009543966501951218,
-0.02786114439368248,
-0.03939395770430565,
-0.05902303755283356,
0.03299783915281296,
0.011360273696482182,
-2.164016957925682e-33,
-0.07592757791280746,
-0.05257463455200195,
0.013369395397603512,
0.13827471435070038,
-0.016322243958711624,
-0.0769265815615654,
-0.03129325434565544,
0.01565113104879856,
0.0007695633685216308,
-0.040929608047008514,
-0.032511670142412186,
-0.11213873326778412,
0.06829509139060974,
-0.018025832250714302,
0.0793086513876915,
-0.050293970853090286,
-0.027612607926130295,
-0.005974629428237677,
-0.007227544207125902,
0.08268573880195618,
0.038543857634067535,
0.10472655296325684,
-0.0383635014295578,
0.05868736654520035,
-0.018983453512191772,
0.027627458795905113,
0.0026197494007647038,
-0.06664123386144638,
-0.029616134241223335,
-0.03493769094347954,
0.06558084487915039,
0.035866908729076385,
-0.04362207278609276,
-0.008463808335363865,
-0.09872899949550629,
-0.006084529682993889,
0.04948737844824791,
-0.030812764540314674,
-0.023652466014027596,
0.13016745448112488,
0.0433778390288353,
0.020203279331326485,
-0.05545756593346596,
-0.015431050211191177,
0.012337400577962399,
0.06557925790548325,
-0.07484409213066101,
-0.02129930630326271,
0.01964772306382656,
0.02735145390033722,
-0.04788495600223541,
-0.008106225170195103,
-0.07789081335067749,
-0.0002309269766556099,
-0.03584889695048332,
-0.20503517985343933,
-0.05196276307106018,
-0.03485796973109245,
0.013226523995399475,
0.025881797075271606,
-0.021319154649972916,
0.03386539965867996,
0.019898032769560814,
-0.03169739991426468,
-0.02287241443991661,
-0.06431417167186737,
0.03809213638305664,
-0.0020639176946133375,
-0.010800166055560112,
0.04541180282831192,
0.06626591831445694,
0.005388672463595867,
0.008107403293251991,
-0.007808650378137827,
0.08306895941495895,
0.05302804708480835,
-0.0028772142250090837,
-0.026186898350715637,
-0.07167091220617294,
-0.13095566630363464,
0.02888033725321293,
0.010774055495858192,
0.059965912252664566,
0.007171507924795151,
0.0245262049138546,
0.025148309767246246,
0.06268563121557236,
0.08973254263401031,
-0.07124486565589905,
0.015513872727751732,
-0.027034791186451912,
-0.06414978951215744,
-0.017785334959626198,
-0.06182364746928215,
-0.0971202626824379,
-4.01733259991488e-8,
-0.01807396113872528,
0.031161921098828316,
-0.05252968519926071,
0.11079944670200348,
-0.06769814342260361,
-0.02271309494972229,
-0.05433002486824989,
0.07558737695217133,
-0.039418578147888184,
-0.009773577563464642,
0.014519811607897282,
-0.005872579291462898,
-0.1269702911376953,
-0.053495343774557114,
0.03227473050355911,
0.04316063970327377,
0.019222941249608994,
0.042473096400499344,
-0.022066764533519745,
-0.05795392394065857,
0.05492504686117172,
0.05909045413136482,
-0.011531392112374306,
0.023482337594032288,
0.011005744338035583,
-0.008711754344403744,
-0.0205878596752882,
-0.02336006797850132,
0.03627840057015419,
-0.031793590635061264,
0.05390533432364464,
0.02999930828809738,
-0.03304765000939369,
-0.04201480373740196,
0.06803997606039047,
0.039148103445768356,
0.021713601425290108,
0.028423132374882698,
0.09364371001720428,
0.05705031752586365,
0.0037349951453506947,
0.05309678614139557,
-0.0806654617190361,
0.02846681699156761,
-0.013101854361593723,
-0.04044501110911369,
-0.09496354311704636,
-0.055314116179943085,
-0.0038458595518022776,
-0.04813297092914581,
0.03266531974077225,
-0.12693703174591064,
-0.022074420005083084,
-0.02757812663912773,
0.045135561376810074,
0.09014499932527542,
0.006217911373823881,
-0.05249957740306854,
0.02350301668047905,
0.008857740089297295,
0.08935429155826569,
-0.03474745526909828,
-0.0188478734344244,
-0.024632392451167107
] | https://github.com/yx-lu/benchmarking-llms--judgments-with-no-gold-standard | 0 | 0 | 0 | 0 |
The Pitfalls of Memorization: When Memorization Hurts Generalization | https://openreview.net/forum?id=vVhZh9ZpIM | [
"Reza Bayat",
"Mohammad Pezeshki",
"Elvis Dohmatob",
"David Lopez-Paz",
"Pascal Vincent"
] | Poster | Neural networks often learn simple explanations that fit the majority of the data while memorizing exceptions that deviate from these explanations. This behavior leads to poor generalization when the learned explanations rely on spurious correlations. In this work, we formalize $\textit{the interplay between memorization and generalization}$, showing that spurious correlations would particularly lead to poor generalization when are combined with memorization. Memorization can reduce training loss to zero, leaving no incentive to learn robust, generalizable patterns. To address this, we propose $\textit{memorization-aware training}$ (MAT), which uses held-out predictions as a signal of memorization to shift a model's logits. MAT encourages learning robust patterns invariant across distributions, improving generalization under distribution shifts. | Memorization, Generalization, Spurious Correlations | null | 12,110 | 2412.07684 | [
-0.06691565364599228,
-0.06034291908144951,
0.048431556671857834,
0.0913611650466919,
0.07472381740808487,
0.040446605533361435,
0.05182977393269539,
-0.03287016972899437,
-0.008716017939150333,
-0.02602720819413662,
-0.01994439773261547,
-0.009349055588245392,
0.028339611366391182,
0.030481398105621338,
-0.022414645180106163,
-0.07588735967874527,
0.021930551156401634,
0.0703616514801979,
-0.040388498455286026,
-0.009675837121903896,
0.030264826491475105,
-0.0022384373005479574,
0.033159039914608,
0.058123502880334854,
0.019891761243343353,
0.005350420717149973,
0.003147799288854003,
-0.032373860478401184,
0.029187070205807686,
-0.002945818705484271,
0.027514031156897545,
0.04364430159330368,
-0.02074245736002922,
0.010684161446988583,
-0.07822943478822708,
0.0004664584994316101,
-0.11388392746448517,
0.10093854367733002,
0.02527112141251564,
0.012915103696286678,
0.03122555837035179,
0.04598899930715561,
0.038465775549411774,
0.019467076286673546,
0.024847479537129402,
-0.011822853237390518,
0.04913132265210152,
-0.07279667258262634,
-0.048038456588983536,
0.015759825706481934,
-0.030221762135624886,
0.037000369280576706,
-0.0626658946275711,
0.01568509452044964,
0.007299761287868023,
-0.004301032517105341,
0.0929885134100914,
0.05556587129831314,
-0.10117747634649277,
0.0018141228938475251,
0.0029586723539978266,
-0.15311050415039062,
-0.06331725418567657,
-0.04330887645483017,
-0.023003166541457176,
-0.04078058898448944,
0.0366910919547081,
0.09878679364919662,
0.0038242903538048267,
0.10114700347185135,
-0.0449540875852108,
0.11861877143383026,
-0.08902734518051147,
0.04355063661932945,
0.051929622888565063,
0.10540473461151123,
-0.02568637579679489,
-0.04085180163383484,
0.05179300159215927,
-0.02357577346265316,
-0.06300380080938339,
0.035652704536914825,
0.006528145633637905,
0.028754666447639465,
0.1367507427930832,
-0.003625442273914814,
0.07179301232099533,
0.014285516925156116,
0.0626397579908371,
-0.027256928384304047,
-0.006528964266180992,
-0.03983169421553612,
0.010633086785674095,
-0.015958551317453384,
-0.0036709667183458805,
0.06948141753673553,
0.016371503472328186,
-0.07226644456386566,
-0.08275264501571655,
0.056567080318927765,
-0.0032037494238466024,
0.09912380576133728,
-0.040239930152893066,
0.00795366708189249,
0.059446271508932114,
-0.06712712347507477,
0.061078719794750214,
-0.011517485603690147,
0.08425993472337723,
-0.12860384583473206,
0.03142934665083885,
0.07032977044582367,
0.004299257881939411,
0.034114956855773926,
-0.0004611686454154551,
0.02300693839788437,
0.03357676789164543,
-0.036271680146455765,
-0.03707544133067131,
0.07498735934495926,
-0.014086591079831123,
0.008525637909770012,
-0.024566305801272392,
0.0670168399810791,
-0.03882889449596405,
-0.006998306605964899,
-0.05341918766498566,
2.866799296205241e-33,
0.06676001846790314,
0.029463915154337883,
0.025237105786800385,
0.01679920218884945,
0.05418248102068901,
-0.05697067454457283,
-0.08456216007471085,
-0.025568168610334396,
0.026775894686579704,
0.055681344121694565,
0.03662523627281189,
0.050755638629198074,
-0.0600411482155323,
0.11191506683826447,
-0.00347710563801229,
0.0663602203130722,
-0.062477365136146545,
0.05947034806013107,
0.011115861125290394,
-0.050912197679281235,
0.0396016463637352,
-0.02061995305120945,
-0.01693408191204071,
-0.0397367924451828,
0.028331762179732323,
-0.013300502672791481,
0.039569687098264694,
-0.011733894236385822,
0.022601861506700516,
0.01019344013184309,
-0.03733516111969948,
-0.004429675173014402,
0.05306186527013779,
0.012392411008477211,
0.025813385844230652,
-0.0411711260676384,
-0.016221430152654648,
-0.06866234540939331,
0.044398173689842224,
-0.024101151153445244,
-0.0630958303809166,
-0.007655635010451078,
-0.0022986584808677435,
-0.0008804242825135589,
0.002165946178138256,
-0.11100656539201736,
0.006771826185286045,
-0.01881401240825653,
-0.04697953909635544,
-0.10275394469499588,
0.0020003223326057196,
0.023997262120246887,
-0.05763646960258484,
-0.04824063554406166,
-0.009499963372945786,
0.014747925102710724,
0.039820995181798935,
0.021308964118361473,
-0.02290499210357666,
0.036306023597717285,
-0.005263127386569977,
0.05944521725177765,
0.01942017488181591,
0.056147653609514236,
-0.07110852748155594,
0.03888864070177078,
-0.016248956322669983,
0.06272505968809128,
0.007339160423725843,
0.012991526164114475,
-0.029348723590373993,
-0.0024190701078623533,
-0.11539141088724136,
-0.07462243735790253,
0.07774923741817474,
0.003402703208848834,
0.07020972669124603,
-0.1087108924984932,
-0.02532072551548481,
-0.013417505659162998,
0.028546493500471115,
-0.06233559176325798,
-0.007281217258423567,
-0.009654992260038853,
-0.07061102986335754,
-0.010193782858550549,
0.0824858620762825,
-0.06577546894550323,
-0.04760425537824631,
-0.028410600498318672,
-0.08904583752155304,
-0.010573276318609715,
0.01153494417667389,
-0.02951623499393463,
-0.05512306094169617,
-3.953978021766281e-33,
-0.06802161782979965,
0.053121019154787064,
-0.020854489877820015,
-0.006546983495354652,
-0.02417999505996704,
-0.053258176892995834,
-0.029942672699689865,
0.013257999904453754,
-0.050417426973581314,
-0.08837733417749405,
-0.04112410917878151,
-0.05147087574005127,
-0.04162828251719475,
0.0641200989484787,
-0.011553512886166573,
-0.06423657387495041,
-0.011010597459971905,
0.07707063108682632,
-0.00885037798434496,
-0.003899905364960432,
0.027493774890899658,
0.028639115393161774,
-0.07299759984016418,
0.029678335413336754,
-0.038642510771751404,
0.0038534486666321754,
-0.09821426868438721,
0.08278220146894455,
-0.023876860737800598,
-0.02839457057416439,
-0.09970080107450485,
-0.030190473422408104,
-0.022554123774170876,
-0.02809579111635685,
-0.02534039504826069,
0.051044855266809464,
-0.020939510315656662,
0.0434907041490078,
-0.0936734527349472,
0.05852820351719856,
0.02194763906300068,
-0.016103409230709076,
-0.045714493840932846,
0.028952673077583313,
0.0626746118068695,
-0.10061805695295334,
-0.03732084110379219,
0.02298862673342228,
0.03508930280804634,
0.03200653940439224,
-0.06471674889326096,
0.021833006292581558,
-0.079544298350811,
0.02271706983447075,
-0.05755103752017021,
0.013178055174648762,
-0.003332365071401,
0.005541400983929634,
0.014490229077637196,
0.02895457111299038,
-0.08252021670341492,
0.01116137020289898,
0.018510842695832253,
-0.054700516164302826,
-0.03698984906077385,
-0.05485401675105095,
-0.01969393528997898,
0.03467778488993645,
0.04762273654341698,
0.03986239433288574,
0.11769615113735199,
0.07452370226383209,
-0.05025960132479668,
-0.015612166374921799,
-0.027287675067782402,
0.018297679722309113,
0.032521750777959824,
-0.04995083808898926,
-0.044273462146520615,
-0.02437634766101837,
0.049438852816820145,
0.03617372736334801,
0.04760356619954109,
0.020782534033060074,
0.02538496069610119,
0.07019124180078506,
0.03414790332317352,
0.09204505383968353,
-0.006141561549156904,
-0.04087959975004196,
-0.06980854272842407,
0.024378206580877304,
-0.0289032943546772,
0.04586396738886833,
-0.10321084409952164,
-4.392009955722642e-8,
-0.09866099059581757,
0.07097993791103363,
0.03395291790366173,
0.043012216687202454,
0.09574129432439804,
-0.06278354674577713,
-0.044696662575006485,
0.07234200090169907,
-0.005425675306469202,
-0.05281179025769234,
0.06412606686353683,
0.015449625439941883,
-0.08002083003520966,
-0.0008037154329940677,
-0.019137689843773842,
0.12096250802278519,
-0.004486170131713152,
0.021062884479761124,
-0.038199663162231445,
-0.028700027614831924,
-0.019993286579847336,
-0.032806653529405594,
0.022051457315683365,
0.004241904243826866,
0.019506938755512238,
-0.0918235182762146,
0.028578203171491623,
0.12231134623289108,
0.04610098525881767,
0.02600124292075634,
0.022545713931322098,
-0.009201260283589363,
0.04684682935476303,
-0.028822725638747215,
0.04370689392089844,
0.12687599658966064,
0.03699330613017082,
-0.0005663665942847729,
0.10968483984470367,
-0.07768301665782928,
-0.016706297174096107,
0.021147020161151886,
0.06812436878681183,
0.04137251898646355,
-0.002144303871318698,
0.010480071417987347,
-0.004408174194395542,
-0.1157320886850357,
0.018106402829289436,
-0.041919782757759094,
0.028585931286215782,
0.03233286365866661,
0.039350468665361404,
0.06578563898801804,
0.06539466977119446,
-0.004294905811548233,
-0.03148084133863449,
-0.0524844229221344,
-0.03706750646233559,
0.06894388049840927,
-0.049849338829517365,
0.04346967115998268,
-0.040024518966674805,
0.021936725825071335
] | https://github.com/facebookresearch/pitfalls-of-memorization | 40 | 0 | 0 | 0 |
Graph Neural Networks Gone Hogwild | https://openreview.net/forum?id=WfxPVtYRlL | [
"Olga Solodova",
"Nick Richardson",
"Deniz Oktay",
"Ryan P Adams"
] | Poster | Graph neural networks (GNNs) appear to be powerful tools to learn state representations for agents in distributed, decentralized multi-agent systems, but generate catastrophically incorrect predictions when nodes update asynchronously during inference.
This failure under asynchrony effectively excludes these architectures from many potential applications where synchrony is difficult or impossible to enforce, e.g., robotic swarms or sensor networks.
In this work we identify ''implicitly-defined'' GNNs as a class of architectures which is provably robust to asynchronous ''hogwild'' inference, adapting convergence guarantees from work in asynchronous and distributed optimization.
We then propose a novel implicitly-defined GNN architecture, which we call an energy GNN.
We show that this architecture outperforms other GNNs from this class on a variety of synthetic tasks inspired by multi-agent systems. | graph neural networks, multi-agent, asynchronous, decentralized | null | 12,109 | 2407.00494 | [
-0.11872434616088867,
-0.05283094197511673,
-0.03520189970731735,
0.0745876133441925,
0.0009125421056523919,
-0.025655709207057953,
-0.03769192844629288,
-0.06598170846700668,
-0.04766884446144104,
-0.046691447496414185,
-0.027737801894545555,
0.014651939272880554,
0.022655632346868515,
0.0202080886811018,
-0.03223671391606331,
0.0015722736716270447,
0.1331651359796524,
0.04884903132915497,
-0.07739169150590897,
-0.06298577785491943,
-0.03145280107855797,
0.041903261095285416,
0.025658823549747467,
0.044998716562986374,
0.044573087245225906,
0.008656203746795654,
-0.059116180986166,
-0.07213637977838516,
-0.015815144404768944,
-0.0002830344601534307,
0.07506263256072998,
-0.0472673699259758,
-0.07415354251861572,
0.016841523349285126,
0.0015528517542406917,
0.09613238275051117,
-0.07411161810159683,
0.009254090487957,
0.08046433329582214,
-0.014695401303470135,
0.03560009226202965,
0.02926124632358551,
0.0071568978019058704,
-0.010621869005262852,
0.036606449633836746,
0.02517305314540863,
0.022909492254257202,
-0.0035968476440757513,
-0.005762397311627865,
-0.017316972836852074,
-0.03390142694115639,
-0.06804432719945908,
-0.011656399816274643,
0.03164006024599075,
0.035907503217458725,
-0.017707478255033493,
0.004143084399402142,
0.027890430763363838,
-0.026131395250558853,
-0.031056899577379227,
0.05840548500418663,
-0.11201769858598709,
0.0017821850487962365,
0.0094329584389925,
-0.024561738595366478,
0.05130443722009659,
0.02459566853940487,
0.029916491359472275,
0.009075561538338661,
0.0032424379605799913,
0.06392140686511993,
0.029832102358341217,
0.007061844225972891,
-0.060923490673303604,
0.01586112007498741,
0.04356645792722702,
0.04995264112949371,
0.024456363171339035,
0.08106988668441772,
-0.03272426873445511,
0.00028798257699236274,
0.00005477556987898424,
0.0052866749465465546,
0.028513344004750252,
0.08777983486652374,
-0.016733819618821144,
-0.0579042062163353,
0.05265985429286957,
0.07001720368862152,
-0.03486234322190285,
-0.05912041664123535,
-0.025872470811009407,
0.0073904539458453655,
-0.005648621823638678,
0.03448988124728203,
0.020679552108049393,
0.05000317841768265,
-0.036898765712976456,
-0.07814917713403702,
0.10101887583732605,
-0.0636715367436409,
0.017893705517053604,
0.01794719323515892,
-0.022645432502031326,
0.0009774366626515985,
-0.006807080935686827,
0.026365822181105614,
0.04559067264199257,
0.07926172018051147,
-0.059484027326107025,
0.006820276379585266,
0.049884725362062454,
0.06980803608894348,
0.0684860348701477,
-0.04392828419804573,
-0.025288039818406105,
0.022982157766819,
0.012314321473240852,
0.0005136837717145681,
0.08304598927497864,
-0.016789274290204048,
0.0021615216974169016,
0.027483269572257996,
0.019017567858099937,
0.05428711697459221,
0.05213503912091255,
-0.07995837181806564,
1.55217361769062e-33,
0.010862817987799644,
0.007971353828907013,
0.09978845715522766,
-0.04913565143942833,
0.07929308712482452,
-0.016866818070411682,
-0.024756960570812225,
-0.03868168219923973,
-0.016630101948976517,
-0.031403541564941406,
0.002790220081806183,
0.0012824424775317311,
-0.044883351773023605,
0.09748329222202301,
0.0017630700021982193,
-0.008551321923732758,
0.01416266243904829,
-0.03970926254987717,
0.06014788895845413,
-0.06870827078819275,
0.09029338508844376,
0.016364485025405884,
-0.04090077057480812,
0.014576978981494904,
0.014768688008189201,
0.03165195882320404,
0.05298764631152153,
-0.01371056493371725,
0.028642825782299042,
0.02117088995873928,
-0.08027411252260208,
0.05979795753955841,
-0.0031052299309521914,
0.14540289342403412,
-0.024152928963303566,
-0.02274468168616295,
-0.07470054924488068,
-0.06677480787038803,
0.045390304177999496,
-0.01022953912615776,
0.014473317191004753,
-0.002200185554102063,
0.014235167764127254,
-0.0812639445066452,
-0.032482653856277466,
-0.07327219843864441,
0.007504334207624197,
0.0062232473865151405,
-0.012656068429350853,
-0.09538780152797699,
-0.04681040346622467,
0.0018641800852492452,
-0.00873104389756918,
-0.03044830821454525,
0.05947943404316902,
-0.03412492573261261,
0.06714988499879837,
0.059446465224027634,
-0.011095982044935226,
0.038754381239414215,
0.04152357578277588,
-0.027160057798027992,
-0.017877234145998955,
-0.005751641001552343,
0.016540544107556343,
0.00003937325527658686,
-0.04981973022222519,
0.10962875187397003,
0.04844188317656517,
-0.03139625862240791,
0.015239509753882885,
0.05540380999445915,
-0.036217205226421356,
-0.023166170343756676,
0.06264158338308334,
-0.01020707469433546,
0.02146427519619465,
-0.06495359539985657,
-0.09995274990797043,
0.0052252039313316345,
-0.005917862057685852,
-0.015970304608345032,
-0.12044352293014526,
0.09863398969173431,
-0.03717951476573944,
-0.05544499307870865,
-0.005073105450719595,
-0.012040968984365463,
0.010198123753070831,
0.0037474622949957848,
-0.0859883725643158,
-0.06106150150299072,
-0.007601783610880375,
-0.010997750796377659,
-0.06199900060892105,
-1.8920843396201928e-33,
-0.04210284724831581,
0.048142749816179276,
-0.0595763623714447,
0.05691543594002724,
0.03546157851815224,
-0.05051779747009277,
-0.05344812944531441,
-0.06800487637519836,
-0.08229103684425354,
0.0077969785779714584,
-0.0950440838932991,
-0.027536971494555473,
0.0229648444801569,
0.025632791221141815,
0.011058494448661804,
-0.023790230974555016,
0.029229070991277695,
-0.043598536401987076,
0.03874889388680458,
-0.07375571131706238,
0.023897284641861916,
0.04984511807560921,
-0.07614272087812424,
0.013257361948490143,
0.09323855489492416,
0.06915312260389328,
-0.08144950866699219,
0.06573040038347244,
0.04349373281002045,
0.054001640528440475,
-0.06448158621788025,
-0.06400539726018906,
-0.03414853662252426,
-0.02706955559551716,
0.06295668333768845,
0.08751904964447021,
0.03612013906240463,
0.004620460793375969,
-0.06562524288892746,
-0.020878443494439125,
0.0507596917450428,
-0.10633203387260437,
-0.08659936487674713,
0.03917166218161583,
0.06516045331954956,
0.009091390296816826,
-0.04205242544412613,
0.06202615797519684,
-0.0794505625963211,
0.040988702327013016,
-0.010990642011165619,
0.04864579066634178,
-0.06440796703100204,
-0.060349393635988235,
-0.10218856483697891,
-0.04876285418868065,
0.07590678334236145,
0.03502148389816284,
0.0843031257390976,
-0.04263467341661453,
-0.051908958703279495,
-0.09640020877122879,
-0.006520713679492474,
0.06071290001273155,
-0.0643400102853775,
0.042289670556783676,
-0.007748713716864586,
0.03195859119296074,
0.10182950645685196,
0.04371143877506256,
0.04878326877951622,
0.026502016931772232,
-0.026792706921696663,
0.0007748978096060455,
0.0363156683743,
0.012351758778095245,
-0.002087767468765378,
-0.01990770362317562,
-0.004474112298339605,
-0.04063641279935837,
-0.02708357386291027,
0.016717052087187767,
0.07172071933746338,
0.05059374123811722,
0.10268624871969223,
0.029355840757489204,
0.02702420949935913,
0.09106078743934631,
0.03108457289636135,
0.016792358830571175,
-0.013629166409373283,
0.029676739126443863,
-0.06989973038434982,
0.06652587652206421,
-0.10697781294584274,
-4.156121491405429e-8,
-0.10069344937801361,
0.0484379306435585,
0.03653780370950699,
0.008408820256590843,
0.059624843299388885,
-0.046901822090148926,
0.08953497558832169,
-0.002786205615848303,
0.02182595804333687,
0.018714167177677155,
0.06099136918783188,
-0.04419349506497383,
0.04260224103927612,
-0.07142585515975952,
0.0022220220416784286,
0.026001732796430588,
-0.0017333164578303695,
-0.054789312183856964,
0.024784713983535767,
-0.037702251225709915,
0.012891022488474846,
-0.02651679702103138,
-0.04523345082998276,
0.029444092884659767,
-0.001816934091039002,
-0.10647711902856827,
-0.039110004901885986,
0.021478038281202316,
0.01189457904547453,
0.09382998198270798,
-0.04003891348838806,
0.0341852642595768,
0.08931450545787811,
-0.06337198615074158,
-0.03854728490114212,
0.061880290508270264,
0.02714366465806961,
-0.006334592122584581,
0.09305618703365326,
-0.08415941894054413,
-0.054790474474430084,
0.07422513514757156,
-0.03903034329414368,
-0.010958467610180378,
-0.016123078763484955,
-0.07224865257740021,
0.027625013142824173,
-0.08034505695104599,
0.07365009933710098,
-0.04659570753574371,
-0.024547260254621506,
0.057063858956098557,
-0.043731939047575,
0.0726582258939743,
0.0286738071590662,
-0.029801957309246063,
0.04323854669928551,
-0.14923571050167084,
0.03356093913316727,
0.1020725890994072,
-0.027194427326321602,
-0.005105193238705397,
-0.07980892062187195,
-0.03267228975892067
] | 0 | 0 | 0 | 0 |
|
Benchmarking Vision Language Model Unlearning via Fictitious Facial Identity Dataset | https://openreview.net/forum?id=0y3hGn1wOk | [
"Yingzi Ma",
"Jiongxiao Wang",
"Fei Wang",
"Siyuan Ma",
"Jiazhao Li",
"Jinsheng Pan",
"Xiujun Li",
"Furong Huang",
"Lichao Sun",
"Bo Li",
"Yejin Choi",
"Muhao Chen",
"Chaowei Xiao"
] | Poster | Machine unlearning has emerged as an effective strategy for forgetting specific information in the training data. However, with the increasing integration of visual data, privacy concerns in Vision Language Models (VLMs) remain underexplored. To address this, we introduce Facial Identity Unlearning Benchmark (FIUBench), a novel VLM unlearning benchmark designed to robustly evaluate the effectiveness of unlearning algorithms under the Right to be Forgotten setting. Specifically, we formulate the VLM unlearning task via constructing the Fictitious Facial Identity VQA dataset and apply a two-stage evaluation pipeline that is designed to precisely control the sources of information and their exposure levels. In terms of evaluation, since VLM supports various forms of ways to ask questions with the same semantic meaning, we also provide robust evaluation metrics including membership inference attacks and carefully designed adversarial privacy attacks to evaluate the performance of algorithms. Through the evaluation of four baseline VLM unlearning algorithms within FIUBench, we find that all methods remain limited in their unlearning performance, with significant trade-offs between model utility and forget quality. Furthermore, our findings also highlight the importance of privacy attacks for robust evaluations. We hope FIUBench will drive progress in developing more effective VLM unlearning algorithms. | Machine Unlearning, Vision Language Model, Privacy | A new benchmark to robustly evaluate vision language model unlearning under the Right to be Forgotten setting. | 12,100 | 2411.03554 | [
-0.0832754373550415,
-0.042848363518714905,
-0.049344226717948914,
0.07396692037582397,
0.10405613481998444,
0.03429168090224266,
-0.006044644396752119,
-0.058552972972393036,
-0.005153378006070852,
-0.06336899846792221,
-0.015651358291506767,
-0.152919739484787,
0.029835356399416924,
0.005847329739481211,
0.000954856164753437,
-0.014685508795082569,
0.05826375260949135,
0.11708241701126099,
-0.11453312635421753,
-0.030169285833835602,
0.009774028323590755,
0.018644459545612335,
0.08109979331493378,
0.0015250708675011992,
0.06251416355371475,
-0.02880595438182354,
0.010745694860816002,
-0.04442352056503296,
0.0027879050467163324,
-0.04900599643588066,
0.050912290811538696,
-0.04404853284358978,
-0.03554726019501686,
0.05794442817568779,
-0.02585575357079506,
0.0060058231465518475,
-0.033848658204078674,
0.035144053399562836,
-0.0014080348191782832,
-0.046911146491765976,
-0.074821338057518,
-0.06485025584697723,
-0.03739490360021591,
-0.0028437806759029627,
0.047721631824970245,
0.03019523248076439,
0.009777303785085678,
0.032787300646305084,
-0.028029363602399826,
-0.046138644218444824,
-0.05208627134561539,
-0.10117866843938828,
-0.047505371272563934,
0.05148348584771156,
-0.06347448378801346,
-0.002656772034242749,
-0.0029558702372014523,
0.007147232536226511,
-0.02852153219282627,
0.08543843030929565,
-0.06940163671970367,
-0.09958905726671219,
-0.03236645832657814,
0.09288771450519562,
-0.11343973129987717,
0.04066793620586395,
0.03678593039512634,
-0.05763285607099533,
-0.025023557245731354,
0.0018508840585127473,
-0.007579296361654997,
0.07522333413362503,
-0.0026158958207815886,
0.01924997940659523,
0.06120197847485542,
0.08599478006362915,
0.017262952402234077,
-0.07245554029941559,
0.04438953101634979,
-0.07737495750188828,
0.024230889976024628,
0.0014365449314936996,
0.04495570436120033,
0.0336107574403286,
0.09018650650978088,
0.03492909297347069,
-0.03135763481259346,
0.02339058555662632,
-0.03120257332921028,
0.02970563806593418,
-0.06028887629508972,
-0.048067741096019745,
0.026459241285920143,
-0.024817438796162605,
0.12231896817684174,
0.01168392226099968,
-0.03436584770679474,
-0.025152655318379402,
-0.05106224864721298,
0.061342157423496246,
-0.045132674276828766,
0.03680436685681343,
-0.013406690210103989,
-0.042010657489299774,
0.032495565712451935,
0.012468867935240269,
0.09534373134374619,
0.017412154003977776,
0.051387205719947815,
-0.040830839425325394,
0.07364458590745926,
0.03304821625351906,
-0.018916891887784004,
0.02823927439749241,
0.04490664228796959,
-0.004746394231915474,
-0.0027607535012066364,
-0.020933929830789566,
0.04918508604168892,
0.008192591369152069,
0.012482433579862118,
0.02382003329694271,
0.08493533730506897,
0.018106477335095406,
-0.02301708608865738,
-0.047257039695978165,
-0.08902289718389511,
1.4871426988092386e-33,
-0.026973895728588104,
0.02751181833446026,
0.03752004727721214,
-0.047009143978357315,
0.055631496012210846,
-0.025869125500321388,
-0.0440676249563694,
0.04162532463669777,
-0.09803128987550735,
0.028629036620259285,
-0.018387410789728165,
-0.008383350446820259,
-0.05654437094926834,
0.0813746228814125,
-0.005879424512386322,
0.058207590132951736,
-0.000703433935996145,
0.07539916038513184,
-0.04605047032237053,
0.008620882406830788,
0.13495154678821564,
-0.0032292550895363092,
0.017170501872897148,
0.0033197335433214903,
0.008055342361330986,
-0.01298039685934782,
0.07807385176420212,
-0.024952376261353493,
-0.017084460705518723,
0.05326630547642708,
-0.11599961668252945,
0.07351166009902954,
0.05170030519366264,
0.05624714121222496,
0.09478898346424103,
-0.015974031761288643,
0.039665013551712036,
-0.020239194855093956,
0.04386081546545029,
0.055599454790353775,
0.02551756612956524,
0.03949442133307457,
-0.0035871986765414476,
-0.05150188133120537,
-0.07476085424423218,
-0.017154397442936897,
0.000967802363447845,
0.0675203800201416,
-0.03214160352945328,
-0.0022245266009122133,
0.040091484785079956,
0.06143379956483841,
-0.09584513306617737,
-0.014962160028517246,
-0.07690165936946869,
-0.020391933619976044,
0.01578429900109768,
0.01864503137767315,
-0.034980956465005875,
0.006530309095978737,
-0.001164925517514348,
0.06097887456417084,
0.01988181471824646,
0.0562426932156086,
-0.05272659286856651,
-0.039313096553087234,
0.007773805409669876,
-0.0054077510721981525,
-0.03224824741482735,
-0.017989052459597588,
0.0061306431889534,
-0.0008591336663812399,
-0.07262151688337326,
-0.05609585717320442,
0.02660279907286167,
-0.01271792408078909,
0.09316448867321014,
0.008802019990980625,
-0.008970124647021294,
0.04254620149731636,
0.08881653100252151,
0.019837111234664917,
-0.003546085674315691,
-0.06878025829792023,
-0.09657644480466843,
-0.09531381726264954,
0.03300251066684723,
-0.03923655301332474,
-0.012976850382983685,
0.026164602488279343,
-0.01804511807858944,
-0.0498654767870903,
-0.04098520427942276,
-0.05343817174434662,
-0.02764841914176941,
-3.1749573443252116e-33,
-0.018180930987000465,
0.0293866116553545,
-0.05351749435067177,
0.02631240338087082,
-0.0048856972716748714,
-0.06531736254692078,
0.05776485055685043,
0.02192792110145092,
0.03689313679933548,
-0.07469438761472702,
-0.026359589770436287,
-0.011380909010767937,
-0.02282051183283329,
0.002956291427835822,
0.05858299136161804,
-0.01524063479155302,
-0.014111446216702461,
-0.008704905398190022,
-0.07004369795322418,
0.03221709281206131,
-0.021874023601412773,
0.096245676279068,
-0.09286823123693466,
0.0013327337801456451,
-0.008159293793141842,
0.03780960291624069,
-0.062665656208992,
0.05607014149427414,
0.020723186433315277,
-0.02163529396057129,
-0.0488290935754776,
0.017455661669373512,
-0.08584578335285187,
-0.08264228701591492,
-0.053173985332250595,
0.04441438615322113,
0.032877836376428604,
-0.007244163192808628,
-0.03478695824742317,
0.09955020993947983,
0.04548531398177147,
-0.014679024927318096,
-0.18228831887245178,
0.008502986282110214,
0.0030307115521281958,
-0.0814894437789917,
-0.03136901557445526,
-0.045208461582660675,
0.015443379990756512,
0.03009851463139057,
-0.03183451667428017,
0.027064312249422073,
-0.03098984621465206,
0.06414417922496796,
-0.09115368127822876,
-0.09182897955179214,
0.021270515397191048,
0.038846563547849655,
0.040250666439533234,
0.004607021808624268,
-0.021952232345938683,
-0.009651903994381428,
-0.0686984434723854,
0.0008671433315612376,
0.016234906390309334,
-0.05913104489445686,
-0.0363616906106472,
0.012661865912377834,
0.02066768705844879,
-0.020345939323306084,
0.06481786072254181,
-0.08481905609369278,
-0.058289431035518646,
0.033428676426410675,
-0.009949208237230778,
-0.030434617772698402,
0.004339119419455528,
-0.0021485143806785345,
-0.050620708614587784,
0.05295310169458389,
0.08193132281303406,
-0.06049741059541702,
0.04903325438499451,
0.042730581015348434,
0.06531596183776855,
0.059473343193531036,
0.09885895252227783,
-0.01435949094593525,
-0.02820775844156742,
0.05450472980737686,
-0.041656285524368286,
-0.0008938260725699365,
-0.06633347272872925,
0.07368141412734985,
-0.038873981684446335,
-3.8886931008619285e-8,
-0.04760593920946121,
0.008648673072457314,
0.03085465170443058,
0.012855874374508858,
0.004193606320768595,
-0.08196307718753815,
-0.07464510947465897,
0.08707955479621887,
-0.07216997444629669,
-0.030297234654426575,
0.046130452305078506,
0.009762524627149105,
0.04028679430484772,
-0.06639296561479568,
-0.019236497581005096,
0.06877932697534561,
0.000657254655379802,
0.03628215566277504,
0.02917034924030304,
-0.01030889805406332,
-0.04946007579565048,
-0.0033126012422144413,
-0.022007327526807785,
-0.05642179027199745,
0.004506963770836592,
-0.02632763423025608,
-0.03656907007098198,
0.07196373492479324,
0.04622483626008034,
0.08739195764064789,
-0.02134426310658455,
0.03802016004920006,
0.018413998186588287,
-0.029920384287834167,
0.020761284977197647,
0.09831038862466812,
-0.03984800726175308,
-0.06744643300771713,
-0.0010768062202259898,
-0.009215522557497025,
0.04510420560836792,
0.06871441751718521,
0.039555490016937256,
0.031560685485601425,
-0.008286506868898869,
0.03438379243016243,
0.04230010509490967,
-0.10190713405609131,
0.05505896732211113,
0.011934739537537098,
-0.030748650431632996,
-0.0019666599109768867,
-0.04160286486148834,
0.1852971762418747,
0.018059512600302696,
-0.021707763895392418,
0.021580835804343224,
-0.022557351738214493,
0.019988542422652245,
0.11144432425498962,
0.050995174795389175,
0.0627206563949585,
-0.034974947571754456,
-0.05550922453403473
] | https://github.com/safolab-wisc/fiubench | 15 | 0 | 0 | 0 |
A Deep Generative Learning Approach for Two-stage Adaptive Robust Optimization | https://openreview.net/forum?id=CKXul9iX77 | [
"Aron Brenner",
"Rahman Khorramfar",
"Jennifer Z Sun",
"Saurabh Amin"
] | Poster | Two-stage adaptive robust optimization (ARO) is a powerful approach for planning under uncertainty, balancing first-stage decisions with recourse decisions made after uncertainty is realized. To account for uncertainty, modelers typically define a simple uncertainty set over which potential outcomes are considered. However, classical methods for defining these sets unintentionally capture a wide range of unrealistic outcomes, resulting in overly-conservative and costly planning in anticipation of unlikely contingencies. In this work, we introduce AGRO, a solution algorithm that performs adversarial generation for two-stage adaptive robust optimization using a variational autoencoder. AGRO generates high-dimensional contingencies that are simultaneously adversarial and realistic, improving the robustness of first-stage decisions at a lower planning cost than standard methods. To ensure generated contingencies lie in high-density regions of the uncertainty distribution, AGRO defines a tight uncertainty set as the image of "latent" uncertainty sets under the VAE decoding transformation. Projected gradient ascent is then used to maximize recourse costs over the latent uncertainty sets by leveraging differentiable optimization methods. We demonstrate the cost-efficiency of AGRO by applying it to both a synthetic production-distribution problem and a real-world power system expansion setting. We show that AGRO outperforms the standard column-and-constraint algorithm by up to 1.8% in production-distribution planning and up to 8% in power system expansion. | robust optimization, stochastic optimization, discrete optimization, deep learning, unsupervised learning | We solve adaptive robust optimization using a variational autoencoder to generate realistic adversarial contingencies and obtain decisions that are sufficiently robust without being over-conservative. | 12,099 | 2409.03731 | [
-0.07372840493917465,
-0.08528641611337662,
0.0430435985326767,
0.07051496207714081,
-0.01520531252026558,
-0.02674325555562973,
-0.08260214328765869,
-0.0309468824416399,
0.008531900122761726,
-0.042989298701286316,
-0.01868375577032566,
-0.08815323561429977,
0.025447435677051544,
-0.02141720801591873,
-0.10913338512182236,
-0.0015498842112720013,
0.09449625015258789,
0.10298998653888702,
-0.11012333631515503,
-0.007307741791009903,
0.003163497429341078,
0.08595584332942963,
-0.012699909508228302,
-0.018100224435329437,
0.09738770872354507,
-0.03513626009225845,
-0.005423343274742365,
0.021926099434494972,
0.013229026459157467,
-0.037064749747514725,
0.07959441095590591,
-0.03751315549015999,
-0.06132372096180916,
0.00043487438233569264,
-0.04111766815185547,
0.08489395678043365,
-0.0999055951833725,
0.08769917488098145,
0.05772353336215019,
-0.06782122701406479,
-0.0004421748162712902,
0.015681667253375053,
-0.060318510979413986,
-0.04950486496090889,
0.016352729871869087,
0.0007847791421227157,
0.011263314634561539,
-0.06471526622772217,
0.03906405717134476,
-0.015426465310156345,
0.020686166360974312,
-0.05542092025279999,
0.0047503020614385605,
-0.03221828490495682,
-0.038658611476421356,
0.021913250908255577,
0.06459616869688034,
-0.014945557340979576,
-0.013230946846306324,
0.0032303309999406338,
-0.041489046066999435,
-0.12654352188110352,
-0.007269309833645821,
-0.06297680735588074,
-0.0566900297999382,
-0.09014176577329636,
0.029062097892165184,
0.039887186139822006,
-0.010167385451495647,
0.035966530442237854,
0.012773364782333374,
0.06276650726795197,
-0.03357400745153427,
-0.0020382185466587543,
0.06427684426307678,
0.10530710220336914,
0.06712940335273743,
0.0033155151177197695,
0.09418197721242905,
0.008805491961538792,
0.07623135298490524,
0.05752287805080414,
0.004866509698331356,
0.06068389490246773,
0.10498543828725815,
-0.0000023305626655201195,
0.01357319112867117,
0.08309706300497055,
0.07651253044605255,
-0.005790831986814737,
-0.08496816456317902,
-0.043591100722551346,
-0.0038666874170303345,
0.01686588115990162,
0.0015226133400574327,
-0.0006459428695961833,
0.013979804702103138,
-0.11094408482313156,
0.011395473964512348,
0.08872269839048386,
-0.025254836305975914,
0.0516674779355526,
0.08670990914106369,
-0.026701979339122772,
0.020059293136000633,
-0.03230920061469078,
0.018206294625997543,
0.030268928036093712,
-0.00211184099316597,
-0.05602891743183136,
0.07169511914253235,
0.0036922823637723923,
0.08649048209190369,
-0.03884931653738022,
-0.04819539934396744,
0.02506830357015133,
-0.03813180699944496,
-0.006782252807170153,
0.015078684315085411,
0.011171014979481697,
-0.06947349011898041,
-0.029129622504115105,
-0.011068405583500862,
0.06834132969379425,
0.05731721594929695,
-0.06789860129356384,
-0.03381772339344025,
3.397242224039752e-33,
-0.008365168236196041,
0.021760186180472374,
0.03673446923494339,
0.04237917438149452,
0.05866622552275658,
0.04009038954973221,
-0.005112707614898682,
0.027422886341810226,
-0.03393799066543579,
0.008158903568983078,
-0.005003031343221664,
-0.037272628396749496,
-0.15053695440292358,
0.058340080082416534,
0.11109867691993713,
-0.05164294317364693,
-0.0552312396466732,
-0.01031228993088007,
0.0402388796210289,
-0.05661246180534363,
-0.018440527841448784,
-0.09115441888570786,
-0.07663998752832413,
-0.003970781806856394,
0.023767663165926933,
-0.025033481419086456,
0.13768547773361206,
-0.021278878673911095,
-0.04240850359201431,
-0.020327577367424965,
-0.10620469599962234,
0.007143727503716946,
0.03736889734864235,
0.052873872220516205,
0.016840213909745216,
-0.06511861085891724,
-0.03563651069998741,
-0.03405800461769104,
-0.028880387544631958,
0.054318297654390335,
0.06534997373819351,
0.11077259480953217,
0.043982867151498795,
0.053867630660533905,
-0.07168468087911606,
-0.07029981911182404,
0.03876698389649391,
-0.007121922913938761,
-0.07360819727182388,
-0.03309863060712814,
-0.027938609942793846,
0.06392402201890945,
0.03687133640050888,
-0.02205200120806694,
-0.02715206705033779,
-0.04604882746934891,
0.00004387683657114394,
0.0105414604768157,
0.024748103693127632,
-0.003428798168897629,
0.004984715953469276,
0.04079968109726906,
-0.016292205080389977,
0.08261457830667496,
-0.01193227805197239,
0.034713730216026306,
-0.01205372717231512,
0.02704693004488945,
0.04221566021442413,
0.0247762743383646,
-0.009490695782005787,
-0.017102636396884918,
0.03770570456981659,
-0.0710005909204483,
0.03408633545041084,
-0.0007666610763408244,
0.044245559722185135,
0.03161042183637619,
0.020869333297014236,
-0.04380635544657707,
-0.02567162923514843,
0.1072554886341095,
-0.05941286310553551,
0.08821310102939606,
-0.04653367027640343,
-0.06395113468170166,
0.03588871657848358,
0.013265727087855339,
-0.034520652145147324,
0.003620099276304245,
-0.06722605973482132,
-0.055957309901714325,
-0.03449032083153725,
0.04463190212845802,
0.014677436091005802,
-2.3779128487371826e-33,
-0.004012774210423231,
0.015357240103185177,
0.009738507680594921,
0.0058562313206493855,
0.06185954809188843,
-0.003939558751881123,
-0.01921783573925495,
-0.004328335169702768,
0.01641797088086605,
-0.09063532948493958,
-0.1461990624666214,
0.07626821100711823,
0.08888373523950577,
-0.04045514017343521,
0.014209193177521229,
-0.03547132387757301,
-0.006727172061800957,
0.04719853028655052,
0.014006526209414005,
0.03297697380185127,
0.046504270285367966,
0.07410130649805069,
-0.07353413105010986,
0.03662395849823952,
0.0038120620884001255,
0.018190259113907814,
-0.0561998151242733,
0.07640280574560165,
-0.042634498327970505,
0.03238072618842125,
-0.0030159589368849993,
-0.005503560416400433,
0.028754791244864464,
-0.008077322505414486,
-0.039976246654987335,
0.05153016746044159,
0.03336876630783081,
-0.018704764544963837,
-0.04257485643029213,
0.08138563483953476,
0.0173680130392313,
-0.05448460951447487,
-0.028055081143975258,
0.012566632591187954,
0.018892820924520493,
0.006372664589434862,
0.02839105762541294,
-0.06720805168151855,
0.0406072698533535,
0.00445684464648366,
-0.017064418643712997,
0.06705589592456818,
-0.11157839000225067,
0.04379827901721001,
-0.025186436250805855,
-0.08154883235692978,
0.03417731076478958,
-0.053166329860687256,
0.1150854229927063,
0.059500955045223236,
-0.04108787700533867,
-0.0309988372027874,
0.019811157137155533,
-0.020167892798781395,
-0.04889892786741257,
0.03438597172498703,
-0.06102481856942177,
0.0019190243910998106,
0.038104694336652756,
0.02719065174460411,
0.01689806953072548,
-0.0040474398992955685,
0.05196129158139229,
0.03670169413089752,
0.03846045583486557,
-0.059687938541173935,
0.030547207221388817,
0.021715693175792694,
0.06139678880572319,
-0.06329570710659027,
0.040113046765327454,
-0.00680264038965106,
0.038565777242183685,
0.06918082386255264,
0.03083418682217598,
0.05337643623352051,
0.033324312418699265,
0.016402829438447952,
-0.012885689735412598,
0.042612429708242416,
-0.0202681552618742,
-0.019468452781438828,
0.032088860869407654,
0.015393003821372986,
0.009369418025016785,
-4.3996092102815965e-8,
-0.02077798917889595,
0.05767447501420975,
0.017399167641997337,
-0.037498701363801956,
0.024074681103229523,
-0.0786980614066124,
-0.04919693246483803,
0.05503300949931145,
0.043016690760850906,
-0.05377311632037163,
0.03999810665845871,
-0.01385529711842537,
0.06160812824964523,
-0.07538259029388428,
-0.0332590714097023,
0.11873627454042435,
0.0283010583370924,
0.012749552726745605,
-0.002188290935009718,
-0.010761614888906479,
-0.013286215253174305,
0.05764852091670036,
-0.054987769573926926,
-0.07455485314130783,
0.08281366527080536,
-0.11865917593240738,
-0.03918714076280594,
0.01853705197572708,
0.035460807383060455,
-0.0005715584265999496,
-0.029551707208156586,
-0.01786264404654503,
0.04740600660443306,
0.03354281187057495,
-0.011486160568892956,
0.0923057496547699,
-0.022669056430459023,
-0.08777640759944916,
-0.006471439730376005,
-0.09397699683904648,
0.046451762318611145,
0.06365659087896347,
-0.06115264445543289,
0.011406105943024158,
-0.03288383409380913,
0.02831329219043255,
0.005240654572844505,
-0.08488036692142487,
0.029660223051905632,
-0.005774079822003841,
0.06141756474971771,
0.03355199471116066,
0.043965891003608704,
0.0471908338367939,
0.06361958384513855,
-0.04412119463086128,
0.030087972059845924,
-0.1278621107339859,
0.04533739387989044,
0.06564880162477493,
-0.0222028698772192,
-0.008804609067738056,
-0.08993514627218246,
-0.015822604298591614
] | 0 | 0 | 0 | 0 |
|
Fundamental Limitations on Subquadratic Alternatives to Transformers | https://openreview.net/forum?id=T2d0geb6y0 | [
"Josh Alman",
"Hantao Yu"
] | Poster | The Transformer architecture is widely deployed in many popular and impactful Large Language Models. At its core is the attention mechanism for calculating correlations between pairs of tokens. Performing an attention computation takes quadratic time in the input size, and had become the time bottleneck for transformer operations. In order to circumvent this, researchers have used a variety of approaches, including designing heuristic algorithms for performing attention computations faster, and proposing alternatives to the attention mechanism which can be computed more quickly. For instance, state space models such as Mamba were designed to replace attention with an almost linear time alternative.
In this paper, we prove that any such approach cannot perform important tasks that Transformer is able to perform (assuming a popular conjecture from fine-grained complexity theory). We focus on document similarity tasks, where one is given as input many documents and would like to find a pair which is (approximately) the most similar. We prove that Transformer is able to perform this task, and we prove that this task cannot be performed in truly subquadratic time by any algorithm. Thus, any model which can be evaluated in subquadratic time – whether because of subquadratic-time heuristics for attention, faster attention replacements like Mamba, or any other reason – cannot perform this task. In other words, in order to perform tasks that (implicitly or explicitly) involve document similarity, one may as well use Transformer and cannot avoid its quadratic running time. | Large Language Models, Transformers, Fine-grained complexity theory, Document similarity, Hardness of Approximation, Fast attention computation | null | 12,091 | 2410.04271 | [
-0.11797571182250977,
-0.010366602800786495,
-0.018209367990493774,
-0.021814322099089622,
-0.019102348014712334,
-0.025504998862743378,
-0.07328404486179352,
0.050603870302438736,
0.0757879689335823,
-0.03361612185835838,
-0.04616688936948776,
0.02298247255384922,
0.05096212029457092,
0.07324185222387314,
0.007151396945118904,
-0.0036101462319493294,
0.01840500719845295,
0.01402751449495554,
-0.10139738768339157,
-0.11267215758562088,
0.07883674651384354,
-0.02644643560051918,
-0.058668918907642365,
-0.015811588615179062,
0.04986049234867096,
0.044026631861925125,
-0.03132689744234085,
-0.10721830278635025,
-0.029402149841189384,
0.020116683095693588,
0.01693163998425007,
0.013330768793821335,
0.02021907828748226,
0.01759159378707409,
-0.0691523402929306,
0.05220497027039528,
-0.04339863359928131,
-0.04047176614403725,
0.05399465188384056,
-0.06889563798904419,
-0.007425707299262285,
-0.018127402290701866,
0.03431160748004913,
0.03416789695620537,
0.023251065984368324,
0.002127084182575345,
0.05058138445019722,
-0.0027276920154690742,
-0.0209613386541605,
-0.012829997576773167,
-0.08099042624235153,
-0.060589488595724106,
-0.042338356375694275,
0.06678419560194016,
-0.013726851902902126,
0.03943224251270294,
0.03837570548057556,
0.03663422912359238,
-0.014022917486727238,
-0.05943452566862106,
-0.08114991337060928,
-0.052816443145275116,
0.0018933506216853857,
-0.06208939477801323,
-0.048365090042352676,
0.018417760729789734,
0.05919494852423668,
-0.0008709912071935833,
-0.03645291179418564,
0.07354719936847687,
-0.011340770870447159,
0.10241164267063141,
-0.016059737652540207,
0.06546996533870697,
0.02705247700214386,
-0.033864106982946396,
0.06960728764533997,
-0.04134424403309822,
0.0756826251745224,
-0.07786019891500473,
0.030269518494606018,
-0.04005097225308418,
0.05251007527112961,
0.01781325228512287,
0.1238166019320488,
-0.03875310719013214,
-0.00008014613558771089,
-0.014863673597574234,
0.037183672189712524,
-0.028947191312909126,
-0.06100493669509888,
-0.10662899911403656,
0.06663250178098679,
-0.047618500888347626,
0.007536254357546568,
0.024919288232922554,
0.014130346477031708,
0.046761371195316315,
0.012001699768006802,
0.08276078850030899,
0.020329734310507774,
0.05105941742658615,
0.08744645863771439,
-0.05481584742665291,
0.015552873723208904,
0.013836151920258999,
0.03428928181529045,
0.020772142335772514,
-0.0014737071469426155,
-0.10596608370542526,
0.015238502994179726,
-0.038831714540719986,
0.007419600151479244,
0.017636431381106377,
0.07839147746562958,
-0.017166152596473694,
0.08299938589334488,
0.006306207738816738,
-0.048916786909103394,
0.04540111497044563,
0.03464176878333092,
0.05228700116276741,
0.01781625859439373,
0.013598250225186348,
0.030333606526255608,
0.016786692664027214,
-0.01760965585708618,
4.561909537206997e-33,
-0.031799860298633575,
0.08021512627601624,
0.00643612165004015,
0.046313922852277756,
-0.0065459455363452435,
-0.025491107255220413,
0.029033878818154335,
0.030287323519587517,
0.002650015987455845,
0.009420862421393394,
-0.013560758903622627,
0.09006191045045853,
-0.011194190941751003,
0.015943914651870728,
0.012778657488524914,
-0.028401492163538933,
0.0189147237688303,
-0.0040300507098436356,
-0.06714598834514618,
-0.12221825122833252,
0.043926987797021866,
-0.017183203250169754,
-0.010163112543523312,
-0.03575008362531662,
-0.047171324491500854,
-0.05392098426818848,
0.07770570367574692,
-0.1512470841407776,
0.00601342273876071,
-0.0357615128159523,
-0.08720225840806961,
0.04481375589966774,
-0.04621367156505585,
0.017555465921759605,
0.032968901097774506,
-0.10008043050765991,
0.028803395107388496,
-0.11116640269756317,
0.0005640496383421123,
-0.05949777364730835,
-0.029051193967461586,
0.08181075751781464,
-0.013164913281798363,
-0.050918612629175186,
-0.02558094635605812,
0.007096502929925919,
0.030964385718107224,
0.020026469603180885,
0.03898385167121887,
-0.03951720893383026,
0.11245401203632355,
0.01738853007555008,
-0.08268778771162033,
0.0453755259513855,
0.10557978600263596,
-0.006490281317383051,
0.035692207515239716,
0.07768631726503372,
0.11503300815820694,
0.11465813964605331,
-0.009134448133409023,
0.00537320040166378,
0.0024823679123073816,
0.03642961382865906,
0.006691566668450832,
0.000710438413079828,
-0.03248368576169014,
0.08556661754846573,
0.05612199380993843,
0.06322253495454788,
-0.011076676659286022,
0.0018554640701040626,
-0.05943375825881958,
-0.07099580764770508,
0.08082553744316101,
0.010078074410557747,
0.100497767329216,
-0.06295979022979736,
-0.03916323184967041,
-0.04567527398467064,
-0.08162622153759003,
-0.055218808352947235,
0.05580953508615494,
-0.026714367792010307,
-0.04126090183854103,
-0.034370459616184235,
0.05572504177689552,
-0.07996141165494919,
-0.07230228930711746,
-0.053927402943372726,
-0.02093176171183586,
0.051604025065898895,
0.0014586589531973004,
-0.016794269904494286,
0.02159329503774643,
-3.9141875379909464e-33,
-0.04699509218335152,
-0.06511110067367554,
-0.020688045769929886,
0.0765811949968338,
-0.014574854634702206,
-0.08171987533569336,
-0.002711840905249119,
-0.026198310777544975,
-0.014336220920085907,
-0.03442014381289482,
-0.09802931547164917,
-0.06793300062417984,
0.09234670549631119,
-0.05825478583574295,
0.07254207134246826,
0.02295733615756035,
0.05475115776062012,
0.015895603224635124,
-0.05373087152838707,
0.049779802560806274,
0.021699465811252594,
0.10082367807626724,
-0.07305239886045456,
-0.04256339743733406,
-0.04115967079997063,
0.058760903775691986,
-0.08445193618535995,
0.004426280036568642,
0.0033835500944405794,
-0.023925576359033585,
-0.10410396754741669,
-0.015619650483131409,
-0.059393398463726044,
-0.046416617929935455,
-0.07020271569490433,
-0.0021708812564611435,
0.0027677228208631277,
0.01860729046165943,
-0.033525727689266205,
0.027987083420157433,
0.015436978079378605,
-0.0525953471660614,
-0.010489615611732006,
0.0155904870480299,
-0.035598065704107285,
-0.029991406947374344,
-0.10311350971460342,
-0.020482942461967468,
-0.01847066543996334,
0.001876689144410193,
0.0031420376617461443,
-0.0003595263697206974,
-0.04509596899151802,
-0.03276916593313217,
-0.032332852482795715,
-0.03443363681435585,
-0.06628905981779099,
-0.00422733323648572,
0.05305600166320801,
-0.012522077187895775,
-0.01818360947072506,
0.000096469433628954,
0.037202246487140656,
-0.044398948550224304,
0.06317044794559479,
-0.037729326635599136,
0.015399204567074776,
-0.03359602391719818,
0.030663440003991127,
0.08503487706184387,
0.07682004570960999,
-0.011132165789604187,
-0.014781533740460873,
-0.02005886472761631,
0.010329893790185452,
0.04656032845377922,
0.11072687804698944,
-0.025629540905356407,
-0.001907605561427772,
-0.051421232521533966,
-0.013889661058783531,
0.0848882719874382,
0.010583487339317799,
-0.04678064212203026,
-0.05957507714629173,
0.07327999174594879,
0.10416091233491898,
-0.02495933696627617,
0.07626601308584213,
0.01520420704036951,
-0.037091922014951706,
0.05849749594926834,
0.06285300850868225,
0.03220708295702934,
0.0027239432092756033,
-4.9622471465227136e-8,
-0.06113959103822708,
0.009996823966503143,
-0.08196850121021271,
-0.005550006870180368,
0.031014110893011093,
-0.05605059862136841,
-0.010281210765242577,
0.08353390544652939,
-0.048334892839193344,
0.016098184511065483,
0.08933655917644501,
-0.05412587523460388,
-0.0017871067393571138,
0.013190967962145805,
-0.040276046842336655,
0.0458475686609745,
0.04731713607907295,
-0.07641701400279999,
-0.052143167704343796,
-0.0010857315501198173,
-0.05380924418568611,
0.06484027951955795,
0.008339022286236286,
0.08015601336956024,
-0.02431563474237919,
-0.008919994346797466,
0.010946228168904781,
0.010601735673844814,
0.0723532885313034,
0.001374573796056211,
-0.017227454110980034,
0.06782898306846619,
0.0027717887423932552,
-0.0252155102789402,
0.05205773562192917,
0.005683493334800005,
0.08334403485059738,
0.03705872595310211,
-0.052438776940107346,
0.007937665097415447,
0.08379258215427399,
0.02440512180328369,
-0.07326582074165344,
0.08622931689023972,
0.12776724994182587,
-0.03698237985372543,
0.032508354634046555,
-0.08224405348300934,
-0.005408284720033407,
-0.04096156358718872,
0.048754818737506866,
-0.0692453533411026,
-0.05142871290445328,
0.01480938121676445,
-0.038679372519254684,
0.053528450429439545,
0.04349223151803017,
-0.014161917380988598,
0.04000838100910187,
0.010525766760110855,
0.04417871683835983,
0.02619076706469059,
-0.008369393646717072,
0.01753176562488079
] | 0 | 0 | 0 | 0 |
|
Cocoon: Robust Multi-Modal Perception with Uncertainty-Aware Sensor Fusion | https://openreview.net/forum?id=DKgAFfCs5F | [
"Minkyoung Cho",
"Yulong Cao",
"Jiachen Sun",
"Qingzhao Zhang",
"Marco Pavone",
"Jeong Joon Park",
"Heng Yang",
"Zhuoqing Mao"
] | Poster | An important paradigm in 3D object detection is the use of multiple modalities to enhance accuracy in both normal and challenging conditions, particularly for long-tail scenarios. To address this, recent studies have explored two directions of adaptive approaches: MoE-based adaptive fusion, which struggles with uncertainties arising from distinct object configurations, and late fusion for output-level adaptive fusion, which relies on separate detection pipelines and limits comprehensive understanding. In this work, we introduce Cocoon, an object- and feature-level uncertainty-aware fusion framework. The key innovation lies in uncertainty quantification for heterogeneous representations, enabling fair comparison across modalities through the introduction of a feature aligner and a learnable surrogate ground truth, termed feature impression. We also define a training objective to ensure that their relationship provides a valid metric for uncertainty quantification. Cocoon consistently outperforms existing static and adaptive methods in both normal and challenging conditions, including those with natural and artificial corruptions. Furthermore, we show the validity and efficacy of our uncertainty metric across diverse datasets. | Multi-modal perception; Sensor fusion; Robustness; Uncertainty quantification | Cocoon is a new object- and feature-level uncertainty-aware multimodal fusion framework designed for 3D OD tasks. At the core of Cocoon are uncertainty quantification and comparison for heterogeneous representations. | 12,089 | 2410.12592 | [
-0.0014624808682128787,
-0.1113685742020607,
0.05022967979311943,
0.07283633947372437,
0.0670790821313858,
-0.0485021136701107,
-0.000888639478944242,
-0.06977047026157379,
0.008640462532639503,
-0.13281020522117615,
-0.038396187126636505,
-0.13646483421325684,
0.0038366480730473995,
0.04846026748418808,
-0.03030114807188511,
-0.010980846360325813,
0.10552872717380524,
0.17027759552001953,
-0.1117330938577652,
0.002283954992890358,
0.031935859471559525,
-0.024103716015815735,
0.006949175149202347,
0.030684003606438637,
0.033460985869169235,
0.04383035749197006,
-0.03923347592353821,
-0.0846201479434967,
0.01222814992070198,
-0.0622800849378109,
0.017638642340898514,
-0.007440078537911177,
-0.01965017430484295,
-0.007499196100980043,
0.05145146697759628,
0.00548684922978282,
0.029608190059661865,
-0.012661888264119625,
0.029399041086435318,
-0.039281655102968216,
-0.017733659595251083,
0.028292136266827583,
-0.00936522800475359,
-0.016382548958063126,
0.016221970319747925,
0.10192105919122696,
0.0436369888484478,
-0.07228147238492966,
0.055283378809690475,
0.05865786224603653,
-0.019240884110331535,
-0.03924454376101494,
-0.00371370161883533,
-0.01492722611874342,
0.004933628253638744,
0.08142243325710297,
0.03089476376771927,
-0.021122710779309273,
0.06481105834245682,
-0.0035465830005705357,
0.027413180097937584,
-0.04636745899915695,
0.02215297892689705,
0.03141419589519501,
0.06424009799957275,
-0.07358450442552567,
0.008781625889241695,
-0.0143020860850811,
0.006174028851091862,
0.01576150208711624,
-0.013584934175014496,
0.03422616049647331,
-0.008213277906179428,
-0.009656058624386787,
-0.02901889942586422,
0.04536306485533714,
0.08611921221017838,
-0.047156188637018204,
0.009199796244502068,
0.009254306554794312,
0.03858349099755287,
0.050096794962882996,
0.021323734894394875,
-0.02100316621363163,
0.13319995999336243,
-0.015079857781529427,
-0.09345381706953049,
0.06543894112110138,
-0.07408903539180756,
0.021588675677776337,
-0.11688682436943054,
-0.07766000181436539,
0.0028682404663413763,
-0.015255958773195744,
0.02902950905263424,
-0.021440481767058372,
0.0035352923441678286,
-0.06604533642530441,
0.05548838526010513,
0.07088328152894974,
-0.03260830417275429,
-0.027622250840067863,
0.006436107214540243,
-0.0060901641845703125,
0.07061906158924103,
-0.02118527889251709,
0.008395751938223839,
0.0302118007093668,
0.03959326446056366,
0.008855915628373623,
0.06037788465619087,
-0.025949930772185326,
0.049513399600982666,
-0.023187140002846718,
0.0008440356468781829,
0.016786273568868637,
-0.024544868618249893,
0.05513959750533104,
0.05559176579117775,
0.024987000972032547,
-0.021640637889504433,
-0.03796686232089996,
0.05001290142536163,
0.04762286692857742,
0.1002284437417984,
-0.027981510385870934,
0.009055058471858501,
2.4272626729785946e-34,
0.026117293164134026,
0.004716929048299789,
0.08024013042449951,
0.004889047238975763,
0.0448622889816761,
0.007930873893201351,
-0.02826310321688652,
0.020977335050702095,
-0.10505634546279907,
0.083065927028656,
-0.006313652731478214,
0.02470659837126732,
-0.07936502993106842,
0.020840153098106384,
0.04798629879951477,
-0.009745893068611622,
-0.08755207806825638,
-0.02614441327750683,
-0.036980729550123215,
0.020964648574590683,
0.02103007398545742,
-0.09766754508018494,
-0.05954092741012573,
0.04475422576069832,
-0.01717047207057476,
0.011112463660538197,
0.07256696373224258,
0.029847316443920135,
-0.03713007643818855,
-0.008572109043598175,
-0.07447170466184616,
0.002031507669016719,
0.03062169812619686,
0.02516036480665207,
0.06946523487567902,
-0.09453818202018738,
-0.05151499807834625,
-0.056775886565446854,
-0.07241851836442947,
-0.013730937615036964,
0.001656153122894466,
0.01595190353691578,
-0.02876446582376957,
-0.02110845036804676,
-0.07204325497150421,
-0.02349221520125866,
0.05794503539800644,
0.00573115237057209,
-0.09672744572162628,
0.007536903955042362,
0.007307595107704401,
0.019602913409471512,
-0.026034489274024963,
-0.02837335132062435,
-0.021399354562163353,
-0.06483427435159683,
-0.03437510505318642,
0.037904027849435806,
0.0716380700469017,
-0.05333757773041725,
-0.03995620831847191,
0.0032735655549913645,
-0.0691828653216362,
-0.0025230671744793653,
0.002587228314951062,
0.1052279993891716,
-0.02126472070813179,
0.06857113540172577,
0.010979445651173592,
0.07858418673276901,
-0.05521922558546066,
0.02050885744392872,
-0.018078817054629326,
-0.02594170719385147,
0.02528557740151882,
-0.000796229753177613,
0.022039448842406273,
-0.018157074227929115,
0.04767701029777527,
0.0025610942393541336,
-0.0325743742287159,
0.058901578187942505,
0.01340122614055872,
-0.013956322334706783,
-0.09506119042634964,
-0.007107939571142197,
0.00463023129850626,
-0.01721869967877865,
-0.08797932416200638,
0.03649505600333214,
-0.04890678822994232,
-0.08104018867015839,
0.017888130620121956,
-0.011276977136731148,
-0.013427654281258583,
-4.868112407613764e-34,
0.0336984246969223,
0.03748202696442604,
0.004177068825811148,
-0.03563303127884865,
-0.010119940154254436,
-0.06360884755849838,
-0.01986393705010414,
0.045351460576057434,
-0.005762310698628426,
-0.11733313649892807,
-0.08565932512283325,
-0.005729347001761198,
0.0871284082531929,
0.0037379474379122257,
-0.043568260967731476,
0.057559140026569366,
-0.08447247743606567,
-0.010849183425307274,
-0.0346435122191906,
0.018646273761987686,
0.051723774522542953,
0.03865734115242958,
-0.08939027041196823,
0.023388754576444626,
-0.08574474602937698,
0.09231378883123398,
-0.06348620355129242,
0.01413718331605196,
0.02136867865920067,
-0.04585021361708641,
0.03448423370718956,
-0.03346174955368042,
0.023127302527427673,
-0.03375723212957382,
-0.04083377867937088,
-0.022137906402349472,
-0.028532693162560463,
-0.049618568271398544,
0.016317548230290413,
0.029282012954354286,
-0.000246548093855381,
-0.05704354867339134,
-0.08120609819889069,
0.01029213611036539,
0.03593144938349724,
-0.010818000882863998,
0.018921636044979095,
-0.015226414427161217,
0.023654216900467873,
0.043619513511657715,
-0.05665126442909241,
0.03303458169102669,
-0.13658860325813293,
0.06657467037439346,
0.013267512433230877,
-0.059429336339235306,
-0.0570761114358902,
-0.023174723610281944,
0.10873828828334808,
0.03265579417347908,
-0.022120071575045586,
-0.04370610788464546,
-0.0018098816508427262,
0.0639495924115181,
0.023130279034376144,
0.07331535965204239,
-0.051172610372304916,
0.0037924956995993853,
0.06378570199012756,
0.05102511867880821,
0.06052182987332344,
-0.012225357815623283,
-0.014648495241999626,
-0.04929187893867493,
-0.020018571987748146,
-0.03975497558712959,
0.07177479565143585,
-0.01646224781870842,
0.02831386588513851,
0.05935361608862877,
0.03030705451965332,
-0.061653982847929,
0.11172804236412048,
0.13436278700828552,
0.10379068553447723,
0.11364123225212097,
0.018591364845633507,
0.04212057590484619,
0.012080936692655087,
0.05032626539468765,
-0.03992954641580582,
0.0012593697756528854,
-0.010271694511175156,
0.02820587158203125,
0.0037527051754295826,
-4.352348881297985e-8,
-0.06001380831003189,
0.053238920867443085,
-0.07025054097175598,
-0.06587442010641098,
-0.020965073257684708,
0.006080699153244495,
-0.0044594407081604,
0.0006772684864699841,
-0.0431194007396698,
-0.07554729282855988,
0.05847513675689697,
-0.06381592154502869,
-0.010536796413362026,
-0.047646984457969666,
0.08068468421697617,
0.06557188928127289,
-0.036011360585689545,
0.01992199383676052,
0.003813657443970442,
0.060570456087589264,
0.023137446492910385,
0.016752198338508606,
0.02704562060534954,
0.014245518483221531,
-0.010204286314547062,
-0.035146232694387436,
-0.0321386493742466,
0.11159010976552963,
0.038195639848709106,
0.005141488276422024,
-0.03495832160115242,
-0.011513558216392994,
0.06404819339513779,
0.016668332740664482,
0.013960453681647778,
0.0938158929347992,
0.012166247703135014,
-0.04092869162559509,
0.010066326707601547,
-0.058951906859874725,
-0.023768603801727295,
0.059199463576078415,
-0.04104457423090935,
0.04834292456507683,
-0.012785958126187325,
0.07655586302280426,
0.061244845390319824,
-0.0940094143152237,
-0.04619584232568741,
0.036536723375320435,
0.02208986133337021,
0.03546213358640671,
-0.06488577276468277,
0.14681071043014526,
0.10865183919668198,
0.03137153014540672,
0.009172875434160233,
-0.08187755942344666,
-0.001732143573462963,
0.08767089992761612,
0.03528791293501854,
-0.05858045443892479,
-0.0414186529815197,
-0.060746319591999054
] | 0 | 0 | 0 | 0 |
|
Decision Tree Induction Through LLMs via Semantically-Aware Evolution | https://openreview.net/forum?id=UyhRtB4hjN | [
"Tennison Liu",
"Nicolas Huynh",
"Mihaela van der Schaar"
] | Poster | Decision trees are a crucial class of models offering robust predictive performance and inherent interpretability across various domains, including healthcare, finance, and logistics. However, current tree induction methods often face limitations such as suboptimal solutions from greedy methods or prohibitive computational costs and limited applicability of exact optimization approaches.
To address these challenges, we propose an evolutionary optimization method for decision tree induction based on genetic programming (GP). Our key innovation is the integration of semantic priors and domain-specific knowledge about the search space into the optimization algorithm.
To this end, we introduce $\texttt{LLEGO}$, a framework that incorporates semantic priors into genetic search operators through the use of Large Language Models (LLMs), thereby enhancing search efficiency and targeting regions of the search space that yield decision trees with superior generalization performance. This is operationalized through novel genetic operators that work with structured natural language prompts, effectively utilizing LLMs as conditional generative models and sources of semantic knowledge. Specifically, we introduce $\textit{fitness-guided}$ crossover to exploit high-performing regions, and $\textit{diversity-guided}$ mutation for efficient global exploration of the search space. These operators are controlled by corresponding hyperparameters that enable a more nuanced balance between exploration and exploitation across the search space. Empirically, we demonstrate across various benchmarks that $\texttt{LLEGO}$ evolves superior-performing trees compared to existing tree induction methods, and exhibits significantly more efficient search performance compared to conventional GP approaches. | decision trees, LLMs, genetic programming | null | 12,082 | 2503.14217 | [
-0.0778694823384285,
0.015735287219285965,
0.016608282923698425,
0.06924469023942947,
0.022259807214140892,
-0.056228045374155045,
0.01571812480688095,
0.0023896025959402323,
-0.047522783279418945,
0.01580018363893032,
-0.07242930680513382,
-0.0950729101896286,
0.026108907535672188,
0.017757195979356766,
-0.008796569891273975,
0.026005279272794724,
0.12395595759153366,
0.06917236000299454,
-0.06467593461275101,
-0.04722995311021805,
0.011805472895503044,
0.11439289152622223,
0.05522779002785683,
0.01942604035139084,
-0.0018981872126460075,
-0.018063662573695183,
0.0335078202188015,
0.01743403635919094,
0.04744870588183403,
0.052078187465667725,
0.0545349158346653,
-0.00017259997548535466,
-0.0013430183753371239,
0.028173131868243217,
-0.06829019635915756,
0.016284378245472908,
-0.09333566576242447,
0.04890792444348335,
0.015302195213735104,
-0.004746206104755402,
0.0004943661042489111,
-0.0046363878063857555,
-0.011436313390731812,
0.015547404997050762,
0.06830517202615738,
-0.0012668977724388242,
-0.061753515154123306,
0.01585724577307701,
-0.0028196752537041903,
0.050388798117637634,
-0.06737980991601944,
-0.07735125720500946,
0.004227634519338608,
-0.0014003196265548468,
-0.018925677984952927,
0.06846757233142853,
0.02632521092891693,
-0.09252890199422836,
-0.09612330794334412,
-0.042111147195100784,
0.0023072981275618076,
-0.14797693490982056,
0.024931859225034714,
-0.06352677196264267,
0.0346490703523159,
-0.010367878712713718,
-0.007089841179549694,
0.05599283427000046,
0.001651874161325395,
-0.026635732501745224,
0.010008597746491432,
0.10837546736001968,
-0.024069810286164284,
0.08927695453166962,
0.013109580613672733,
0.07952146977186203,
0.10133980959653854,
0.02361365035176277,
0.0623977966606617,
-0.06636995077133179,
-0.0391395166516304,
0.019246824085712433,
-0.0032079247757792473,
0.04571731388568878,
0.06862773001194,
0.009016675874590874,
0.04164004698395729,
0.024566449224948883,
0.034218113869428635,
0.06713198125362396,
-0.044037945568561554,
-0.10780970752239227,
0.05666513741016388,
-0.003514904761686921,
-0.04164700582623482,
0.030371904373168945,
-0.060117464512586594,
-0.1071334034204483,
0.012416383251547813,
-0.006780991330742836,
0.03591305762529373,
0.06313181668519974,
0.07685381919145584,
-0.07286031544208527,
-0.07646960020065308,
-0.0650668740272522,
-0.03921041265130043,
0.01621606945991516,
0.02864956110715866,
-0.061711668968200684,
-0.0013871982228010893,
0.10541306436061859,
0.05531134456396103,
0.010637469589710236,
-0.11114311963319778,
0.04321306198835373,
-0.02996857650578022,
-0.006634363904595375,
0.024104557931423187,
0.08618316054344177,
-0.009929873049259186,
0.030464844778180122,
-0.08730009198188782,
0.02259327471256256,
-0.0007453548605553806,
-0.06376200914382935,
-0.09390957653522491,
-4.247985872967236e-34,
-0.020487014204263687,
-0.019459962844848633,
0.05245426669716835,
0.11010377109050751,
0.06015045568346977,
0.0007119260262697935,
-0.019775880500674248,
0.05474467948079109,
-0.02921527810394764,
0.01630418188869953,
-0.14717881381511688,
0.0342988520860672,
-0.05318307504057884,
0.002354166703298688,
0.05316004902124405,
-0.003278931602835655,
-0.10521198064088821,
0.05314774438738823,
-0.07204467803239822,
-0.03677116334438324,
0.053972043097019196,
0.0004465485690161586,
0.0004756665730383247,
-0.03179549053311348,
0.08345558494329453,
0.0007213268545456231,
0.03705652430653572,
-0.06802944093942642,
0.07299219071865082,
0.008304279297590256,
-0.10960482060909271,
-0.04509297013282776,
0.008227773942053318,
0.03717326000332832,
0.008184070698916912,
-0.03385547548532486,
-0.06192480027675629,
-0.022564606741070747,
-0.019357243552803993,
0.10333634167909622,
-0.1010826900601387,
-0.0013037463650107384,
0.034430138766765594,
0.028036944568157196,
-0.054868921637535095,
-0.0705462247133255,
-0.015818094834685326,
0.016227511689066887,
-0.006340222433209419,
-0.0024193639401346445,
-0.007564119528979063,
-0.016750602051615715,
-0.006466665305197239,
-0.040558215230703354,
-0.006641366519033909,
-0.01503394078463316,
-0.021957896649837494,
0.0257261972874403,
-0.021660832688212395,
-0.04351602494716644,
0.02056519128382206,
0.0256541408598423,
0.04014446213841438,
0.063010074198246,
-0.04349958524107933,
-0.015238185413181782,
0.01368075143545866,
0.04949098080396652,
0.023158768191933632,
0.02764621563255787,
0.03110172227025032,
0.04593294486403465,
-0.037815533578395844,
-0.02043650671839714,
0.03561269864439964,
0.04065729305148125,
0.09572511166334152,
-0.03516440466046333,
-0.005008716601878405,
-0.03162567317485809,
-0.051447588950395584,
0.06993281096220016,
-0.00246064574457705,
0.042404137551784515,
0.026669491082429886,
-0.05957493185997009,
0.038141097873449326,
-0.03912802413105965,
-0.01407299842685461,
-0.04727618768811226,
-0.13518153131008148,
-0.001260888297110796,
-0.005820412654429674,
0.03643801808357239,
0.09213729202747345,
-1.2113439897864902e-33,
-0.040839504450559616,
-0.0776815116405487,
0.08587383478879929,
-0.03473416343331337,
-0.04104062169790268,
-0.07224718481302261,
-0.03368423134088516,
-0.037406012415885925,
0.003227436915040016,
-0.022603556513786316,
-0.02298704907298088,
0.06556091457605362,
0.051033779978752136,
0.011875465512275696,
0.0055339583195745945,
0.011591622605919838,
-0.02398357167840004,
0.028197383508086205,
-0.017128635197877884,
0.0894017145037651,
0.033262949436903,
0.13871346414089203,
-0.12795977294445038,
0.0369597002863884,
-0.021077996119856834,
-0.01207734551280737,
-0.029877901077270508,
0.10827776044607162,
0.04108145087957382,
0.009598988108336926,
-0.07674117386341095,
0.004060703329741955,
-0.020448757335543633,
-0.002020788611844182,
-0.06016717851161957,
0.0016193672781810164,
0.08340604603290558,
-0.018435563892126083,
-0.031284503638744354,
0.08693663030862808,
-0.0053185285069048405,
-0.02282111346721649,
-0.028926387429237366,
0.024310272186994553,
-0.024263862520456314,
-0.027048218995332718,
-0.029772920534014702,
-0.059468481689691544,
0.10866530984640121,
0.038812972605228424,
0.07918573170900345,
0.003505973145365715,
-0.02133353054523468,
0.06273461878299713,
-0.0398554801940918,
-0.06784261763095856,
-0.008099190890789032,
-0.040686026215553284,
-0.0039169667288661,
-0.029644446447491646,
-0.04012656956911087,
0.05365496501326561,
0.07803498953580856,
0.02214467152953148,
-0.06551504880189896,
-0.07073602080345154,
-0.027920793741941452,
0.0012248795246705413,
-0.04533965513110161,
-0.0140319112688303,
-0.010937592945992947,
-0.018388589844107628,
0.018120085820555687,
0.0022817126009613276,
0.014991732314229012,
0.01697632484138012,
0.011280977167189121,
-0.018750550225377083,
-0.013591897673904896,
-0.06708993762731552,
0.0793280303478241,
-0.038493767380714417,
0.11489519476890564,
0.09899339079856873,
0.013992641121149063,
-0.046628426760435104,
0.062458738684654236,
0.09576366096735,
0.07876356691122055,
0.004747537896037102,
-0.05388153716921806,
-0.03721441328525543,
-0.03791762888431549,
-0.0031826249323785305,
-0.07477578520774841,
-4.1953608587164126e-8,
-0.041127774864435196,
-0.009055948816239834,
-0.031456295400857925,
0.04747160151600838,
0.044337090104818344,
-0.016062112525105476,
-0.07683169096708298,
0.08336164057254791,
-0.022633537650108337,
-0.043588802218437195,
0.0050191874615848064,
0.08780142664909363,
0.018839837983250618,
-0.04909271001815796,
0.01407712884247303,
0.04371333867311478,
-0.07788565009832382,
-0.04327817261219025,
0.022254960611462593,
0.053418803960084915,
-0.0001697909174254164,
0.05587388575077057,
0.016348719596862793,
0.04778992757201195,
0.06948406249284744,
-0.12200519442558289,
0.0016065165400505066,
-0.007415916305035353,
0.08849957585334778,
-0.01580551639199257,
-0.0439629852771759,
-0.031652890145778656,
-0.0661843940615654,
0.05567343533039093,
0.02498459629714489,
0.06340444833040237,
-0.027647094801068306,
-0.00972724612802267,
-0.06419070810079575,
-0.04012879356741905,
0.08685366809368134,
0.0009042106103152037,
-0.002902104752138257,
0.026387320831418037,
-0.0005110393394716084,
-0.08516857028007507,
-0.015600243583321571,
-0.008358776569366455,
0.02682422287762165,
0.025921935215592384,
-0.023633942008018494,
-0.032702621072530746,
0.04307897016406059,
-0.013425848446786404,
0.04783857241272926,
-0.009242716245353222,
-0.048297397792339325,
-0.02633475326001644,
0.013852381147444248,
0.06827720254659653,
0.05936722829937935,
-0.01076136901974678,
0.026253508403897285,
0.017743322998285294
] | 0 | 0 | 0 | 0 |
|
Efficient Imitation under Misspecification | https://openreview.net/forum?id=fn36V5qsCw | [
"Nicolas Espinosa-Dice",
"Sanjiban Choudhury",
"Wen Sun",
"Gokul Swamy"
] | Poster | Interactive imitation learning (IL) is a powerful paradigm for learning to make sequences of decisions from an expert demonstrating how to perform a task. Prior work in efficient imitation learning has focused on the realizable setting, where the expert's policy lies within the learner's policy class (i.e. the learner can perfectly imitate the expert in all states). However, in practice, perfect imitation of the expert is often impossible due to differences in state information and action space expressiveness (e.g. morphological differences between robots and humans.) In this paper, we consider the more general misspecified setting, where no assumptions are made about the expert policy's realizability. We introduce a novel structural condition, reward-agnostic policy completeness, and prove that it is sufficient for interactive IL algorithms to efficiently avoid the quadratically compounding errors that stymie offline approaches like behavioral cloning. We address an additional practical constraint—the case of limited expert data—and propose a principled method for using additional offline data to further improve the sample-efficiency of interactive IL algorithms. Finally, we empirically investigate the optimal reset distribution in efficient IRL under misspecification with a suite of continuous control tasks. | Inverse Reinforcement Learning, Imitation Learning, Distribution Shift, Policy Completeness | We address interactive imitation learning in the misspecified setting and prove a structural condition under which our efficient imitation algorithm can avoid compounding errors. | 12,080 | 2503.13162 | [
-0.08152967691421509,
-0.09159286320209503,
0.04110334813594818,
0.011235958896577358,
-0.037564415484666824,
-0.04532967135310173,
0.06134093925356865,
-0.00498611805960536,
-0.013365104794502258,
0.07744117081165314,
0.022315753623843193,
-0.005330149084329605,
0.04730159416794777,
0.0098031647503376,
0.024722885340452194,
0.005980688147246838,
0.08278346061706543,
0.0762927234172821,
-0.08499985933303833,
-0.010762718506157398,
0.05459468066692352,
-0.052396222949028015,
0.0229458250105381,
-0.01654515042901039,
0.01480826549232006,
0.012007727287709713,
0.012803195975720882,
-0.09297320246696472,
0.017262298613786697,
-0.03347500413656235,
-0.020522210747003555,
-0.02059660665690899,
0.05735703930258751,
-0.12284734845161438,
0.009550830349326134,
0.0873912051320076,
-0.10771171003580093,
0.003655956592410803,
0.028321396559476852,
-0.0482565201818943,
-0.010803919285535812,
-0.0169308353215456,
-0.022812776267528534,
0.03563975915312767,
0.023821331560611725,
0.046534910798072815,
-0.008435743860900402,
-0.04950887709856033,
0.0575740821659565,
0.05675363168120384,
-0.08927972614765167,
-0.004276118241250515,
0.021115727722644806,
-0.016704251989722252,
0.008763205260038376,
0.015504968352615833,
0.038382843136787415,
0.03108568675816059,
0.055331967771053314,
0.008348586037755013,
-0.0358545258641243,
-0.05750277265906334,
0.009843463078141212,
0.010369876399636269,
-0.0268336720764637,
-0.015361367724835873,
-0.020549258217215538,
-0.025053007528185844,
0.03695132955908775,
0.05958092585206032,
-0.0571138858795166,
0.03578297793865204,
-0.03226073086261749,
0.04232113063335419,
0.01284592505544424,
-0.0631459504365921,
0.008592728525400162,
-0.02960517257452011,
-0.04421509802341461,
-0.04759838804602623,
-0.029331926256418228,
-0.015518064610660076,
-0.0007588855805806816,
0.04478434845805168,
0.03066886216402054,
-0.0123006422072649,
-0.0022971052676439285,
0.05730960890650749,
0.1613498479127884,
0.018598070368170738,
-0.04739911109209061,
-0.004924234934151173,
0.02733715996146202,
-0.050678227096796036,
0.024997210130095482,
0.037391677498817444,
0.019966168329119682,
-0.02071109041571617,
-0.02173546887934208,
0.09278187155723572,
0.02090270258486271,
0.011752444319427013,
-0.002550633391365409,
-0.019904565066099167,
-0.02941115014255047,
-0.06906553357839584,
0.0027874980587512255,
-0.003727337345480919,
0.035244330763816833,
-0.03823297843337059,
-0.002357212593778968,
-0.008872431702911854,
0.0649719387292862,
0.0380493700504303,
0.05329607427120209,
-0.0009693729225546122,
0.012209602631628513,
-0.047215621918439865,
0.005604834295809269,
0.03465526923537254,
0.053688544780015945,
-0.021272709593176842,
0.09174121171236038,
0.04885585606098175,
0.07999230176210403,
-0.03774245083332062,
-0.01580420695245266,
-6.158892448537986e-34,
0.05928400158882141,
0.04000169783830643,
0.057194508612155914,
0.01967124082148075,
0.018361631780862808,
-0.013834971934556961,
0.06990299373865128,
-0.022901376709342003,
-0.03380857780575752,
-0.03219221904873848,
0.04604936018586159,
-0.01579415798187256,
-0.0288232509046793,
0.09839155524969101,
0.043855004012584686,
0.036146827042102814,
-0.00537231657654047,
0.054750822484493256,
-0.03486960753798485,
-0.0016697446117177606,
0.08946292102336884,
-0.025757163763046265,
-0.009132033213973045,
-0.05459129810333252,
-0.00505463732406497,
-0.0395352840423584,
-0.015315324999392033,
-0.030865566805005074,
0.026342734694480896,
-0.0012722262181341648,
-0.08296643942594528,
-0.0361286886036396,
0.0002061189734376967,
0.07850908488035202,
0.023040246218442917,
0.012318700551986694,
0.031152861192822456,
-0.08059507608413696,
-0.04902210831642151,
0.0011912856716662645,
0.023056834936141968,
0.019639261066913605,
0.043748315423727036,
0.03794104978442192,
-0.05256359651684761,
-0.11164462566375732,
0.017421171069145203,
0.022134047001600266,
-0.033958952873945236,
-0.015560168772935867,
0.06072515249252319,
0.07084672152996063,
0.038878586143255234,
-0.11231572180986404,
-0.0192208644002676,
0.015457428060472012,
-0.04756704717874527,
0.022147072479128838,
-0.03044433705508709,
0.012629411183297634,
0.010896038264036179,
0.010710536502301693,
0.020903300493955612,
0.09318473190069199,
-0.009300668723881245,
0.085418701171875,
-0.02567385323345661,
0.009538233280181885,
0.06674829870462418,
-0.0013073401059955359,
-0.012973698787391186,
-0.017590364441275597,
-0.0867149829864502,
-0.08163642883300781,
0.07321883738040924,
-0.02995235100388527,
0.01874113455414772,
-0.08935512602329254,
0.005204938352108002,
0.01493870373815298,
-0.04677101969718933,
-0.019104843959212303,
-0.10028710216283798,
-0.08591239899396896,
-0.015789059922099113,
-0.12588930130004883,
0.04522297531366348,
-0.032430972903966904,
-0.0020827772095799446,
-0.01028361078351736,
-0.08879425376653671,
-0.0694541335105896,
-0.13448910415172577,
0.0456596277654171,
0.019109994173049927,
-1.3160213025497341e-33,
0.028262119740247726,
-0.0525992251932621,
0.09393364191055298,
0.09212604910135269,
-0.08093558251857758,
-0.027829047292470932,
0.01639449968934059,
-0.024005888029932976,
-0.04086682200431824,
-0.057511698454618454,
-0.07401910424232483,
0.020162027329206467,
0.060215383768081665,
-0.022305889055132866,
-0.05070667713880539,
0.004112424794584513,
-0.05440415441989899,
0.04774376004934311,
0.009188065305352211,
0.021516339853405952,
0.04452605918049812,
0.023327618837356567,
-0.0769146978855133,
0.10364927351474762,
-0.07040895521640778,
0.057201772928237915,
-0.10292007774114609,
0.1176762729883194,
0.01945040374994278,
0.03352341428399086,
-0.045348092913627625,
-0.05143854394555092,
-0.066156305372715,
-0.025852318853139877,
-0.03846445679664612,
0.1095513179898262,
0.046890825033187866,
0.102629654109478,
-0.08426433056592941,
0.06770550459623337,
0.0473741739988327,
-0.09244436770677567,
-0.05769523233175278,
-0.026666607707738876,
0.026049263775348663,
-0.07770296186208725,
-0.04604419693350792,
-0.011860130354762077,
-0.0034405351616442204,
-0.0038912007585167885,
-0.0038339111488312483,
0.05552633851766586,
-0.052197523415088654,
-0.01131142396479845,
-0.05695963650941849,
-0.038245175033807755,
-0.0159803107380867,
0.04020354896783829,
0.015789145603775978,
-0.0393148772418499,
-0.020191391929984093,
-0.01465973537415266,
-0.028094053268432617,
0.06404490023851395,
-0.03826067969202995,
-0.0058169495314359665,
0.005699483677744865,
0.06051132455468178,
0.021914508193731308,
-0.017101053148508072,
-0.004283308517187834,
-0.001671392354182899,
-0.04687302187085152,
-0.05763775110244751,
0.11582786589860916,
-0.01571056991815567,
0.03397781401872635,
0.028685109689831734,
-0.008700706996023655,
0.016686515882611275,
0.02274264208972454,
0.010834760963916779,
0.057139672338962555,
0.038803018629550934,
-0.02576645277440548,
0.11790099740028381,
-0.027188051491975784,
0.08074217289686203,
0.024239860475063324,
0.01939377747476101,
0.01627584919333458,
0.037324171513319016,
-0.03652632236480713,
0.046305928379297256,
-0.048093490302562714,
-4.711167633786317e-8,
-0.07949163019657135,
0.07494959980249405,
0.13438351452350616,
0.030596980825066566,
0.08102251589298248,
-0.0076816207729279995,
-0.0878673866391182,
-0.05208398774266243,
-0.07443065196275711,
-0.05983123928308487,
0.032623372972011566,
0.047363363206386566,
0.052605580538511276,
-0.10210955142974854,
-0.0027035840321332216,
0.04457039758563042,
-0.014987869188189507,
-0.02786703035235405,
-0.0000586820715398062,
0.0660376325249672,
0.025346558541059494,
-0.050964195281267166,
-0.04807494580745697,
-0.02798750437796116,
-0.0022084324155002832,
-0.04718577116727829,
-0.031258851289749146,
0.005276670679450035,
0.024086888879537582,
0.07646774500608444,
-0.0076955100521445274,
-0.025064431130886078,
0.06454058736562729,
-0.011078810319304466,
0.08046936243772507,
-0.054031334817409515,
0.06390994787216187,
-0.10974842309951782,
-0.04938243329524994,
-0.085992231965065,
0.0033791849855333567,
0.04546360298991203,
0.02566615678369999,
-0.026717668399214745,
0.05953507497906685,
-0.0006225985125638545,
-0.07930180430412292,
-0.138752743601799,
0.06462296098470688,
-0.04645755514502525,
0.02116469107568264,
-0.03899853676557541,
-0.05425221100449562,
0.10300146043300629,
0.10435733944177628,
0.005352668464183807,
-0.0049666184931993484,
0.013156655244529247,
0.027002671733498573,
0.08847847580909729,
0.09142030030488968,
0.05687098577618599,
-0.03145742416381836,
-0.027454521507024765
] | 0 | 0 | 0 | 0 |
|
SV-RAG: LoRA-Contextualizing Adaptation of MLLMs for Long Document Understanding | https://openreview.net/forum?id=FDaHjwInXO | [
"Jian Chen",
"Ruiyi Zhang",
"Yufan Zhou",
"Tong Yu",
"Franck Dernoncourt",
"Jiuxiang Gu",
"Ryan A. Rossi",
"Changyou Chen",
"Tong Sun"
] | Poster | Multimodal large language models (MLLMs) have recently shown great progress in text-rich image understanding, yet they still struggle with complex, multi-page visually-rich documents. Traditional methods using document parsers for retrieval-augmented generation suffer from performance and efficiency limitations, while directly presenting all pages to MLLMs leads to inefficiencies, especially with lengthy ones. In this work, we present a novel framework named **S**elf-**V**isual **R**etrieval-**A**ugmented **G**eneration (SV-RAG), which can broaden horizons of *any* MLLM to support long-document understanding. We demonstrate that **MLLMs themselves can be an effective multimodal retriever** to fetch relevant pages and then answer user questions based on these pages. SV-RAG is implemented with two specific MLLM adapters, one for evidence page retrieval and the other for question answering. Empirical results show state-of-the-art performance on public benchmarks, demonstrating the effectiveness of SV-RAG. | Large Multimodal Models | SV-RAG enhances long-document understanding by adapting MLLMs for self-visual retrieval-augmented generation, optimizing both evidence retrieval and question answering with specialized LoRA adapters. | 12,071 | null | [
-0.07854734361171722,
-0.07497343420982361,
-0.028721584007143974,
-0.018774595111608505,
0.12183745205402374,
-0.03830011561512947,
-0.0871594250202179,
0.015477217733860016,
0.013701488263905048,
-0.008119914680719376,
-0.020337700843811035,
-0.013595037162303925,
0.041860952973365784,
0.01849173754453659,
-0.008831818588078022,
0.010598350316286087,
0.040973324328660965,
0.08477921038866043,
-0.12438327074050903,
0.017378779128193855,
-0.04792400822043419,
0.007088838145136833,
0.10812336951494217,
-0.06412824988365173,
0.038275912404060364,
0.03451259061694145,
-0.04565439000725746,
-0.091863252222538,
0.0370611809194088,
-0.10125632584095001,
0.07508071511983871,
0.030013466253876686,
-0.014765167608857155,
0.062198199331760406,
-0.02387416362762451,
0.07533527165651321,
-0.0395958386361599,
0.03226770833134651,
0.011804758571088314,
-0.052330613136291504,
-0.05799025669693947,
-0.01948866806924343,
-0.0005179467261768878,
-0.011367988772690296,
0.09460392594337463,
0.022979889065027237,
-0.05439549684524536,
0.0028723611030727625,
0.06198298931121826,
0.010235334746539593,
-0.10792148113250732,
-0.044225044548511505,
0.02066316269338131,
0.023215748369693756,
-0.04473331943154335,
-0.04319067671895027,
0.0399080254137516,
-0.05307069420814514,
-0.058388225734233856,
-0.039057131856679916,
-0.048019200563430786,
-0.07327236235141754,
-0.04245207831263542,
0.012930857948958874,
-0.0347091406583786,
-0.04164497181773186,
0.07686809450387955,
-0.06472010910511017,
-0.01342094037681818,
-0.003726477734744549,
-0.07151292264461517,
0.006652954500168562,
0.006896149832755327,
0.09458953142166138,
0.004629368428140879,
0.053386982530355453,
0.05642690137028694,
-0.06197121739387512,
0.0534500889480114,
-0.07944850623607635,
0.09009519219398499,
0.016929227858781815,
0.07242662459611893,
0.009659787639975548,
0.09669992327690125,
-0.0074148098938167095,
0.0015417529502883554,
0.017368463799357414,
-0.03171424940228462,
0.052227847278118134,
0.00972344633191824,
-0.057893410325050354,
0.07048221677541733,
-0.04170743376016617,
0.0033221766352653503,
0.06139906868338585,
0.095815509557724,
-0.06001214310526848,
-0.0239495150744915,
0.09874353557825089,
0.009433665312826633,
0.06144579127430916,
0.018524643033742905,
-0.08921639621257782,
-0.022957755252718925,
-0.03725944086909294,
0.05220666155219078,
0.026203613728284836,
0.003939615562558174,
-0.0653187707066536,
0.08545002341270447,
-0.010951591655611992,
0.014359092339873314,
-0.017555154860019684,
0.00005845016858074814,
-0.07760145515203476,
0.0700899288058281,
-0.02706114575266838,
0.06384889781475067,
-0.004574902355670929,
-0.018247712403535843,
0.07095495611429214,
-0.013156693428754807,
0.009528720751404762,
0.008352454751729965,
0.010054005309939384,
-0.005569550208747387,
-7.1306722508303285e-34,
0.06533502042293549,
0.042588066309690475,
-0.0071240318939089775,
0.037511106580495834,
-0.03632405772805214,
0.007039553485810757,
-0.05454955995082855,
0.03262167051434517,
-0.07008620351552963,
-0.040279630571603775,
-0.03714422509074211,
0.028983917087316513,
-0.034086283296346664,
0.10247446596622467,
-0.015496693551540375,
0.011563058011233807,
-0.04041145369410515,
-0.015920545905828476,
-0.045007284730672836,
-0.006850442849099636,
-0.024966644123196602,
-0.06739062815904617,
0.04067379608750343,
0.028594328090548515,
0.04595944657921791,
0.017369912937283516,
0.08310503512620926,
-0.11914306879043579,
-0.03172111511230469,
-0.0026542830746620893,
-0.07135060429573059,
-0.05492379888892174,
0.00588136026635766,
0.002655950840562582,
0.01793663389980793,
-0.05224737524986267,
-0.024781515821814537,
-0.04961191490292549,
-0.030506419017910957,
-0.010396427474915981,
-0.10338177531957626,
-0.04320218414068222,
-0.007290662731975317,
-0.0563712976872921,
-0.09232266992330551,
-0.0007380351889878511,
-0.00735542131587863,
0.031059959903359413,
-0.023590339347720146,
-0.025764647871255875,
0.06296084821224213,
0.01617552526295185,
-0.04908439889550209,
-0.026237335056066513,
0.019379323348402977,
-0.0014228252694010735,
0.041187822818756104,
0.05709008499979973,
0.05447358265519142,
0.010472822934389114,
0.010436874814331532,
0.09181718528270721,
0.05784744396805763,
0.07320918887853622,
0.011950393207371235,
-0.032819848507642746,
-0.012084761634469032,
0.09014251083135605,
0.00003201282743248157,
0.0037187286652624607,
-0.01088256947696209,
-0.0014117518439888954,
0.07120633125305176,
-0.06682170182466507,
0.008543136529624462,
0.02107272855937481,
-0.029353098943829536,
-0.06044469028711319,
0.08233577758073807,
0.007166814059019089,
0.0033305720426142216,
-0.004090792499482632,
0.03975355997681618,
-0.09219831228256226,
-0.030950365588068962,
-0.02772560529410839,
0.050110917538404465,
-0.1154242679476738,
-0.019465573132038116,
-0.06612005829811096,
0.028005177155137062,
-0.007456535939127207,
-0.004476774949580431,
-0.07417292147874832,
0.028635350987315178,
-1.8064300381609816e-35,
0.059391144663095474,
-0.07431373745203018,
-0.04277970269322395,
0.05998045206069946,
-0.010637066327035427,
0.026583775877952576,
-0.05011587217450142,
0.0670304223895073,
-0.06789851188659668,
-0.025490136817097664,
0.014435037039220333,
-0.000797289190813899,
-0.010999811813235283,
0.003832796821370721,
0.08614623546600342,
-0.043622180819511414,
0.04369020462036133,
-0.019671807065606117,
-0.07644936442375183,
0.08780732750892639,
-0.0049237036146223545,
0.05090363696217537,
0.012921025976538658,
0.0029201104771345854,
-0.02363171987235546,
0.056193288415670395,
-0.005438826512545347,
0.02102510631084442,
-0.10516209900379181,
-0.03976648673415184,
-0.03245993331074715,
0.012291528284549713,
0.012335947714745998,
-0.04775078967213631,
-0.02640189416706562,
0.07919684052467346,
0.06414157897233963,
-0.027168413624167442,
-0.08520855009555817,
0.10194015502929688,
0.056853245943784714,
0.0010621841065585613,
-0.021462664008140564,
-0.04993782192468643,
-0.0657411515712738,
-0.04715077579021454,
-0.11374539881944656,
-0.007866223342716694,
0.02190840244293213,
0.04584140703082085,
-0.0012906830525025725,
-0.035995565354824066,
-0.061724334955215454,
-0.03110663965344429,
-0.009055105969309807,
-0.0686790943145752,
-0.10058163106441498,
-0.030164921656250954,
-0.08997400850057602,
0.004237060435116291,
-0.0811322033405304,
-0.038357146084308624,
-0.06049201637506485,
0.05190562084317207,
0.03937876969575882,
-0.04295995458960533,
-0.03346638008952141,
-0.11461997032165527,
0.006509472616016865,
0.0178156029433012,
0.0009711896418593824,
-0.09241870790719986,
0.013203499838709831,
-0.013566273264586926,
0.07026566565036774,
0.081199511885643,
0.03709696978330612,
-0.035566627979278564,
0.05423421412706375,
-0.09210245311260223,
-0.007618722040206194,
-0.05707418546080589,
0.05070725083351135,
0.08924353867769241,
0.07564227283000946,
0.029850823804736137,
-0.043478842824697495,
0.07727042585611343,
0.0648093894124031,
0.011346071027219296,
0.008125909604132175,
0.01019662618637085,
0.030856583267450333,
0.10216210037469864,
0.06880433112382889,
-3.884569110823577e-8,
-0.057805050164461136,
0.039872005581855774,
-0.05562880262732506,
0.0231252983212471,
0.03089335560798645,
-0.024261001497507095,
-0.02824392355978489,
0.05657808482646942,
-0.04640736058354378,
0.02266397327184677,
0.048564210534095764,
-0.019541963934898376,
-0.0178395826369524,
-0.03577557951211929,
0.030329028144478798,
0.06974877417087555,
0.04004557058215141,
0.01825876347720623,
-0.01954825036227703,
-0.006719992961734533,
0.07051245868206024,
0.011448848992586136,
0.003614746266975999,
0.04110328480601311,
0.00042237876914441586,
0.01662076637148857,
-0.03334237262606621,
0.0893135666847229,
0.08634079247713089,
-0.09752633422613144,
0.008868026547133923,
0.0026567126624286175,
-0.04675546661019325,
-0.043980926275253296,
0.04245741665363312,
0.017043307423591614,
0.018267212435603142,
0.0011578116100281477,
-0.029559796676039696,
0.043883029371500015,
0.033002350479364395,
0.0644536092877388,
-0.05227919667959213,
0.03152408450841904,
0.030424436554312706,
0.021072251722216606,
-0.04425571486353874,
-0.017378663644194603,
-0.007060552015900612,
-0.04027850180864334,
-0.046690601855516434,
-0.11530101299285889,
0.08033161610364914,
0.11700642108917236,
0.032942887395620346,
-0.0280644278973341,
0.01468922384083271,
0.09940990805625916,
0.08816925436258316,
0.08183127641677856,
0.06531395018100739,
0.010107758454978466,
-0.024743320420384407,
0.040574412792921066
] | 0 | 0 | 0 | 0 |
|
Intelligence at the Edge of Chaos | https://openreview.net/forum?id=IeRcpsdY7P | [
"Shiyang Zhang",
"Aakash Patel",
"Syed A Rizvi",
"Nianchen Liu",
"Sizhuang He",
"Amin Karbasi",
"Emanuele Zappala",
"David van Dijk"
] | Poster | We explore the emergence of intelligent behavior in artificial systems by investigating how the complexity of rule-based systems influences the capabilities of models trained to predict these rules. Our study focuses on elementary cellular automata (ECA), simple yet powerful one-dimensional systems that generate behaviors ranging from trivial to highly complex. By training distinct Large Language Models (LLMs) on different ECAs, we evaluated the relationship between the complexity of the rules' behavior and the intelligence exhibited by the LLMs, as reflected in their performance on downstream tasks. Our findings reveal that rules with higher complexity lead to models exhibiting greater intelligence, as demonstrated by their performance on reasoning and chess move prediction tasks. Both uniform and periodic systems, and often also highly chaotic systems, resulted in poorer downstream performance, highlighting a sweet spot of complexity conducive to intelligence. We conjecture that intelligence arises from the ability to predict complexity and that creating intelligence may require only exposure to complexity. | Large Language Models (LLMs), Elementary Cellular Automata (ECA), Emergent Intelligence, Complex Systems, Complexity Theory | Training Large Language Models (LLMs) on elementary cellular automata (ECA) rules shows that models exposed to more complex rule behaviors perform better on downstream tasks. | 12,068 | 2410.02536 | [
-0.011209700256586075,
-0.06072380766272545,
0.06942246109247208,
-0.07370096445083618,
0.004485568962991238,
-0.061350833624601364,
-0.041988808661699295,
-0.038434870541095734,
0.05053100734949112,
0.03865133598446846,
-0.09864307194948196,
-0.06165122985839844,
0.04980696365237236,
0.015795599669218063,
-0.07525266706943512,
-0.05206697806715965,
0.018100755289196968,
-0.05809815973043442,
-0.0880245715379715,
-0.09808933734893799,
0.06007148325443268,
-0.04514544829726219,
-0.018863707780838013,
-0.052671223878860474,
0.009177224710583687,
0.07365437597036362,
-0.10194618999958038,
-0.03340446203947067,
0.0006191280554048717,
-0.04402109608054161,
0.050218332558870316,
-0.059627555310726166,
0.07645657658576965,
0.0014761375496163964,
-0.024036327376961708,
0.04621167853474617,
-0.012637571431696415,
0.027079617604613304,
0.07317915558815002,
-0.01251776423305273,
0.04515248164534569,
-0.035676583647727966,
0.062144938856363297,
0.06913059949874878,
0.05366479232907295,
0.030218273401260376,
-0.023044319823384285,
-0.09223029017448425,
-0.07143957912921906,
-0.03855874761939049,
-0.146925151348114,
-0.069916732609272,
0.010346837341785431,
0.042386848479509354,
0.03160033002495766,
0.040589798241853714,
0.01067203376442194,
0.11014071106910706,
0.01979721523821354,
-0.048094045370817184,
0.006582282483577728,
-0.10316857695579529,
0.003516391385346651,
-0.011780455708503723,
-0.049764156341552734,
0.007500224746763706,
-0.016611699014902115,
0.034417975693941116,
-0.08380448818206787,
-0.01789303496479988,
0.023522410541772842,
0.05658769980072975,
0.002588733332231641,
0.07778114080429077,
0.07632406055927277,
0.025129592046141624,
-0.07457538694143295,
-0.034831415861845016,
0.010797685943543911,
0.02138070948421955,
-0.07768985629081726,
-0.05873142182826996,
0.00035966854193247855,
0.020721662789583206,
0.04986991733312607,
-0.021473783999681473,
-0.0019889241084456444,
0.04262330383062363,
0.13767464458942413,
0.03240066021680832,
0.042898524552583694,
-0.05157141014933586,
0.025710538029670715,
-0.01821974851191044,
0.014723004773259163,
0.06299425661563873,
-0.00038730044616386294,
-0.04573787748813629,
0.012656105682253838,
0.01927211508154869,
0.006927489768713713,
0.017839323729276657,
-0.004643734078854322,
0.06051308289170265,
0.060099873691797256,
0.02889169380068779,
0.07104943692684174,
0.019174739718437195,
0.011839272454380989,
-0.0036114677786827087,
-0.05880739539861679,
0.06999717652797699,
0.060369133949279785,
0.09439098834991455,
0.007894331589341164,
0.02273694798350334,
-0.04942963644862175,
0.06256696581840515,
0.011580665595829487,
0.08005574345588684,
0.07025066018104553,
-0.04258730262517929,
-0.009824790060520172,
0.03030047006905079,
0.025894729420542717,
0.044509973376989365,
-0.06542620062828064,
-1.2335106908361292e-33,
-0.030004680156707764,
0.044105105102062225,
0.009365100413560867,
0.013915708288550377,
0.03740952908992767,
-0.05153936147689819,
0.006604174617677927,
-0.05692868307232857,
0.05560866370797157,
0.07665383815765381,
-0.0911998599767685,
0.07918313890695572,
-0.019440919160842896,
0.06903734803199768,
0.02046450786292553,
0.08758996427059174,
-0.010926729068160057,
-0.04879327118396759,
-0.057788506150245667,
-0.10343200713396072,
0.07585204392671585,
-0.026904471218585968,
0.03568210452795029,
-0.12885324656963348,
-0.0304822139441967,
0.00219453196041286,
0.03740294277667999,
-0.045775674283504486,
-0.0023923050612211227,
0.0012531706597656012,
-0.10699450969696045,
-0.01252895686775446,
-0.030717013403773308,
0.0846521183848381,
0.03881348669528961,
0.032272569835186005,
0.0064843688160181046,
-0.027095969766378403,
0.08316365629434586,
-0.01714208349585533,
-0.017172066494822502,
-0.021835805848240852,
0.0035277199931442738,
-0.0170237198472023,
0.017184928059577942,
-0.009649042040109634,
-0.002426804741844535,
-0.04053283482789993,
-0.0772261768579483,
-0.06299635022878647,
0.046063248068094254,
-0.008260216563940048,
0.031560663133859634,
-0.08688032627105713,
0.07202711701393127,
0.034514352679252625,
0.03235548734664917,
0.07239663600921631,
-0.028242776170372963,
0.1252482831478119,
-0.04513169080018997,
0.017345193773508072,
-0.0037265964783728123,
0.07160275429487228,
0.03714119270443916,
-0.004434544593095779,
-0.07087055593729019,
0.03930968418717384,
0.08539443463087082,
-0.022185441106557846,
0.07766567915678024,
-0.021168291568756104,
-0.03265242278575897,
0.0015882571460679173,
0.07126758992671967,
-0.08005518466234207,
0.10858552902936935,
-0.11683328449726105,
-0.10287489742040634,
-0.015782902017235756,
-0.04392634704709053,
-0.06265703588724136,
-0.03403059020638466,
0.018248477950692177,
-0.03324628993868828,
-0.07213136553764343,
0.007732411380857229,
-0.01733240857720375,
-0.05088610574603081,
0.030781880021095276,
-0.05658185854554176,
-0.09218629449605942,
0.058454010635614395,
0.00653644697740674,
-0.07796572148799896,
-2.1789284711078474e-33,
-0.06255199760198593,
-0.0653257742524147,
-0.04018675535917282,
0.037163570523262024,
0.01843375898897648,
0.007545118220150471,
-0.0522993765771389,
0.05209042504429817,
0.030881434679031372,
-0.02090558037161827,
-0.04902833700180054,
0.06384161114692688,
0.04381183162331581,
0.022073065862059593,
0.057191528379917145,
-0.08267957717180252,
0.05855649337172508,
0.004314010497182608,
0.06361434608697891,
0.015548020601272583,
-0.01601254940032959,
0.0365002267062664,
-0.09126611053943634,
0.034381505101919174,
-0.021149903535842896,
0.040142614394426346,
-0.08251676708459854,
0.029827285557985306,
-0.04248504340648651,
0.11520620435476303,
-0.030854325741529465,
-0.0684138759970665,
0.0277686920017004,
0.03863672539591789,
0.044683996587991714,
0.030803563073277473,
0.011507326737046242,
-0.029709376394748688,
-0.00680937385186553,
0.03114444389939308,
-0.021115226671099663,
-0.04144912213087082,
0.00783859845250845,
-0.031131388619542122,
0.012632024474442005,
0.08122462779283524,
-0.016641948372125626,
0.03944873437285423,
-0.026393650099635124,
0.022987380623817444,
0.046123068779706955,
0.03832630813121796,
-0.04892988130450249,
-0.06848384439945221,
-0.045294564217329025,
0.027209796011447906,
0.0011578514240682125,
0.03955419734120369,
0.03803905472159386,
-0.0035513455513864756,
-0.04329918697476387,
0.020603694021701813,
-0.0007076868787407875,
-0.04652085527777672,
-0.025101667270064354,
0.01861351542174816,
-0.08533041924238205,
-0.0029806471429765224,
0.051545921713113785,
0.0037948698736727238,
0.022254224866628647,
0.051981449127197266,
-0.08353358507156372,
-0.014643348753452301,
-0.012935774400830269,
0.028758110478520393,
-0.06611833721399307,
-0.02476431056857109,
-0.017349284142255783,
-0.018491869792342186,
0.027261950075626373,
0.08504738658666611,
0.007302707526832819,
-0.05027469992637634,
-0.0051755765452980995,
0.05029315873980522,
0.06436453759670258,
0.03474860265851021,
-0.00672678928822279,
-0.007819831371307373,
0.07354956120252609,
-0.02662961371243,
0.0010091773001477122,
0.04999031126499176,
-0.10042019933462143,
-4.174734513640033e-8,
0.01187075674533844,
0.018665367737412453,
0.049432214349508286,
0.0017224145121872425,
0.039373572915792465,
-0.00000977859235717915,
0.012815861962735653,
0.029566267505288124,
0.029550064355134964,
0.008549160324037075,
0.07879190146923065,
0.0005967789911665022,
0.02660922147333622,
0.06854241341352463,
0.07496337592601776,
0.047550879418849945,
-0.03769102320075035,
0.0010021902853623033,
-0.03822379559278488,
0.024847107008099556,
0.057495132088661194,
-0.026412982493638992,
-0.07525654137134552,
-0.017389846965670586,
-0.00767199881374836,
-0.11144261807203293,
-0.09928939491510391,
0.011185859329998493,
-0.05492815375328064,
0.07019300758838654,
0.03160744160413742,
0.005664341151714325,
0.03157281503081322,
0.005814948119223118,
-0.015088915824890137,
0.02141820453107357,
0.0005398074863478541,
-0.10123401135206223,
0.028376147150993347,
-0.07810455560684204,
0.028539417311549187,
0.08648382872343063,
-0.015561864711344242,
-0.005673397798091173,
0.01387912966310978,
-0.029755955561995506,
-0.028932977467775345,
-0.16985924541950226,
0.0002650658134371042,
-0.036090072244405746,
0.06162279471755028,
-0.007391368504613638,
-0.024952828884124756,
-0.06510403007268906,
0.07621759176254272,
-0.02445795387029648,
-0.018206287175416946,
-0.03889475390315056,
-0.008740141987800598,
0.0011540562845766544,
0.035726964473724365,
0.06856405735015869,
0.03818698599934578,
-0.07744970917701721
] | 0 | 0 | 0 | 0 |
|
Towards a learning theory of representation alignment | https://openreview.net/forum?id=DShqJA1Z64 | [
"Francesco Insulla",
"Shuo Huang",
"Lorenzo Rosasco"
] | Poster | It has recently been argued that AI models' representations are becoming aligned as their scale and performance increase. Empirical analyses have been designed to support this idea and conjecture the possible alignment of different representations toward a shared statistical model of reality. In this paper, we propose a learning-theoretic perspective to representation alignment. First, we review and connect different notions of alignment based on metric, probabilistic, and spectral ideas. Then, we focus on stitching, a particular approach to understanding the interplay between different representations in the context of a task. Our main contribution here is to relate the properties of stitching to the kernel alignment of the underlying representation. Our results can be seen as a first step toward casting representation alignment as a learning-theoretic problem. | learning theory, representation learning, model stitching, representation alignment | properties of stitching and connection to kernel alignment to better understand multimodal representation learning. | 12,067 | 2502.14047 | [
-0.0431431382894516,
-0.0980735570192337,
-0.04447491839528084,
-0.04625270143151283,
-0.00952743086963892,
0.052392296493053436,
0.006368102040141821,
-0.05128394439816475,
0.1107294112443924,
-0.02058246172964573,
-0.030033648014068604,
-0.05117949843406677,
0.022852309048175812,
0.05745730921626091,
-0.042872507125139236,
-0.03908870369195938,
0.06428199261426926,
0.058687712997198105,
-0.0985754057765007,
0.022315654903650284,
-0.019476864486932755,
-0.004479812458157539,
-0.03282476216554642,
0.031874943524599075,
0.047916412353515625,
0.06468554586172104,
0.08468751609325409,
-0.049989622086286545,
-0.008862457238137722,
-0.03813360258936882,
0.005889138672500849,
0.009087175130844116,
-0.012827004306018353,
0.02596820704638958,
-0.01988837495446205,
0.07689423859119415,
-0.08391715586185455,
0.03381378948688507,
0.005082784220576286,
-0.09300224483013153,
0.0389120876789093,
-0.006319511216133833,
-0.016600660979747772,
-0.0252994354814291,
0.05981418117880821,
0.08900182694196701,
-0.009132043458521366,
-0.0794709101319313,
-0.00967963133007288,
0.020421253517270088,
-0.18974661827087402,
0.031038235872983932,
-0.035257577896118164,
0.04427230730652809,
0.017847534269094467,
0.03490706533193588,
0.025865569710731506,
-0.00042655039578676224,
0.018113309517502785,
0.009424516931176186,
0.03409862145781517,
-0.05495944991707802,
0.0041655744425952435,
0.07224395871162415,
0.043492916971445084,
-0.01566675864160061,
0.024859441444277763,
0.009395533241331577,
-0.0862070620059967,
0.005548237822949886,
-0.09079107642173767,
0.10308192670345306,
-0.04150887951254845,
0.04623069614171982,
0.028706474229693413,
0.07426100224256516,
0.06099977344274521,
-0.02422957308590412,
0.055539652705192566,
-0.06508985906839371,
0.01357263047248125,
0.07449400424957275,
0.07017436623573303,
0.05968702584505081,
0.10893484950065613,
-0.009514616802334785,
0.02044498734176159,
-0.017020754516124725,
0.041454073041677475,
-0.05058499798178673,
-0.05589047074317932,
-0.06625227630138397,
0.0181793961673975,
0.010888028889894485,
0.012950832955539227,
0.0861789658665657,
0.06194652244448662,
-0.010705349035561085,
0.01745622418820858,
0.058332253247499466,
-0.06169985979795456,
0.03313573822379112,
0.006662872154265642,
-0.004741326905786991,
0.024007009342312813,
-0.04490501061081886,
-0.033537790179252625,
-0.047926366329193115,
0.029651539400219917,
-0.11710742115974426,
0.0437907911837101,
-0.10580851137638092,
-0.06694705039262772,
0.0810398980975151,
-0.029255729168653488,
-0.04181063175201416,
-0.013079793192446232,
0.023865599185228348,
-0.008771149441599846,
-0.011096864007413387,
-0.04098447412252426,
0.0066476622596383095,
-0.03407098725438118,
0.05580812320113182,
-0.015823090448975563,
-0.09059750288724899,
-0.043842535465955734,
2.30758816451527e-34,
0.017812907695770264,
0.04703456163406372,
0.07668761163949966,
0.03041263297200203,
0.06914176791906357,
-0.05256356671452522,
-0.07834967970848083,
0.018437596037983894,
0.10746447741985321,
0.032665930688381195,
-0.02078596130013466,
0.00925188697874546,
0.010412583127617836,
0.10537008196115494,
-0.048344727605581284,
0.03066694550216198,
-0.10266668349504471,
0.012771878391504288,
-0.09807871282100677,
-0.05513244494795799,
-0.02607765980064869,
0.028378507122397423,
-0.05426529422402382,
-0.014301914721727371,
-0.040881168097257614,
0.02444283850491047,
-0.013932224363088608,
-0.0596536286175251,
-0.08182010054588318,
-0.029035722836852074,
-0.05639266222715378,
-0.007912777364253998,
-0.019129468128085136,
-0.04743598401546478,
0.06331880390644073,
-0.05695657059550285,
0.028879385441541672,
-0.05124747380614281,
-0.044464077800512314,
0.010003352537751198,
0.03234726935625076,
0.08679714053869247,
-0.03343541920185089,
-0.049440108239650726,
-0.023160239681601524,
-0.02571261115372181,
0.004600184503942728,
-0.020900260657072067,
-0.04819507151842117,
-0.09397826343774796,
0.027274630963802338,
0.014329588040709496,
0.035408865660429,
-0.09141828119754791,
0.009772911667823792,
0.06646760553121567,
-0.02033422142267227,
0.01629435457289219,
0.015333510003983974,
0.005256199277937412,
-0.06031191721558571,
-0.04438571259379387,
0.03670114651322365,
0.03134413808584213,
0.02514309249818325,
0.07053488492965698,
0.06033124774694443,
0.07135280966758728,
0.12172207981348038,
0.030239906162023544,
0.0017940161051228642,
0.031030111014842987,
-0.0744592696428299,
-0.059749726206064224,
-0.03868835046887398,
0.02798963151872158,
0.0012084880145266652,
-0.07484287768602371,
-0.024353360757231712,
0.10565830022096634,
-0.07193666696548462,
-0.006704396102577448,
-0.028464684262871742,
-0.09910160303115845,
-0.09875934571027756,
-0.09470444917678833,
-0.00499472813680768,
-0.08972517400979996,
0.008881011977791786,
0.05891653895378113,
-0.04776832461357117,
-0.009718242101371288,
-0.018041228875517845,
0.010111596435308456,
-0.02180563658475876,
-2.7240972207512924e-33,
-0.0005200866144150496,
0.015231052413582802,
0.013651976361870766,
-0.012242453172802925,
-0.020241202786564827,
-0.05869973078370094,
0.00582939013838768,
0.010723324492573738,
-0.04120706766843796,
0.023621225729584694,
-0.052493803203105927,
-0.030376233160495758,
-0.013119921088218689,
-0.022821538150310516,
-0.020618418231606483,
0.04448648914694786,
-0.013452072627842426,
0.07225769758224487,
0.010704593732953072,
0.005164214409887791,
0.023612622171640396,
0.05623393505811691,
-0.05653538554906845,
0.001608686288818717,
0.09407693147659302,
0.056529659777879715,
0.07153981924057007,
0.0504034087061882,
0.01667446829378605,
0.036520641297101974,
-0.0013922828948125243,
-0.09861502796411514,
-0.06169906631112099,
0.024860963225364685,
-0.1061878427863121,
0.0708438828587532,
-0.019844401627779007,
-0.0022235121577978134,
0.015879664570093155,
0.038270920515060425,
0.022676337510347366,
-0.033418599516153336,
-0.08909585326910019,
-0.03676849976181984,
-0.03750527277588844,
-0.04351617395877838,
0.04093513637781143,
0.02123212069272995,
-0.05534199997782707,
-0.056952640414237976,
-0.004991884343326092,
0.13185323774814606,
-0.09981440007686615,
-0.002387287560850382,
-0.1112416461110115,
0.024779843166470528,
-0.0445011630654335,
-0.014483723789453506,
0.039478376507759094,
0.012417786754667759,
0.025782408192753792,
-0.038127344101667404,
0.07298262417316437,
-0.009438605047762394,
-0.00512061920017004,
0.03825466334819794,
0.008874659426510334,
-0.02952738292515278,
0.029675843194127083,
0.022243641316890717,
0.02497921697795391,
-0.08960912376642227,
-0.021565288305282593,
0.07336273044347763,
-0.025385122746229172,
-0.0594603456556797,
0.07988585531711578,
-0.006670802365988493,
0.011985689401626587,
-0.02013177052140236,
-0.0395188070833683,
-0.06424947082996368,
0.030358973890542984,
0.053723521530628204,
0.010386318899691105,
0.07476407289505005,
0.043482303619384766,
0.04549998417496681,
0.06795407086610794,
0.0030495326500386,
-0.002236052183434367,
0.07672621309757233,
-0.025418486446142197,
0.053578492254018784,
-0.0926676094532013,
-4.502077288748296e-8,
-0.05723036080598831,
0.010706840083003044,
0.08117342740297318,
0.022362858057022095,
0.020124316215515137,
0.023058617487549782,
-0.0010280804708600044,
0.03917267173528671,
-0.10436779260635376,
0.02618309296667576,
0.04882432147860527,
-0.023020697757601738,
-0.01709437556564808,
-0.058589790016412735,
0.032902080565690994,
0.017067305743694305,
-0.044280897825956345,
0.05152805894613266,
-0.008879057131707668,
0.09658359736204147,
0.06459075212478638,
0.0360894538462162,
-0.0272358525544405,
0.006895069032907486,
-0.012267103418707848,
-0.07984375953674316,
-0.09477153420448303,
0.057763926684856415,
-0.013573773205280304,
0.06641240417957306,
-0.03520116209983826,
0.09387558698654175,
0.08287285268306732,
-0.08438277244567871,
0.021264029666781425,
0.00021162467601243407,
0.025618689134716988,
-0.054968107491731644,
-0.022620834410190582,
-0.06260889768600464,
0.04616967588663101,
0.10787958651781082,
-0.014797916635870934,
0.04623939469456673,
0.06693883240222931,
0.041649457067251205,
0.03615299612283707,
-0.02692851983010769,
0.02868136763572693,
-0.001526149339042604,
0.03983714431524277,
0.017196981236338615,
0.0000074688350650831126,
0.046911828219890594,
0.034718096256256104,
-0.006256286986172199,
0.015391972847282887,
-0.04467711225152016,
0.0320885069668293,
0.05782422795891762,
0.036000918596982956,
0.034794844686985016,
-0.04174446687102318,
-0.0014868557918816805
] | 0 | 0 | 0 | 0 |
|
A Unified Framework for Forward and Inverse Problems in Subsurface Imaging using Latent Space Translations | https://openreview.net/forum?id=yIlyHJdYV3 | [
"Naveen Gupta",
"Medha Sawhney",
"Arka Daw",
"Youzuo Lin",
"Anuj Karpatne"
] | Poster | In subsurface imaging, learning the mapping from velocity maps to seismic waveforms (forward problem) and waveforms to velocity (inverse problem) is important for several applications. While traditional techniques for solving forward and inverse problems are computationally prohibitive, there is a growing interest to leverage recent advances in deep learning to learn the mapping between velocity maps and seismic waveform images directly from data. Despite the variety of architectures explored in previous works, several open questions still remain unanswered such as the effect of latent space sizes, the importance of manifold learning, the complexity of translation models, and the value of jointly solving forward and inverse problems. We propose a unified framework to systematically characterize prior research in this area termed the Generalized Forward-Inverse (GFI) framework, building on the assumption of manifolds and latent space translations. We show that GFI encompasses previous works in deep learning for subsurface imaging, which can be viewed as specific instantiations of GFI. We also propose two new model architectures within the framework of GFI: Latent U-Net and Invertible X-Net, leveraging the power of U-Nets for domain translation and the ability of IU-Nets to simultaneously learn forward and inverse translations, respectively. We show that our proposed models achieve state-of-the-art (SOTA) performance for forward and inverse problems on a wide range of synthetic datasets, and also investigate their zero-shot effectiveness on two real-world-like datasets. The code is available at https://github.com/KGML-lab/Generalized-Forward-Inverse-Framework-for-DL4SI | Machine Learning, Inverse Problems, Full-Waveform Inversion, Seismic Imaging, ML4Science, OpenFWI | Latent space Models for jointly solving forward and inverse problems for subsurface imaging | 12,054 | 2410.11247 | [
-0.06549190729856491,
-0.08412454277276993,
0.12746720016002655,
-0.011700463481247425,
0.03731447830796242,
-0.04554952681064606,
-0.06422481685876846,
-0.0607169084250927,
0.01087596919387579,
-0.0006156410090625286,
-0.00671839015558362,
-0.02402590401470661,
-0.031670182943344116,
0.04412209615111351,
-0.04659019038081169,
-0.005667643155902624,
0.009063060395419598,
0.04432135820388794,
-0.05885835736989975,
-0.01587766408920288,
-0.023822490125894547,
0.07992001622915268,
-0.014900267124176025,
-0.024705758318305016,
0.10577365010976791,
0.0037908905651420355,
0.059022173285484314,
0.010751352645456791,
-0.024839580059051514,
0.030981073155999184,
0.08448626846075058,
0.002949062269181013,
-0.05356019735336304,
-0.018600115552544594,
-0.040009044110774994,
0.1421840935945511,
-0.031878966838121414,
0.03779701516032219,
-0.022193653509020805,
0.03509194776415825,
-0.013530749827623367,
0.06443237513303757,
0.0019061286002397537,
0.03904133662581444,
0.0900338664650917,
-0.0007573777111247182,
0.005646166857331991,
-0.10517912358045578,
0.04633734002709389,
0.04243754595518112,
-0.0009198856423608959,
0.010209045372903347,
-0.10454947501420975,
0.024442065507173538,
0.04855504631996155,
0.034512970596551895,
0.005475588608533144,
-0.044617727398872375,
-0.0024919055867940187,
0.015375928953289986,
0.0416361466050148,
-0.04243108630180359,
-0.041180726140737534,
-0.04123343899846077,
0.04354498162865639,
-0.010431145317852497,
-0.0009944491321220994,
0.02588433027267456,
-0.07681326568126678,
0.057747092097997665,
-0.05965611711144447,
0.07646716386079788,
0.0062935203313827515,
-0.019248096272349358,
-0.002926494926214218,
-0.00038027652772143483,
0.040238793939352036,
0.049587324261665344,
0.038593094795942307,
-0.055751983076334,
0.1095653548836708,
0.04423031210899353,
0.037509430199861526,
0.025918645784258842,
0.05840011313557625,
0.039465151727199554,
-0.03285450488328934,
0.05016966909170151,
-0.034604281187057495,
-0.04451172798871994,
0.007768124341964722,
-0.12042701244354248,
-0.012140393257141113,
-0.0011089388281106949,
0.07292381674051285,
-0.07309452444314957,
0.007918167859315872,
-0.0829438641667366,
0.015218350104987621,
0.08189985901117325,
0.003134967526420951,
0.0147633608430624,
0.022917762398719788,
0.08816109597682953,
-0.0045937118120491505,
-0.03693418204784393,
0.029592718929052353,
0.08921854943037033,
0.02262607403099537,
-0.08850019425153732,
0.03409472852945328,
-0.026618972420692444,
0.045138806104660034,
-0.014972774311900139,
0.019029833376407623,
-0.11523614078760147,
0.01934671960771084,
-0.02916022203862667,
-0.004044955596327782,
-0.05470765009522438,
-0.08832016587257385,
-0.0014920743415132165,
-0.00909516867250204,
0.035067472606897354,
0.044537678360939026,
-0.09312541782855988,
-0.03396337479352951,
1.68020185022646e-33,
0.037031855434179306,
0.09677013754844666,
0.045220233500003815,
0.05095816031098366,
0.05016714334487915,
0.05298645421862602,
-0.03451300412416458,
0.028908491134643555,
0.014527739025652409,
0.06562203913927078,
-0.07397358119487762,
0.03659824654459953,
-0.04367177560925484,
0.0719493106007576,
-0.06805332750082016,
0.021474527195096016,
0.011033058166503906,
0.028784986585378647,
-0.046490442007780075,
-0.06161288917064667,
0.019017543643712997,
-0.014697671867907047,
-0.034712690860033035,
-0.035866878926754,
-0.03261757642030716,
-0.0015169823309406638,
-0.004864509683102369,
-0.0491129606962204,
-0.0543217808008194,
-0.034022945910692215,
-0.16787050664424896,
0.046355389058589935,
-0.01307947002351284,
0.030459875240921974,
-0.008558416739106178,
-0.020054390653967857,
0.0563030019402504,
-0.08005626499652863,
-0.0074199954979121685,
-0.010754476301372051,
0.0014136922545731068,
0.07091166079044342,
-0.01245004404336214,
-0.04938272759318352,
-0.0443643182516098,
-0.05526706576347351,
0.05416594818234444,
0.03892437741160393,
-0.0026497659273445606,
-0.004407062195241451,
0.04080471023917198,
-0.041263315826654434,
-0.07827922701835632,
-0.07682505249977112,
-0.04257986322045326,
0.0570765919983387,
0.012442312203347683,
-0.009546607732772827,
0.0642724558711052,
0.041225332766771317,
0.04528306797146797,
-0.03959265723824501,
0.00042668156675063074,
0.05285343900322914,
-0.040279485285282135,
0.061132002621889114,
0.03451085090637207,
-0.0011909729801118374,
0.03920397907495499,
0.04894560948014259,
-0.040483780205249786,
0.05789593607187271,
-0.06437929719686508,
-0.009728279896080494,
0.0741029605269432,
0.016563013195991516,
-0.024541497230529785,
-0.027707695960998535,
-0.009038319811224937,
0.03431439399719238,
-0.1098034605383873,
0.02396247163414955,
-0.06253275275230408,
-0.0010562869720160961,
-0.04611121490597725,
0.05369199812412262,
0.0780864730477333,
-0.09174348413944244,
0.040708571672439575,
-0.0225206445902586,
-0.11716138571500778,
-0.04958178475499153,
-0.011835851706564426,
-0.009098962880671024,
0.023195479065179825,
-1.5816612608229817e-33,
-0.017907584086060524,
0.05104070529341698,
-0.10193558037281036,
-0.026762837544083595,
0.0284735057502985,
-0.038409966975450516,
0.10577090829610825,
0.05818072706460953,
-0.0008343079243786633,
-0.060130082070827484,
0.011824729852378368,
-0.024922313168644905,
0.04828643426299095,
-0.01314459927380085,
-0.02256365492939949,
0.030852263793349266,
-0.005287828855216503,
-0.06894072890281677,
-0.039009686559438705,
-0.06866975873708725,
-0.027666917070746422,
0.0705738216638565,
0.023662220686674118,
0.04065064340829849,
-0.028105460107326508,
0.04617729410529137,
0.0004633628122974187,
0.028495125472545624,
-0.05448294058442116,
0.016354797407984734,
-0.050773896276950836,
-0.026228005066514015,
0.0032540042884647846,
-0.022828668355941772,
-0.06258545070886612,
0.1733296662569046,
0.04322534799575806,
-0.024856166914105415,
-0.05105414241552353,
-0.019576814025640488,
0.012644206173717976,
-0.003180336207151413,
0.024342169985175133,
0.033553604036569595,
-0.04554224759340286,
-0.009399555623531342,
0.018378855660557747,
-0.003704774659126997,
0.0344093032181263,
0.06799141317605972,
-0.045349277555942535,
-0.036783020943403244,
-0.020110636949539185,
-0.01082120556384325,
-0.0792013630270958,
0.012621451169252396,
0.022470960393548012,
-0.021595412865281105,
0.08525166660547256,
0.03784065693616867,
-0.04551718011498451,
-0.0676870197057724,
0.010079216212034225,
0.009065556339919567,
-0.018497053533792496,
0.023000875487923622,
-0.04167066514492035,
0.007445730734616518,
0.032829400151968,
0.08713458478450775,
-0.0035180416889488697,
-0.020048560574650764,
0.023726169019937515,
0.017865071073174477,
0.058970287442207336,
-0.07648036628961563,
0.025307053700089455,
0.05570311099290848,
-0.005124583840370178,
-0.0619637705385685,
0.12595629692077637,
-0.03775806352496147,
0.027348747476935387,
0.12477541714906693,
-0.0031924417708069086,
0.025974225252866745,
0.0380622074007988,
-0.04126649722456932,
0.019609997048974037,
-0.009557897225022316,
0.014720327220857143,
-0.01446455530822277,
-0.053862232714891434,
-0.02426762320101261,
-0.017473263666033745,
-3.677035564919606e-8,
-0.10874617099761963,
0.043282851576805115,
0.07788387686014175,
-0.11117898672819138,
-0.02943054586648941,
-0.048940639942884445,
0.07165426760911942,
0.05712675675749779,
0.0032860231585800648,
-0.04211044684052467,
0.034164197742938995,
-0.09101838618516922,
-0.05891268327832222,
-0.10769390314817429,
-0.09330885857343674,
0.0655861347913742,
0.04080880805850029,
0.009752350859344006,
-0.025259243324398994,
0.014300239272415638,
-0.013002912513911724,
0.07227134704589844,
0.00497244531288743,
-0.018796928226947784,
0.1104983314871788,
-0.05018412321805954,
-0.035675108432769775,
-0.004222827963531017,
0.06231828033924103,
-0.02807314321398735,
-0.030938776209950447,
0.03160080313682556,
0.024955550208687782,
0.04116801917552948,
-0.0431801937520504,
0.03720395267009735,
0.056098394095897675,
-0.013760041445493698,
-0.06672023981809616,
-0.02428678423166275,
-0.08313167095184326,
0.1076507717370987,
-0.002555983141064644,
-0.0177337434142828,
-0.01201185304671526,
0.09492742270231247,
0.03810373321175575,
-0.002380172023549676,
0.02468489110469818,
0.02010650746524334,
0.09736847877502441,
-0.011071800254285336,
0.05908094719052315,
0.08813004940748215,
0.11147714406251907,
0.053555652499198914,
0.016952209174633026,
-0.05664459615945816,
0.03530389070510864,
0.03901312127709389,
-0.0811525359749794,
0.048629216849803925,
0.00034799313289113343,
-0.04497025907039642
] | https://github.com/KGML-lab/Generalized-Forward-Inverse-Framework-for-DL4SI | 2 | 0 | 0 | 0 |
Cauchy-Schwarz Regularizers | https://openreview.net/forum?id=KZu3xhPhke | [
"Sueda Taner",
"Ziyi Wang",
"Christoph Studer"
] | Poster | We introduce a novel class of regularization functions, called Cauchy–Schwarz (CS) regularizers, which can be designed to induce a wide range of properties in solution vectors of optimization problems. To demonstrate the versatility of CS regularizers, we derive regularization functions that promote discrete-valued vectors, eigenvectors of a given matrix, and orthogonal matrices. The resulting CS regularizers are simple, differentiable, and can be free of spurious stationary points, making them suitable for gradient-based solvers and large-scale optimization problems. In addition, CS regularizers automatically adapt to the appropriate scale, which is, for example, beneficial when discretizing the weights of neural networks. To demonstrate the efficacy of CS regularizers, we provide results for solving underdetermined systems of linear equations and weight quantization in neural networks. Furthermore, we discuss specializations, variations, and generalizations, which lead to an even broader class of new and possibly more powerful regularizers. | Regularizer, optimization, discretization, quantization | We propose a new family of regularization functions that can be used to discretize vectors. | 12,042 | null | [
-0.04287702590227127,
-0.0178766380995512,
0.017282959073781967,
-0.017706915736198425,
0.0067036631517112255,
0.094181127846241,
0.014314730651676655,
-0.11566857993602753,
-0.04053070396184921,
-0.050442442297935486,
0.001638941583223641,
0.030741583555936813,
-0.051756344735622406,
0.0510651133954525,
-0.05363387614488602,
0.018094506114721298,
0.0029913003090769053,
0.07928644120693207,
-0.18468810617923737,
0.046247780323028564,
0.015358427539467812,
0.002663244493305683,
-0.0496550016105175,
-0.019999412819743156,
0.018077341839671135,
-0.0043725622817873955,
0.04938191547989845,
0.023727206513285637,
0.014567206613719463,
0.02233954705297947,
0.01457220409065485,
-0.07874191552400589,
-0.0034156772308051586,
-0.02583678998053074,
-0.08407055586576462,
0.07057895511388779,
0.026054907590150833,
0.07345225661993027,
-0.06156789883971214,
-0.005457404535263777,
0.06571684777736664,
0.022893665358424187,
-0.02179277315735817,
0.06533464789390564,
0.09715066850185394,
-0.019143011420965195,
0.06615934520959854,
-0.052880533039569855,
0.010202845558524132,
-0.050097882747650146,
-0.008565832860767841,
-0.0263373963534832,
-0.04248131066560745,
0.04373829811811447,
0.03787029907107353,
-0.03562534600496292,
0.03292230889201164,
0.03238920494914055,
-0.0032641044817864895,
-0.014370997436344624,
-0.0195612832903862,
-0.10735699534416199,
0.01236940361559391,
-0.024822711944580078,
0.03531265631318092,
-0.02379644475877285,
0.02909550815820694,
0.03148515522480011,
0.04012730345129967,
0.07200781255960464,
-0.037553224712610245,
0.04430808126926422,
-0.04269573092460632,
0.08395365625619888,
0.0537906140089035,
0.003602063050493598,
0.05660117045044899,
-0.026738207787275314,
0.048562973737716675,
0.02969091758131981,
-0.04009722173213959,
-0.003044238081201911,
-0.011738204397261143,
0.013136289082467556,
0.11701370775699615,
0.04431662708520889,
-0.061315540224313736,
-0.023158542811870575,
0.08189766108989716,
0.005857197102159262,
-0.047094836831092834,
0.03457598388195038,
-0.025247633457183838,
-0.039277203381061554,
-0.04919828847050667,
-0.042694225907325745,
0.024768440052866936,
-0.08073533326387405,
0.031435854732990265,
0.10066159069538116,
0.06212349981069565,
-0.019998013973236084,
0.07554025948047638,
0.05909905955195427,
0.030188729986548424,
0.02126852236688137,
0.007060073781758547,
0.06663291156291962,
0.02928158827126026,
-0.10684365779161453,
0.010804328136146069,
0.034580741077661514,
-0.018574241548776627,
0.036639101803302765,
0.030314834788441658,
-0.04644237458705902,
0.0771900862455368,
-0.03686823695898056,
0.06079353019595146,
0.0035124749410897493,
-0.09283600747585297,
-0.0062508187256753445,
-0.03896337375044823,
0.0032719422597438097,
0.02524082362651825,
0.04076828435063362,
-0.12201070785522461,
2.068585734751121e-33,
0.0054319268092513084,
0.044160764664411545,
0.008506943471729755,
-0.031017176806926727,
0.008574449457228184,
0.006599122658371925,
0.024611826986074448,
0.05215676873922348,
0.003815185045823455,
-0.0024564634077250957,
0.008021722547709942,
0.056360479444265366,
-0.08834892511367798,
0.1121533140540123,
0.059723228216171265,
-0.06816209852695465,
-0.0063078682869672775,
0.0018314458429813385,
0.0829038918018341,
-0.1809622347354889,
0.05634709820151329,
-0.08629479259252548,
-0.027398407459259033,
0.04486098140478134,
-0.03937526047229767,
-0.09968967735767365,
0.12297806888818741,
-0.000823855574708432,
-0.047578390687704086,
-0.03597617149353027,
-0.03935903683304787,
-0.04722044616937637,
0.04252002015709877,
0.0034884584601968527,
0.04074002429842949,
0.03339717537164688,
0.09281911700963974,
0.006930147297680378,
0.06910328567028046,
0.012453295290470123,
-0.013557713478803635,
-0.021029755473136902,
0.017294038087129593,
0.0035279234871268272,
-0.014503128826618195,
-0.0349765308201313,
-0.01807175576686859,
0.02366790547966957,
-0.010551503859460354,
-0.12062442302703857,
-0.011996865272521973,
-0.004260310437530279,
-0.00047340005403384566,
-0.02834230661392212,
0.01620667241513729,
-0.018110061064362526,
0.03170565888285637,
-0.023453345522284508,
0.01830967888236046,
0.0029039066284894943,
0.013352633453905582,
0.008617771789431572,
0.04324500635266304,
0.003914571832865477,
-0.026229098439216614,
0.003923050593584776,
0.006685731932520866,
0.034617193043231964,
0.012142224237322807,
0.012189783155918121,
-0.007917860522866249,
0.018140777945518494,
0.06681277602910995,
-0.05576186627149582,
0.12020237743854523,
-0.0013029813999310136,
0.04164191707968712,
-0.03063460811972618,
-0.030808735638856888,
0.022219186648726463,
-0.007818099111318588,
0.09241359680891037,
-0.013871791772544384,
-0.018490949645638466,
-0.10733658075332642,
-0.07961360365152359,
0.03620061278343201,
-0.11773236840963364,
-0.03618361055850983,
-0.055768825113773346,
-0.08804568648338318,
0.0559266097843647,
-0.00896916352212429,
0.018506474792957306,
-0.09796246886253357,
-2.553829619099569e-33,
-0.04698887839913368,
0.03231746330857277,
-0.021930072456598282,
0.008718911558389664,
0.008440418168902397,
-0.0550394132733345,
-0.055895477533340454,
0.015560093335807323,
-0.03898210451006889,
-0.08829336613416672,
0.013479859568178654,
-0.04009294509887695,
0.05897503346204758,
-0.01976342871785164,
0.0053835040889680386,
0.08363717794418335,
-0.08345907181501389,
0.02462020143866539,
-0.026977552101016045,
-0.004472896922379732,
0.004191415384411812,
0.09017682820558548,
-0.003683174028992653,
0.06332580000162125,
-0.012135326862335205,
-0.031907565891742706,
-0.036270931363105774,
0.0837285965681076,
0.01727651059627533,
0.014859672635793686,
-0.04166480526328087,
-0.06161634996533394,
-0.08251634985208511,
-0.018389783799648285,
-0.02862747758626938,
0.03846145421266556,
-0.04580583795905113,
0.002087742555886507,
-0.006930781062692404,
0.09178651124238968,
-0.013691017404198647,
-0.07030028849840164,
-0.06030470132827759,
-0.010454211384057999,
0.04712013527750969,
-0.0053106676787137985,
0.05141708627343178,
-0.023362038657069206,
-0.04885239154100418,
0.03556881472468376,
0.011318172328174114,
0.0018475037068128586,
-0.03992348164319992,
0.07364645600318909,
-0.10092310607433319,
-0.008520415052771568,
0.04784329980611801,
0.007266078144311905,
0.0459672212600708,
0.0515693835914135,
-0.0751061961054802,
-0.08560915291309357,
0.006954073905944824,
-0.04170788452029228,
-0.05264464020729065,
0.0043997857719659805,
0.031910937279462814,
0.046639442443847656,
0.0551484115421772,
-0.013272671960294247,
0.018061520531773567,
-0.04663410037755966,
0.05364355817437172,
0.02863549254834652,
-0.13972052931785583,
-0.09874574840068817,
0.011591421440243721,
-0.010844191536307335,
-0.051290113478899,
0.07462770491838455,
0.05513913184404373,
0.007844316773116589,
0.11852025985717773,
0.02699832059442997,
0.025941554456949234,
0.07089439034461975,
0.0009119870373979211,
0.07760792225599289,
0.047019537538290024,
0.009811364114284515,
0.0017246276838704944,
-0.0015305785927921534,
0.020978780463337898,
-0.0029196173418313265,
-0.04068991169333458,
-4.194142277924584e-8,
-0.14187459647655487,
0.0474729910492897,
0.028488969430327415,
-0.01158857997506857,
0.009135041385889053,
-0.054903432726860046,
0.002353791380301118,
-0.0099767055362463,
0.008449533954262733,
-0.0555882453918457,
0.025188278406858444,
-0.013678678311407566,
-0.07718934863805771,
-0.042551375925540924,
-0.06998802721500397,
0.058375243097543716,
-0.025139804929494858,
-0.00570632703602314,
0.03085996024310589,
-0.0658235251903534,
-0.08444872498512268,
0.0553404875099659,
-0.019837751984596252,
-0.005978934001177549,
0.0317353792488575,
-0.0467195063829422,
-0.05773555487394333,
0.04943219944834709,
0.1218172237277031,
0.08704926073551178,
-0.020484259352087975,
0.06509383767843246,
0.047516848891973495,
0.04181777685880661,
0.007357822265475988,
0.10053325444459915,
0.0033626663498580456,
-0.021044693887233734,
-0.10389180481433868,
-0.0248132161796093,
0.0033685448579490185,
0.02116524428129196,
0.00814766064286232,
0.027778921648859978,
0.055150583386421204,
-0.005039369221776724,
0.0003578370378818363,
-0.011827033013105392,
0.05659441649913788,
-0.035798780620098114,
0.10190452635288239,
0.03160326927900314,
0.012364025227725506,
0.05574357509613037,
0.06897411495447159,
-0.02911580540239811,
-0.04259476065635681,
-0.06633390486240387,
-0.02884731814265251,
0.07099860906600952,
-0.014304501004517078,
0.045498158782720566,
-0.024733973667025566,
-0.0017887064022943377
] | 0 | 0 | 0 | 0 |
|
Adversarially Robust Anomaly Detection through Spurious Negative Pair Mitigation | https://openreview.net/forum?id=t8fu5m8R5m | [
"Hossein Mirzaei",
"Mojtaba Nafez",
"Jafar Habibi",
"Mohammad Sabokrou",
"Mohammad Hossein Rohban"
] | Poster | Despite significant progress in Anomaly Detection (AD), the robustness of existing detection methods against adversarial attacks remains a challenge, compromising their reliability in critical real-world applications such as autonomous driving. This issue primarily arises from the AD setup, which assumes that training data is limited to a group of unlabeled normal samples, making the detectors vulnerable to adversarial anomaly samples during testing. Additionally, implementing adversarial training as a safeguard encounters difficulties, such as formulating an effective objective function without access to labels. An ideal objective function for adversarial training in AD should promote strong perturbations both within and between the normal and anomaly groups to maximize margin between normal and anomaly distribution. To address these issues, we first propose crafting a pseudo-anomaly group derived from normal group samples. Then, we demonstrate that adversarial training with contrastive loss could serve as an ideal objective function, as it creates both inter- and intra-group perturbations. However, we notice that spurious negative pairs compromise the conventional contrastive loss for achieving robust AD. Spurious negative pairs are those that should be mapped closely but are erroneously separated. These pairs introduce noise and misguide the direction of inter-group adversarial perturbations. To overcome the effect of spurious negative pairs, we define opposite pairs and adversarially pull them apart to strengthen inter-group perturbations. Experimental results demonstrate our superior performance in both clean and adversarial scenarios, with a 26.1% improvement in robust detection across various challenging benchmark datasets. | Anomaly Detection, Adversarially Robust Anomaly Detection, Mitigating Spurious Negative Pairs, Anomaly Aware Contrastive Learning | null | 12,040 | null | [
-0.0720873698592186,
-0.04028183966875076,
-0.016562407836318016,
0.07987231761217117,
0.08934380859136581,
-0.002100186189636588,
0.0396522581577301,
-0.08663150668144226,
-0.026942351832985878,
-0.07294735312461853,
0.0039039915427565575,
-0.06902885437011719,
0.04724840074777603,
0.025000443682074547,
-0.10254339873790741,
-0.007947471924126148,
0.10739517211914062,
0.09226607531309128,
-0.07568737864494324,
0.03945303335785866,
-0.05601832643151283,
0.06212478131055832,
-0.028669383376836777,
0.009626329876482487,
-0.02857298217713833,
-0.0416586771607399,
0.00960626546293497,
-0.050815679132938385,
0.01572871394455433,
-0.03875764086842537,
0.03353092446923256,
-0.0397116020321846,
-0.04442382603883743,
-0.018916750326752663,
-0.006320595741271973,
-0.0091554569080472,
-0.01132280845195055,
0.017371123656630516,
0.060875725001096725,
-0.06083964556455612,
0.04344915226101875,
-0.03334781527519226,
-0.012207725085318089,
-0.028454307466745377,
-0.06470271199941635,
-0.0593511238694191,
-0.00013671157648786902,
-0.020598217844963074,
0.059374772012233734,
-0.028045516461133957,
0.04287794977426529,
-0.003764878259971738,
-0.02991778589785099,
0.044003747403621674,
-0.06455644965171814,
0.001740128151141107,
-0.023656710982322693,
0.02685714326798916,
-0.05320058390498161,
0.058466993272304535,
0.04100454971194267,
-0.04586470127105713,
-0.0012188683031126857,
-0.03700481355190277,
-0.006466915365308523,
-0.027241453528404236,
0.030902186408638954,
0.08175396919250488,
-0.0025915333535522223,
0.1338300257921219,
0.05217709392309189,
0.11044634878635406,
-0.03642011061310768,
-0.010773773305118084,
0.07037271559238434,
0.12158419191837311,
-0.036127008497714996,
-0.024851428344845772,
0.016199743375182152,
-0.07113482058048248,
0.0790296271443367,
0.04491083323955536,
0.03635573759675026,
0.07244518399238586,
0.13816110789775848,
0.02183377556502819,
-0.05596158280968666,
0.029439767822623253,
0.051509544253349304,
0.04793857783079147,
-0.10104000568389893,
-0.0121148731559515,
0.004956972785294056,
-0.005695183761417866,
0.1222178041934967,
0.0058243307285010815,
0.01127060316503048,
-0.041994981467723846,
0.0136801078915596,
0.0992475152015686,
-0.01932721585035324,
0.01717367023229599,
-0.03387925773859024,
-0.006597514729946852,
0.04685831815004349,
-0.019922306761145592,
0.032933980226516724,
-0.03706453740596771,
0.10595675557851791,
-0.04427044838666916,
0.05545567721128464,
-0.0036254124715924263,
0.049687184393405914,
-0.05212928354740143,
-0.011493729427456856,
-0.03214501589536667,
-0.022152284160256386,
0.07861999422311783,
-0.02776069939136505,
-0.00857024360448122,
-0.032158251851797104,
0.017871540039777756,
0.027337150648236275,
0.0005964612937532365,
0.0851319432258606,
0.0176331028342247,
-0.007244243752211332,
1.8603937466059932e-33,
0.03966934606432915,
0.03431316465139389,
0.014867709949612617,
-0.023892544209957123,
0.03014279343187809,
-0.05301472917199135,
-0.11691156029701233,
0.03955787792801857,
-0.022813349962234497,
0.06934653967618942,
0.009802475571632385,
-0.005132222548127174,
-0.03265015408396721,
0.01577182114124298,
0.009823773056268692,
0.10475463420152664,
0.04989536851644516,
-0.024962132796645164,
-0.03721366450190544,
0.010274190455675125,
0.007049437612295151,
-0.02774793654680252,
-0.05423484370112419,
-0.064795583486557,
-0.007388348691165447,
0.011441892944276333,
0.024068793281912804,
0.02086728811264038,
0.005487618502229452,
0.03082442283630371,
-0.11209108680486679,
0.06833815574645996,
-0.014431723393499851,
0.03646604344248772,
0.0062885344959795475,
-0.009853304363787174,
-0.05094967037439346,
-0.06733617931604385,
-0.0226972084492445,
0.013430670835077763,
0.00941083300858736,
0.04444875940680504,
-0.00019135925685986876,
-0.04698185250163078,
0.01941593363881111,
-0.011496740393340588,
0.07339014112949371,
0.03444701433181763,
-0.09153243154287338,
-0.07298774272203445,
-0.05262500420212746,
0.01957130618393421,
-0.00046443421160802245,
-0.023955756798386574,
-0.0899737998843193,
-0.05904018133878708,
0.008264916948974133,
0.0408739410340786,
-0.04483909159898758,
0.041505467146635056,
-0.024093642830848694,
-0.011534418910741806,
-0.01421351544559002,
0.03519011661410332,
-0.06852669268846512,
0.02569400705397129,
-0.008401788771152496,
0.006760722957551479,
0.02103452757000923,
0.051909931004047394,
0.0012275241315364838,
-0.008404523134231567,
-0.03179265931248665,
-0.03280368447303772,
0.04182101786136627,
0.02590278908610344,
0.11514781415462494,
0.006926130503416061,
0.006491129286587238,
-0.010082793422043324,
-0.024859696626663208,
-0.03413870185613632,
0.02605261467397213,
-0.03147020563483238,
-0.1360304057598114,
0.022073684260249138,
-0.012448407709598541,
-0.02024826779961586,
-0.026148177683353424,
0.023020794615149498,
-0.10363045334815979,
-0.014170187525451183,
-0.11757614463567734,
0.013791946694254875,
-0.013009118847548962,
-3.400649320947214e-33,
-0.04972051829099655,
0.05495332553982735,
-0.07846316695213318,
-0.021331828087568283,
-0.004512008745223284,
-0.02266472950577736,
0.03876878321170807,
0.09187936037778854,
-0.04725200682878494,
0.0010444188956171274,
-0.07972828298807144,
0.0321887731552124,
0.04170644283294678,
-0.008575653657317162,
0.02431674301624298,
-0.06616107374429703,
-0.03230363503098488,
0.0823589563369751,
0.005341201089322567,
0.0006424161838367581,
0.06924034655094147,
0.14017125964164734,
-0.04415171593427658,
0.07264430075883865,
0.039978768676519394,
0.00015856133541092277,
-0.07106123864650726,
0.08942627906799316,
-0.02998369187116623,
0.0013541006483137608,
0.008683961816132069,
0.0775270015001297,
0.004416486248373985,
-0.032221321016550064,
-0.006393632851541042,
0.09394586086273193,
0.061186037957668304,
-0.006849060766398907,
-0.06580746918916702,
0.05935141071677208,
-0.010677321814000607,
0.04396076500415802,
-0.012576167471706867,
0.05826658010482788,
0.062298405915498734,
-0.021772351115942,
0.059216104447841644,
-0.04253082722425461,
-0.0036580644082278013,
0.0045869178138673306,
-0.027042610570788383,
-0.0066264537163078785,
-0.048531580716371536,
0.0636242926120758,
-0.060891248285770416,
-0.03303716704249382,
0.005738873966038227,
-0.0171870868653059,
0.035275477916002274,
0.12938587367534637,
-0.046307239681482315,
0.01918475143611431,
-0.02894066646695137,
-0.01642858237028122,
-0.05608309432864189,
-0.045526690781116486,
0.004635009448975325,
-0.01606794074177742,
0.03891350328922272,
0.02342015691101551,
0.025353258475661278,
0.030703622847795486,
-0.03627992048859596,
0.06025135889649391,
-0.06504060328006744,
-0.0351376086473465,
-0.015287934802472591,
-0.06813888996839523,
-0.009959413670003414,
0.009878038428723812,
0.08977252244949341,
-0.003086835378780961,
0.05521826446056366,
0.07032216340303421,
0.027265368029475212,
0.08279772847890854,
0.07259092479944229,
-0.017346765846014023,
-0.02143404260277748,
0.01592741161584854,
-0.022489016875624657,
0.0234312005341053,
-0.10406705737113953,
-0.008161528035998344,
-0.024755332618951797,
-3.934514225534258e-8,
-0.09058695286512375,
0.061664655804634094,
0.013068074360489845,
-0.03722088038921356,
0.03531808406114578,
-0.04409538954496384,
-0.10253134369850159,
0.04964955523610115,
-0.041476331651210785,
-0.03056236542761326,
0.007851247675716877,
-0.10416032373905182,
-0.06196437031030655,
-0.09104005992412567,
-0.04990949109196663,
0.06042250990867615,
-0.009999771602451801,
0.010697590187191963,
-0.06374416500329971,
0.014818056486546993,
0.000587415648624301,
0.021970853209495544,
-0.013195142149925232,
-0.07302519679069519,
0.025225309655070305,
-0.1019645631313324,
0.03419067710638046,
-0.019706018269062042,
0.09217558801174164,
-0.023960577324032784,
-0.06494589149951935,
-0.022442921996116638,
0.044114213436841965,
0.06547601521015167,
-0.07591385394334793,
0.1422061324119568,
-0.06970635056495667,
-0.026954200118780136,
0.035250477492809296,
-0.04452362656593323,
-0.025555536150932312,
0.036306336522102356,
0.025171231478452682,
-0.011940373107790947,
-0.08989010006189346,
-0.013579399324953556,
-0.00951513834297657,
-0.09304942935705185,
0.025092247873544693,
-0.04948890581727028,
0.04798179864883423,
-0.061013951897621155,
0.02462216094136238,
0.1257631629705429,
0.04181007295846939,
-0.045792218297719955,
-0.024244580417871475,
-0.059761643409729004,
0.0006488038343377411,
0.08764227479696274,
0.046136219054460526,
0.03992883488535881,
-0.05515825375914574,
-0.062196820974349976
] | 0 | 0 | 0 | 0 |
|
Holographic Node Representations: Pre-training Task-Agnostic Node Embeddings | https://openreview.net/forum?id=tGYFikNONB | [
"Beatrice Bevilacqua",
"Joshua Robinson",
"Jure Leskovec",
"Bruno Ribeiro"
] | Poster | Large general purpose pre-trained models have revolutionized computer vision and natural language understanding. However, the development of general purpose pre-trained Graph Neural Networks (GNNs) lags behind other domains due to the lack of suitable generalist node representations. Existing GNN architectures are often tailored to specific task orders, such as node-level, link-level, or higher-order tasks, because different tasks require distinct permutation symmetries, which are difficult to reconcile within a single model. In this paper, we propose _holographic node representations_, a new blueprint for node representations capable of solving tasks of any order. Holographic node representations have two key components: (1) a task-agnostic expansion map, which produces highly expressive, high-dimensional embeddings, free from node-permutation symmetries, to be fed into (2) a reduction map that carefully reintroduces the relevant permutation symmetries to produce low-dimensional, task-specific embeddings. We show that well-constructed expansion maps enable simple and efficient reduction maps, which can be adapted for any task order. Empirical results show that holographic node representations can be effectively pre-trained and reused across tasks of varying orders, yielding up to 100% relative performance improvement, including in cases where prior methods fail entirely. | GNN, symmetries, pretraining | We propose holographic node representations, a new blueprint for node representations capable of solving graph tasks of any order. | 12,039 | null | [
0.0025753017980605364,
-0.06322864443063736,
0.1359260231256485,
-0.025951309129595757,
0.05976758897304535,
0.04081171378493309,
-0.044898226857185364,
-0.0832488089799881,
-0.015221645124256611,
-0.009290902875363827,
-0.04466791823506355,
0.015011460520327091,
-0.028261376544833183,
0.09186732769012451,
-0.026774311438202858,
0.02516268752515316,
0.025382375344634056,
0.07200873643159866,
-0.049051038920879364,
-0.01256762258708477,
0.020619405433535576,
-0.03669453039765358,
-0.028163431212306023,
0.014365233480930328,
0.06202990561723709,
-0.03038599155843258,
-0.019956273958086967,
-0.028956448659300804,
0.03644814342260361,
0.020752890035510063,
0.0616496317088604,
-0.016344822943210602,
-0.001567948143929243,
0.022895336151123047,
-0.10297458618879318,
0.10172589123249054,
-0.022828424349427223,
-0.023655571043491364,
-0.015552639961242676,
-0.0752045139670372,
0.07468090206384659,
0.06486918777227402,
0.010639293119311333,
0.094252809882164,
0.08305016905069351,
0.008845985867083073,
-0.020523013547062874,
-0.0710405483841896,
-0.0639769583940506,
-0.004785132594406605,
-0.05741318315267563,
-0.025222672149538994,
-0.05290033668279648,
-0.02345055341720581,
-0.0084199458360672,
0.05395732447504997,
0.0037079204339534044,
-0.0697687640786171,
-0.01627449505031109,
-0.004009781870990992,
0.002386852167546749,
-0.08702366799116135,
0.008758381940424442,
-0.08790521323680878,
0.01961665041744709,
0.009673168882727623,
0.06016823276877403,
0.02033322863280773,
0.02134573645889759,
0.07403039932250977,
0.04895021766424179,
0.08706755191087723,
-0.032027386128902435,
0.015255018137395382,
0.035223234444856644,
0.0678175613284111,
0.07053826749324799,
-0.01752477139234543,
0.05483169108629227,
-0.024852678179740906,
-0.012597281485795975,
0.0651620551943779,
0.07628083974123001,
0.06368748843669891,
0.07547254860401154,
-0.018484894186258316,
-0.12078334391117096,
0.027124272659420967,
-0.020255841314792633,
-0.015393784269690514,
-0.06261084228754044,
-0.0760611742734909,
0.027927827090024948,
-0.05405664071440697,
0.04748322442173958,
-0.003330186242237687,
0.03578570485115051,
-0.013510148040950298,
-0.05419362708926201,
0.0652141124010086,
0.018640931695699692,
0.028073394671082497,
0.06685010343790054,
-0.10818423330783844,
0.04933281987905502,
0.06260433793067932,
-0.0035657265689224005,
0.01175452396273613,
-0.029077209532260895,
-0.07176816463470459,
-0.020947830751538277,
0.03232565522193909,
-0.011106133460998535,
0.02318805269896984,
-0.05970078706741333,
0.0071261003613471985,
-0.004992843605577946,
-0.046296894550323486,
0.055147428065538406,
0.05208663269877434,
-0.03875342756509781,
0.06142458692193031,
-0.034790992736816406,
0.06859000027179718,
-0.03569354489445686,
-0.064297616481781,
-0.06798374652862549,
3.021833225754277e-33,
0.0487421415746212,
0.059301648288965225,
0.1137501448392868,
-0.010344009846448898,
0.12227671593427658,
0.02347744256258011,
-0.04160090535879135,
0.0528884120285511,
-0.024123230949044228,
0.03130674734711647,
-0.10260022431612015,
0.0462607741355896,
-0.028382208198308945,
0.10183992236852646,
-0.059962231665849686,
-0.07685641944408417,
0.008316567167639732,
-0.013456068933010101,
-0.012138299643993378,
-0.04760055989027023,
-0.042253538966178894,
-0.03097054548561573,
-0.04174255579710007,
0.021373063325881958,
-0.0456625260412693,
0.0031208747532218695,
0.03482871130108833,
-0.044986993074417114,
-0.04072209820151329,
-0.034527797251939774,
-0.07409686595201492,
-0.003896518610417843,
0.017991088330745697,
0.02929386869072914,
-0.029604516923427582,
-0.06918848305940628,
0.02069753408432007,
-0.04067644104361534,
0.05285007134079933,
-0.06250227987766266,
0.05803954228758812,
0.0733015239238739,
0.025627808645367622,
-0.0035691154189407825,
-0.01622277870774269,
-0.03241885453462601,
0.021073633804917336,
0.025107353925704956,
-0.03438297286629677,
0.014514252543449402,
-0.004192184191197157,
-0.03634116053581238,
-0.05822696164250374,
-0.06594855338335037,
0.022711627185344696,
-0.07939056307077408,
0.08272736519575119,
0.06100643053650856,
0.05830689147114754,
-0.014142367988824844,
0.049908362329006195,
-0.0044321538880467415,
0.001557609997689724,
0.11170170456171036,
-0.021125437691807747,
0.013429921120405197,
0.02812957391142845,
0.015143921598792076,
0.03648688644170761,
-0.016968127340078354,
-0.05299493297934532,
0.05837110057473183,
-0.03163176402449608,
-0.012545418925583363,
0.012574110180139542,
0.03733513131737709,
-0.04423651844263077,
-0.1633087396621704,
-0.05546066164970398,
0.05685865879058838,
-0.09823361039161682,
-0.06991708278656006,
-0.03414887189865112,
-0.011212811805307865,
-0.014173362404108047,
-0.06078372150659561,
0.05066835507750511,
-0.028110453858971596,
0.05695006623864174,
-0.10558141767978668,
-0.06224081665277481,
-0.06285412609577179,
-0.00045104665332473814,
0.016158008947968483,
0.003593842964619398,
-2.6976092922654354e-33,
-0.03694438561797142,
0.07758139818906784,
-0.014685817062854767,
-0.0037731537595391273,
0.02188795432448387,
-0.01691807620227337,
0.05947286635637283,
-0.008535505272448063,
-0.018944092094898224,
0.05047744885087013,
-0.01817067712545395,
-0.06844483315944672,
0.00731961103156209,
-0.06865337491035461,
0.06360681354999542,
-0.017814548686146736,
-0.08027952164411545,
0.021578243002295494,
0.003116594860330224,
0.03309240564703941,
-0.0014694324927404523,
0.1130577102303505,
-0.09220141917467117,
0.07894466072320938,
0.05538293719291687,
0.05833665281534195,
0.0867949053645134,
0.0007961760857142508,
0.03285003826022148,
0.06298669427633286,
-0.051480766385793686,
-0.08445955812931061,
0.03198232874274254,
0.04021444544196129,
0.03905729204416275,
0.003925880882889032,
0.04311605915427208,
0.03316352143883705,
-0.04811200872063637,
0.03977365791797638,
-0.014176861383020878,
-0.07706250250339508,
-0.02856229618191719,
0.056554775685071945,
-0.026004159823060036,
0.031711649149656296,
-0.06605524569749832,
0.008265757001936436,
-0.030403725802898407,
0.07975644618272781,
-0.07932987809181213,
0.04514213651418686,
-0.04308093339204788,
-0.04086679220199585,
-0.06286797672510147,
-0.05165451020002365,
-0.00297705945558846,
0.015195121988654137,
0.07895392924547195,
-0.050163235515356064,
0.033945199102163315,
-0.05958327651023865,
0.0641857236623764,
-0.05139822140336037,
-0.028424758464097977,
0.009733346290886402,
-0.0703275054693222,
0.050834640860557556,
-0.007463091518729925,
0.005605529993772507,
-0.07473447918891907,
0.0398724339902401,
-0.016147159039974213,
0.009690022096037865,
0.009311700239777565,
-0.008134161122143269,
0.006897527724504471,
-0.030710462480783463,
0.0033017939422279596,
-0.05786023661494255,
-0.04563235864043236,
-0.024981969967484474,
0.021828841418027878,
0.09794560074806213,
0.08581733703613281,
0.034352339804172516,
0.01905781216919422,
0.08247608691453934,
0.05085429176688194,
0.0378190316259861,
-0.01626277156174183,
0.06719803065061569,
-0.03025156445801258,
0.1106995940208435,
-0.011974348686635494,
-4.27895088250807e-8,
-0.15310929715633392,
0.06856006383895874,
0.03797229379415512,
-0.012813162989914417,
-0.008122707717120647,
-0.04983242228627205,
0.0812714695930481,
0.09702962636947632,
-0.04319417104125023,
0.013687011785805225,
0.060726672410964966,
-0.023274581879377365,
-0.007367235608398914,
-0.11714933812618256,
0.029629521071910858,
-0.010733465664088726,
0.033208779990673065,
-0.03509730100631714,
0.009181068278849125,
0.030591387301683426,
-0.0013838291633874178,
0.026651112362742424,
0.015500332228839397,
0.002958158729597926,
-0.02922048047184944,
-0.11488213390111923,
-0.0727544054389,
0.1334894299507141,
0.01195074338465929,
-0.008079797960817814,
0.021830836310982704,
0.04344302788376808,
0.06186668574810028,
0.024719350039958954,
-0.03897195681929588,
0.06343355774879456,
0.00200312165543437,
-0.020730098709464073,
-0.0419045016169548,
-0.05852215737104416,
-0.018283989280462265,
0.05056272819638252,
-0.0036928937770426273,
-0.005519218742847443,
0.021103590726852417,
-0.05190558359026909,
0.06944354623556137,
-0.06413711607456207,
0.0546751543879509,
-0.006503242067992687,
0.0007508064736612141,
0.03473961353302002,
-0.07304557412862778,
0.05894257873296738,
-0.029145298525691032,
0.03614289313554764,
0.0050629861652851105,
-0.07612315565347672,
0.02272009290754795,
0.05646318569779396,
-0.003542789723724127,
0.031468190252780914,
-0.01370245311409235,
-0.02002478949725628
] | 0 | 0 | 0 | 0 |
|
The Ramanujan Library - Automated Discovery on the Hypergraph of Integer Relations | https://openreview.net/forum?id=EyaH1wzmao | [
"Itay Beit Halachmi",
"Ido Kaminer"
] | Poster | Fundamental mathematical constants appear in nearly every field of science, from physics to biology.
Formulas that connect different constants often bring great insight by hinting at connections between previously disparate fields.
Discoveries of such relations, however, have remained scarce events, relying on sporadic strokes of creativity by human mathematicians.
Recent developments of algorithms for automated conjecture generation have accelerated the discovery of formulas for specific constants.
Yet, the discovery of connections between constants has not been addressed.
In this paper, we present the first library dedicated to mathematical constants and their interrelations. This library can serve as a central repository of knowledge for scientists from different areas, and as a collaborative platform for development of new algorithms.
The library is based on a new representation that we propose for organizing the formulas of mathematical constants:
a hypergraph, with each node representing a constant and each edge representing a formula.
Using this representation, we propose and demonstrate a systematic approach for automatically enriching this library using PSLQ, an integer relation algorithm based on QR decomposition and lattice construction. During its development and testing, our strategy led to the discovery of 75 previously unknown connections between constants, including a new formula for the `first continued fraction' constant $C_1$, novel formulas for natural logarithms, and new formulas connecting $\pi$ and $e$.
The latter formulas generalize a century-old relation between $\pi$ and $e$ by Ramanujan, which until now was considered a singular formula and is now found to be part of a broader mathematical structure.
The code supporting this library is a public, open-source API that can serve researchers in experimental mathematics and other fields of science. | Continued Fractions, Mathematical Constants, Integer Relations, Experimental Mathematics, Riemann Zeta Function, Irrational Number, PSLQ, AI In Mathematics, Automated Conjecture Generation | Automated generation of nonlinear formulas connecting mathematical constants, forming a publicly-available resource of constants and connections. | 12,034 | 2412.12361 | [
-0.10058002918958664,
0.029765227809548378,
-0.05605937913060188,
0.053104471415281296,
-0.038540102541446686,
-0.03853169456124306,
-0.02519688569009304,
-0.04634547978639603,
0.020147519186139107,
-0.06794245541095734,
-0.02864481694996357,
-0.06667742133140564,
0.010125009343028069,
0.10347484052181244,
-0.02319120429456234,
0.07011587917804718,
-0.015740834176540375,
0.06030583381652832,
-0.05310162901878357,
-0.0375957190990448,
0.004231988452374935,
0.020574618130922318,
-0.0538376048207283,
0.017249885946512222,
0.06230659782886505,
-0.020303335040807724,
0.0484275296330452,
0.022001223638653755,
0.0951552465558052,
0.06365051120519638,
0.04392116516828537,
0.00338547769933939,
0.013688520528376102,
-0.002705097896978259,
-0.014377148821949959,
0.008166100829839706,
-0.024210408329963684,
-0.015375880524516106,
-0.0016632596962153912,
-0.0001322357711615041,
0.014259940013289452,
0.01948906108736992,
-0.005495529156178236,
0.11041256785392761,
-0.002350602298974991,
0.043496113270521164,
-0.06369127333164215,
0.05270640552043915,
-0.02764318697154522,
-0.030072538182139397,
-0.07059533894062042,
-0.036425475031137466,
-0.062171924859285355,
0.008392603136599064,
0.07725207507610321,
-0.029635990038514137,
0.0011054450878873467,
-0.05978129804134369,
-0.011875958181917667,
-0.044759687036275864,
0.029346905648708344,
-0.023246869444847107,
0.0012681534281000495,
-0.0635928213596344,
0.00029493807232938707,
0.016076941043138504,
0.06349613517522812,
0.030244968831539154,
0.07615546882152557,
0.09552273154258728,
-0.0485258586704731,
0.10888703912496567,
-0.04538155347108841,
0.06125980615615845,
-0.004476587753742933,
0.09220948070287704,
-0.03347429633140564,
0.0021268336568027735,
-0.029787780717015266,
-0.04643857106566429,
-0.012031864374876022,
-0.02626297064125538,
-0.03996729850769043,
0.024268651381134987,
0.054828472435474396,
-0.06793271750211716,
0.020317386835813522,
0.06587889790534973,
0.10914992541074753,
0.010577661916613579,
-0.008180955424904823,
-0.05010215565562248,
0.0456353984773159,
-0.059239454567432404,
-0.05894548445940018,
0.08020920306444168,
0.057243868708610535,
-0.0034740972332656384,
0.016170430928468704,
0.04303527995944023,
0.017104417085647583,
0.0517449751496315,
0.03185661509633064,
-0.002660381142050028,
-0.0191139355301857,
0.005236483179032803,
-0.006450154818594456,
0.0611291229724884,
0.02015017345547676,
-0.032715216279029846,
-0.008641478605568409,
-0.007714101113379002,
0.002486067358404398,
0.044217973947525024,
0.005618181545287371,
-0.002244349103420973,
0.06147030368447304,
0.019391346722841263,
0.03739374130964279,
0.032638970762491226,
0.06270589679479599,
0.029624829068779945,
-0.008648080751299858,
0.034503038972616196,
0.03523479029536247,
0.02724238857626915,
-0.16803772747516632,
1.4304665141908305e-33,
0.08636030554771423,
0.1103576198220253,
0.0733104720711708,
0.07471844553947449,
0.060687195509672165,
-0.033109940588474274,
-0.030680648982524872,
-0.012890822254121304,
-0.05016358569264412,
0.06370259821414948,
-0.07711703330278397,
0.09982800483703613,
0.06284730136394501,
-0.06100962683558464,
-0.00637308694422245,
0.01324781309813261,
0.007170388475060463,
0.0011981009738519788,
-0.02228091098368168,
-0.08852170407772064,
0.024184752255678177,
-0.023577779531478882,
-0.04076393321156502,
0.022627612575888634,
0.024553311988711357,
0.09129732102155685,
0.03699266538023949,
-0.019145110622048378,
0.037189140915870667,
0.0011810626601800323,
-0.07077398151159286,
-0.007214323617517948,
-0.029670333489775658,
0.00965231005102396,
0.033796608448028564,
-0.03976377472281456,
-0.005944759119302034,
-0.07037901133298874,
0.02175210788846016,
-0.0679081380367279,
0.013579031452536583,
0.029730744659900665,
-0.009350246749818325,
-0.04128187894821167,
-0.011277339421212673,
-0.030634954571723938,
0.02035181038081646,
0.06994295120239258,
0.017150023952126503,
-0.030022088438272476,
-0.05061498284339905,
-0.04827192798256874,
0.009073449298739433,
-0.02978634648025036,
-0.03311868757009506,
-0.009576933458447456,
-0.023770494386553764,
0.05765550211071968,
0.005382447503507137,
0.09056154638528824,
0.014538072980940342,
0.04908839985728264,
0.007735663093626499,
0.08193790912628174,
-0.06220260262489319,
0.06897320598363876,
-0.06568770110607147,
-0.005471908953040838,
0.057739343494176865,
0.09505537152290344,
-0.07752422988414764,
0.07987967878580093,
-0.03194409981369972,
-0.02870228886604309,
0.04484075680375099,
-0.08702951669692993,
-0.039605315774679184,
-0.12324246019124985,
-0.05941566079854965,
0.0051277452148497105,
-0.08799281716346741,
0.043026965111494064,
0.0321233868598938,
0.0043879542499780655,
-0.07835203409194946,
-0.13571414351463318,
-0.006953742355108261,
0.046185098588466644,
-0.0393076054751873,
-0.09498010575771332,
0.014146903529763222,
0.015324369072914124,
0.08776777982711792,
-0.02095925435423851,
-0.025749169290065765,
-2.6187817431252467e-33,
-0.07478256523609161,
0.0010866138618439436,
0.07057585567235947,
0.016029292717576027,
0.01616627350449562,
-0.04311459884047508,
-0.05451013892889023,
-0.061979990452528,
0.015278493985533714,
-0.04568299278616905,
0.0623207725584507,
-0.059110790491104126,
-0.050005942583084106,
-0.0006121291080489755,
0.04560445249080658,
0.0158170685172081,
-0.09083722531795502,
0.0047552576288580894,
-0.03091573342680931,
0.03996143117547035,
-0.058005061000585556,
0.059630755335092545,
-0.04991576820611954,
-0.04324640706181526,
0.023004213348031044,
0.024775076657533646,
0.06830042600631714,
-0.02937893196940422,
-0.06936587393283844,
0.12299299240112305,
-0.02503160573542118,
-0.06553210318088531,
-0.1014752984046936,
-0.14022496342658997,
-0.062222085893154144,
-0.007289236411452293,
0.06544680148363113,
-0.007284393068403006,
0.033221982419490814,
0.002099312376230955,
-0.02010739967226982,
0.012963879853487015,
0.030065475031733513,
0.018308421596884727,
0.04954209551215172,
0.011273120529949665,
0.031674791127443314,
0.010410214774310589,
-0.019177481532096863,
-0.07878275215625763,
0.0026664058677852154,
-0.08271101117134094,
-0.04180188849568367,
-0.0009472431265749037,
-0.01955091767013073,
0.005305628292262554,
-0.07480604201555252,
0.108753502368927,
0.00859734695404768,
-0.06123262643814087,
-0.04935446009039879,
-0.048048973083496094,
0.029991047456860542,
0.07006219029426575,
-0.07177851349115372,
-0.078483447432518,
0.015649350360035896,
0.018673108890652657,
-0.023707738146185875,
-0.047127481549978256,
0.03090917505323887,
0.02890288643538952,
-0.050938792526721954,
0.09706033021211624,
-0.052930042147636414,
-0.038563892245292664,
0.019610408693552017,
0.03561853989958763,
-0.005371357314288616,
0.03755078464746475,
0.0401359461247921,
0.08591828495264053,
-0.003646932076662779,
-0.05178585648536682,
-0.004961760714650154,
-0.04015280678868294,
0.06274629384279251,
0.10992352664470673,
-0.027533026412129402,
-0.031158221885561943,
-0.06647656112909317,
0.035783421248197556,
-0.05573994293808937,
0.11688750237226486,
0.020520629361271858,
-4.124876440414482e-8,
0.0025060551706701517,
0.019415302202105522,
-0.08487873524427414,
-0.09232412278652191,
0.06651049852371216,
0.03876030817627907,
0.016495145857334137,
0.07478104531764984,
-0.011425627395510674,
0.038392867892980576,
0.10049709677696228,
-0.04116804152727127,
0.017549928277730942,
0.030141429975628853,
-0.056496184319257736,
0.042273420840501785,
0.062084540724754333,
-0.01994672603905201,
-0.05973083898425102,
-0.05066302418708801,
-0.010123584419488907,
0.06679049879312515,
0.029009908437728882,
-0.0010150454472750425,
0.09576327353715897,
-0.07731126248836517,
-0.09805773198604584,
-0.006555605214089155,
-0.022384006530046463,
-0.009182721376419067,
0.021706473082304,
-0.04518985003232956,
0.06987088918685913,
0.018726317211985588,
-0.006077825091779232,
0.09796905517578125,
0.039923470467329025,
-0.01867966167628765,
0.034921303391456604,
0.044439684599637985,
-0.04635189101099968,
0.009791008196771145,
-0.08313247561454773,
0.008429966866970062,
0.05485224723815918,
0.005084121599793434,
0.009971808642148972,
0.004906223155558109,
-0.09610029309988022,
0.04246889427304268,
-0.004335773177444935,
-0.008576716296374798,
-0.028620382770895958,
-0.0842750295996666,
-0.02046317607164383,
0.026051726192235947,
-0.11353583633899689,
-0.0034381095319986343,
-0.008627701550722122,
-0.04468809440732002,
0.0367751270532608,
0.007325084414333105,
-0.012556137517094612,
0.016066256910562515
] | https://github.com/ramanujanmachine/lirec | 3 | 0 | 0 | 0 |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.