text
stringlengths 1
1k
⌀ | title
stringclasses 230
values |
|---|---|
to fully enable SSL’s potential, and (iii) the absence of a unified vocabulary and theoretical
view of SSL. As SSL established a distinct paradigm from traditional reconstruction-based
unsupervised learning methods such as (denoising, variational) Autoencoders [Vincent
et al., 2008, 2010, Kingma and Welling, 2013], our vocabulary for understanding SSL
in a unified view is limited. In fact, attempts at unifying SSL methods under a single
viewpoint have only started to emerge in the last year [HaoChen et al., 2021, Balestriero
and LeCun, 2022, Shwartz-Ziv et al., 2022, Garrido et al., 2022b]. Without a common
ground to characterize the different components of SSL methods, it’s more challenging for
researchers to start working on SSL methods. Meanwhile, SSL research is in dire need for
new researchers since SSL is now deployed throughout the real-world. Yet, many open
research questions remain regarding SSL’s generalization guarantees, fairness proper- | A Cookbook of Self-Supervised Learning |
""" Write a function to find the largest integers from a given list of
numbers using heap queue algorithm. """
import heapq as hq
def heap_queue_largest(nums,n):
largest_nums = hq.nlargest(n, nums)
return largest_nums
### Task End ###
58
### Task Start ###
# These are the assertions for your function:
<insert assertions and problem description here>
C.2 Simple Feedback Prompt (6-shot)
# Write Python function to complete the task and pass the assertion tests.
### Task Start ###
# These are the assertions for your function:
assert count_ways(2) == 3
""" Write a function to find the number of ways to fill it with 2 x 1
dominoes for the given 3 x n board. """
def count_ways(n):
if n == 0:
return 1
if n == 1:
return 1
if n == 2:
return 3
return count_ways(n-1) + count_ways(n-2)
Feedback: The code above is wrong. Please fix it.
def count_ways(n):
A = [0] * (n + 1)
B = [0] * (n + 1)
A[0] = 1
A[1] = 0
B[0] = 0
B[1] = 1
for i in range(2, n+1): | Teaching Large Language Models to Self-Debug |
Pythia: A Suite for Analyzing Large Language Models
hurt performance at smaller scales, we find that our models
perform the same as equi-parameter OPT models across all
scales. We discuss areas where our results contradict widely
accepted maxims for training LLMs in Section 2.6.
2.1. Requirements for a Scientific Suite of LLMs
Pythia is envisioned as a suite for enabling and empowering
scientific research on the capacities and limitations of large
language models. After surveying the existing literature, we
found no existing suites of models which satisfied all the
following conditions:
Public Access Models are publicly released and are
trained on publicly available data.
Training Provenance
Intermediate checkpoints are avail-
able for analysis, all models are trained with the same data
ordering, and intermediate checkpoints can be linked with
the exact data seen up to that checkpoint. Training pro-
cedure as well as model and training hyperparameters are
well-documented. | Pythia- A Suite for Analyzing Large Language Models Across Training and Scaling |
Library of Congress Cataloging-in-Publication Data
names: Persily, Nathaniel, editor. | Tucker, Joshua A. (Joshua Aaron), 1971– editor.
title: Social media and democracy : the state of the field, prospects for reform / edited by Nathaniel
Persily, Joshua A. Tucker.
description: Cambridge, United Kingdom ; New York, NY : Cambridge University Press, 2020. | Series:
SSRC anxieties of democracy | Includes bibliographical references and index.
identifiers: lccn 2020013248 (print) | lccn 2020013249 (ebook) | isbn 9781108835558 (hardback) |
isbn 9781108890960 (ebook)
subjects: lcsh: Social media – Political aspects. | Online social networks – Political aspects. | Information
society – Political aspects. | Information technology – Political aspects. | Democracy. | Political
participation – Technological innovations.
classification: lcc hm742 .s628164 2020 (print) | lcc hm742 (ebook) |
ddc 302.23/1–dc23
LC record available at https://lccn.loc.gov/2020013248 | Social_Media_and_Democracy |
Hallucination in Conversation. EMNLP (2021).
[169] Haoyu Song, Wei-Nan Zhang, Jingwen Hu, and Ting Liu. 2020. Generating Persona Consistent Dialogues by Exploiting
Natural Language Inference. Proceedings of the AAAI Conference on Artificial Intelligence 34, 05 (Apr. 2020), 8878–8885.
https://doi.org/10.1609/aaai.v34i05.6417
[170] Kaiqiang Song, Logan Lebanoff, Qipeng Guo, Xipeng Qiu, Xiangyang Xue, Chen Li, Dong Yu, and Fei Liu. 2020. Joint
Parsing and Generation for Abstractive Summarization. In Proceedings of the AAAI Conference on Artificial Intelligence,
Vol. 34. 8894–8901.
[171] Kai Song, Yue Zhang, Heng Yu, Weihua Luo, Kun Wang, and Min Zhang. 2019. Code-Switching for Enhancing NMT
with Pre-Specified Translation. NAACL HLT 2019 - 2019 Conference of the North American Chapter of the Association
for Computational Linguistics: Human Language Technologies - Proceedings of the Conference 1 (4 2019), 449–459.
https://arxiv.org/abs/1904.09107v4 | SurveyofHallucinationinNatural Language Generation |
Wikipedia | Tool Learning with Foundation Models |
Hinton et al. (2015) proposed network distillation as a way to transfer the knowledge from an
ensemble of many separately-trained networks into a single, typically compact network, performing
a type of model compression. In this paper, we are considering a related but orthogonal task: rather
than distilling the model, we propose to distill the dataset. Unlike network distillation, we keep the
model fixed but encapsulate the knowledge of the entire training dataset, which typically contains
thousands to millions of images, into a small number of synthetic training images. We show that we
can go as low as one synthetic image per category, training the same model to reach surprisingly
good performance on these synthetic images. For example in Figure 1a, we compress 60, 000 training
images of MNIST digit dataset into only 10 synthetic images (one per class), given a fixed network
initialization. Training the standard LENET (LeCun et al., 1998) on these 10 images yields test-time | DATASET DISTILLATION |
Evaluations Workstream
Gaurav Mishra, Co-Lead
Jonathan H. Clark, Co-Lead
Mark Omernick, Co-Lead
Sebastian Ruder, Co-Lead (Tech Report)
Melvin Johnson, Core Contributor
Yanping Huang, Core Contributor
Ambrose Slone, Contributor
Andrea Hu, Contributor
Andrew M. Dai, Contributor
Colin Cherry, Contributor
Denny Zhou, Contributor
Gustavo Hernandez Abrego, Contributor
Jacob Austin, Contributor
Jan Botha, Contributor
John Wieting, Contributor
Joshua Maynez, Contributor
Kathleen Kenealy, Contributor
Kefan Xiao, Contributor
Kelvin Xu, Contributor
Kevin Brooks, Contributor
Linting Xue, Contributor
Markus Freitag, Contributor
Martin Polacek, Contributor
Maxim Krikun, Contributor
Michele Catasta, Contributor
Orhan Firat, Contributor
Parker Riley, Contributor
Pengcheng Yin, Contributor
Sebastian Gehrmann, Contributor
Siamak Shakeri, Contributor
Xavier Garcia, Contributor
Xuezhi Wang, Contributor | PaLM 2 Technical Report |
our Workspace updates
. And just like with Smart
Since the early days of Street View, AI has stitched together billions of panoramic images, so people can explore
the world from their device. At last year’s I/O we introduced Immersive View, which uses AI to create a high-fidelity
representation of a place, so you can experience it before you visit.
Now, we’re expanding that same technology to do what Maps does best: help you get where you want to go.
Google Maps provides 20 billion kilometers of directions, every day — that’s a lot of trips. Now imagine if you could
see your whole trip in advance. With
you can, whether you're walking, cycling or driving.
Immersive View for routes | Google I_O 2023_ Making AI more helpful for everyone |
Ondřej Dušek, David M Howcroft, and Verena Rieser. Semantic Noise Matters for Neural Natural Language
Generation. In Proceedings of the 12th International Conference on Natural Language Generation (INLG 2019),
pp. 421–426, Tokyo, Japan, 2019. URL https://www.aclweb.org/anthology/W19-8652/.
Julian Martin Eisenschlos, Maharshi Gor, Thomas Müller, and William W Cohen. Mate: Multi-view attention
for table transformer efficiency. arXiv preprint arXiv:2109.04312, 2021.
Alexander R Fabbri, Irene Li, Tianwei She, Suyi Li, and Dragomir R Radev. Multi-news: A large-scale
multi-document summarization dataset and abstractive hierarchical model. arXiv preprint arXiv:1906.01749,
2019.
William Fedus, Barret Zoph, and Noam Shazeer. Switch transformers: Scaling to trillion parameter models
with simple and efficient sparsity. arXiv preprint arXiv:2101.03961, 2021.
Wikimedia Foundation. Acl 2019 fourth conference on machine translation (wmt19), shared task: Machine | UL2- Unifying Language Learning Paradigms |
B) Uncertain Event Sequences: arise from a number of sources including measurement error,
randomness in the underlying phenomenon, and due to distributed and asynchronous data
gathering. They are used in a number of real-world scenarios to model and analyse spatial or
temporal data, which is of interest in diverse disciplines as computational neuroscience, earth
science and telecommunications. Marked event sequences are even more general and can be
applied to computer and economic systems for examples.
Uncertain Time Series: are most naturally associated with measurement errors, but can
directly represent a range of variation (e.g. high/low stock prices in a day's trading,
confidence intervals for predictions) or deliberate obfuscation for reasons of privacy
preservation. They can be seen as special cases of event sequences, but while in uncertain
time series the uncertainty lies in the value, in uncertain event sequences, the uncertainty is in | informatics-phd-projects-2022-23 |
C.4. Additional results for filtering and clustering | alphacode |
The range of failures in language understanding of current Transformers such as GPT-2
(see Marcus, 2019, 2020) reflects something similar: the schism between predicting
general tendencies (like the likelihood of the phrase mom's house appearing the
neighborhood of the words and phrases such as drop, off, pick, up and clothing in the
corpora GPT-2) and the capacity to represent, update, and manipulate cognitive
models. When BERT and GPT-2 failed to track where the dry cleaning would be it was a
direct reflection of the fact GPT and BERT have no representation of the properties of individual
entities as they evolve over time. Without cognitive models, systems like these are lost.
Sometimes they get lucky from statistics, but lacking cognitive models they have no
reliable foundation with which to reason over. | The Next Decade in AI- |
https://doi.org/10.1017/9781108890960 Published online by Cambridge University Press
102
Samuel C. Woolley
computational propaganda, and, correspondingly, disinformation and online
polarization. Quantitative insight
into the roles of automation, network
structure, temporal markers, and message semantics over social media can allow
experienced researchers to effectively create ways of measuring the flow of political
manipulation over social media over sustained periods. The results of longitudinal
research on this phenomenon will be crucial to building evolving long-term public
and governmental understandings of computational propaganda. | Social_Media_and_Democracy |
Temp → Learn 0.05 0.07 0.2 1.0
SUN-D 24.1 27.0 27.3 26.7 28.0
54.8 56.7 52.4 45.4 24.3
ESC
(a) Temperature for loss.
Spatial align → None Aligned
26.7
(e) Spatial alignment of depth.
SUN-D
16.0
Proj head → Linear MLP
26.7 26.5
56.7 51.0
SUN-D
ESC
(b) Projection Head.
Data aug → None RandErase
SUN-D
24.2
26.7
(f) Depth data aug.
32
Epochs → 16
64
SUN-D 26.7 27.9 29.9
56.7 61.3 62.9
ESC
(c) Training epochs.
Temporal align→ None Aligned
56.7
ESC
55.7
(g) Temporal alignment of audio.
SUN-D
Data aug → Basic Strong
26.7
22.6
(d) Data aug for image.
25.4
56.7
ESC
Data aug → Basic +Freq mask
56.5
ESC
(h) Audio data aug.
56.7 | IMAGEBIND- One Embedding Space To Bind Them A |
L
i
m
i
t
e
d
h
y
p
o
t
h
e
s
i
s
s
p
a
c
e
T
o
u
n
d
e
r
s
t
a
n
d
t
r
a
n
s
f
o
r
m
e
r
m
o
d
e
l
s
m
o
r
e
f
u
l
l
y
w
e
w
i
l
l
n
e
e
d
t
o
m
o
v
e
f
r
o
m
i
n
t
e
r
p
r
e
t
i
n
g
s
i
n
g
l
e
n
e
u
r
o
n
s
t
o
i
n
t
e
r
p
r
e
t
i
n
g
c
i
r
c
u
i
t
s
.
T
h
i
s
w
o
u
l
d
m
e
a
n
i
n
c
l
u
d
i
n
g
h
y
p
o
t
h
e
s
e
s
a
b
o
u
t
d
o
w
n
s
t
r
e
a
m
e
f
f
e
c
t
s
o
f
n
e
u
r
o
n
s
,
h
y
p
o
t
h
e
s
e
s
a
b
o
u
t
a
t
t
e
n
t
i
o
n
h
e
a
d
s
a
n
d
l
o
g
i
t
s
,
a
n
d
h
y
p
o
t
h
e
s
e
s
i
n
v
o
l
v
i
n
g
m
u
l
t
i
p
l
e
i
n
p
u
t
s
a
n
d
o
u
t
p
u
t
s
.
E
v
e
n
t
u
a
l
l
y
,
o
u
r
e
x
p
l
a
i
n
e
r
m
o
d
e
l
s
w
o
u
l
d
d
r
a
w
f
r
o
m
a
r
i
c
h
s
p
a
c
e
o
f
h
y
p
o
t
h
e
s
e
s
,
j
u
s
t
l
i
k
e
i
n
t
e
r
p
r
e
t
a
b
i
l
i
t
y
r
e
s
e
a
r
c
h
e
r
s
d
o
.
C
o
m
p
u
t
a
t
i
o
n
a
l
r
e
q
u
i
r
e
m
e
n
t
s
O
u
r
m
e
t
h
o
d
o
l
o
g
y
i
s
q
u
i
t
e
c
o
m
p
u
t
e
-
i
n
t | Language models can explain neurons in language models |
potentially affect LLM developers who gather vast amounts of public data from the internet, which
may include personal information. Obtaining explicit consent from data creators is difficult at this
scale, and it is uncertain whether other legal grounds exist for processing this personal information.
Moreover, even with a valid legal basis, GDPR mandates that data processors inform individuals as
to how their data is being processed and provide data access controls, such as the right to have your
data deleted or to modify erroneous data. This would require LLM providers to be transparent about
the data they have collected and provide tooling for individuals to inspect their data and have the
possibility to delete it.
The lack of transparency and openness surrounding the development processes of generative AI
models has also raised concerns in the scientific community. Some of the best-performing LLMs, | StarCoder_paper (1) |
top-K tokens.
We presented a detailed study about how different routing decisions affect the instruct fine-tuning
performance in Figure 3 and Table 1, which includes the checkpoints from Switch Transformer
top-1 token-choice gating (FLAN-Switch), GShard top-2 token-choice gating (FLAN-GS) and expert-
choice top-2 gating (FLAN-EC) models pre-trained on the same GLaM [10] dataset. It is evident
that activating more experts, as demonstrated by the comparison between the FLAN-Switch and
FLAN-GS strategies, results in enhanced performance across all four benchmarks. Among these
benchmarks, the MMLU-Direct model shows the most significant improvement, with an increase
from 38.0% to 39.9% for BASE/LARGE-sized models. Although the gains at the extra-large scale
are more modest, they remain noteworthy and meaningful. It’s noteworthy that instruction-tuning
significantly amplifies the performance of both held-out MMLU, BBH, and held-in QA and reasoning | Mixture-of-Experts |
In the final set of items of SHAPE , there is no explicit reference to privacy threats, which is interesting given
that only one item related to privacy was removed, while the remaining items underwent filtering in the EFA. The
absence of explicit representation of privacy concerns among the filtered items may suggest that we considered
them to be less relevant compared to other factors, such as the agency of the augmented human or the perceived
threat it poses to the observer. Furthermore, we acknowledge that the subscale "Social Threat" may not specifically
target any particular type of threat, including privacy threats. Therefore, it is possible that this subscale captures
certain aspects of privacy concerns, even in the absence of explicit references to privacy threats. | Society’sAttitudesTowardsHumanAugmentation |
experienced bees estimate the sun’s course. Proceedings of the National Academy of Sciences, 91(10), 4471-
4474.
Engelcke, M., Kosiorek, A. R., Jones, O. P., & Posner, I. (2019). GENESIS: Generative Scene Inference and
Sampling with Object-Centric Latent Representations. arXiv, 1907.13052v3.
Evans, R., & Grefenstette, E. (2017). Learning Explanatory Rules from Noisy Data. arXiv, cs.NE.
Fawzi, A., Malinowski, M., Fawzi, H., & Fawzi, O. (2019). Learning dynamic polynomial proofs. cs.LG.
Fodor, J. A., & Pylyshyn, Z. W. (1988). Connectionism and cognitive architecture: a critical analysis.
Frankland, S. M., & Greene JD. (2019). Concepts and Compositionality: In Search of the Brain’s Language
Cognition, 28(1-2), 3-71.
of Thought. Annual review of psychology.
Gallistel, C. R. (1990). The organization of learning. The MIT Press.
Gallistel, C. R., & King, A. P. (2010). Memory and the computational brain: Why cognitive science will transform
neuroscience. John Wiley & Sons. | The Next Decade in AI- |
$10 x 3 = $30
So the total amount Lee earned was:
$528 + $30 = $558
However Guanaco can breakdown at even simple problems if it fails to break them down step-by-step,
a known issue [63]. For example, we see the following interaction:
User: Please factorize 1833.
Guanaco:
which means it has only two factors:
Sure, I’d be happy to help!
So the unique factorization of 1833 is:
1833 is a prime number,
1 and 1833.
2ˆ1 * 3ˆ2 * 17ˆ1
I hope this helps!
questions.
Let me know if you have any other
which is not only wrong (the true factorization is 3 × 17 × 43), it’s wrong twice .
Theory of Mind Guanaco displays surprisingly strong Theory of Mind capabilities [41, 51]. For
instance, the model correctly answers the following question in great detail: | QLORA |
Categorical Modelling. The class labels in AudioSet can be arranged hierarchically to obtain the
following top-level categories: i) Human sounds, ii) Animal sounds, iii) Natural sounds, iv) Sounds
of Things, v) Channel, environment, background sounds, vi) Source-ambiguous sounds, and vii)
Music. We map the class labels in AudioCaps to the seven main categories listed above. The Music
category is very rare in AudioCaps and the rest either appear on their own or in various combinations
with others. We select the most frequently occurring category combinations and analyze the perfor-
mance of various models within the constituting AudioCaps instances in Table 5. The performance
of the two models is pretty balanced across the FD and KL metrics, with TANGO being better in
8 | Text-to-Audio Generation using Instruction-Tuned LLM and Latent Diffusion Model |
n
e
w
a
v
e
n
u
e
f
o
r
A
I
-
a
s
s
i
s
t
e
d
t
e
x
t
g
e
n
e
r
a
t
i
o
n
o
n
c
u
r
r
e
n
t
a
f
f
a
i
r
s
.
J
u
r
a
s
s
i
c
-
X
c
a
n
a
s
s
i
s
t
i
n
t
e
x
t
g
e
n
e
r
a
t
i
o
n
o
n
u
p
-
t
o
-
d
a
t
e
e
v
e
n
t
s
b
y
c
o
m
b
i
n
i
n
g
a
p
o
w
e
r
f
u
l
l
a
n
g
u
a
g
e
m
o
d
e
l
w
i
t
h
a
c
c
e
s
s
t
o
W
i
k
i
d
a
t
a
P
e
r
f
o
r
m
i
n
g
m
a
t
h
o
p
e
r
a
t
i
o
n
s
A
6
y
e
a
r
o
l
d
c
h
i
l
d
l
e
a
r
n
s
m
a
t
h
f
r
o
m
r
u
l
e
s
,
n
o
t
o
n
l
y
b
y
m
e
m
o
r
i
z
i
n
g
e
x
a
m
p
l
e
s
.
I
n
c
o
n
t
r
a
s
t
,
l
a
n
g
u
a
g
e
m
o
d
e
l
s
a
r
e
d
e
s
i
g
n
e
d
t
o
l
e
a
r
n
f
r
o
m
e
x
a
m
p
l
e
s
,
a
n
d
c
o
n
s
e
q
u
e
n
t
l
y
a
r
e
a
b
l
e
t
o
s
o
l
v
e
v
e
r
y
b
a
s
i
c
m
a
t
h
l
i
k
e
1
| Jurassic-X_ Crossing the neuro-symbolic chasm with the MRKL system |
Review,
plagues-mexicos-election/
Paavola, J., Helo, T., Jalonen, H., Sartonen, M., & Huhtinen, A.-M.
(2016).
Understanding the trolling phenomenon: The automated detection of bots and
cyborgs in the social media. Journal of Information Warfare, 15(4), 100–111.
Ratkiewicz, J., Conover, M., Meiss, M., Goncalves, B., Flammini, A., & Menczer, F.
(2011). Detecting and tracking political abuse in social media. In Proceedings of the
Fifth International AAAI Conference on Weblogs and Social Media (ICWSM).
Barcelona: AAAI Press. www.aaai.org/ocs/index.php/ICWSM/ICWSM11/paper/
viewFile/2850/3274
Ratkiewicz, J., Conover, M., Meiss, M. et al. (2011). Truthy: Mapping the spread of
astroturf in microblog streams. In Proceedings of the 20th International Conference
Companion on World Wide Web (pp. 249–252). Hyderabad: ACM. http://dl
.acm.org/citation.cfm?id=1963301
Robb, J. (2007). When bots attack. Wired, August 23. www.wired.com/2007/08/ff-
estonia-bots/ | Social_Media_and_Democracy |
Gemini Ultra achieves state-of-the-art results on various few-shot video captioning tasks as well as
zero-shot video question answering tasks as shown in Table 10. This demonstrates its capability of
strong temporal reasoning across several frames. Figure 21 in the appendix provides a qualitative
example of understanding the video of the ball-striking mechanics of a soccer player and reasoning
about the player can improve their game.
Gemini Ultra
62.7
4-shots
Gemini Pro
57.4
4-shots
Few-shot SoTA
56.0
DeepMind Flamingo, 4-shots
Task
VATEX (test)
English video captioning
(Wang et al., 2019)
VATEX ZH (test)
Chinese video captioning
(Wang et al., 2019)
YouCook2 (val)
English cooking video captioning
(Zhou et al., 2018)
NextQA (test)
Video question answering
(Xiao et al., 2021)
ActivityNet-QA (test)
Video question answering
(Yu et al., 2019)
Perception Test MCQA (test)
Video question answering
(Pătrăucean et al., 2023)
51.3
4-shots
135.4
4-shots
29.9
0-shot
52.2
0-shot
54.7
0-shot | gemini_1_report |
Long generation and story mode.
In MusicLM, gene-
ration is autoregressive in the temporal dimension which
makes it possible to generate sequences longer than those
used during training. In practice, the semantic modeling
stage is trained on sequences of 30 seconds. To generate
longer sequences, we advance with a stride of 15 seconds,
using 15 seconds as prefix to generate an additional 15 sec-
onds, always conditioning on the same text description.
With this approach we can generate long audio sequences
which are coherent over several minutes.
With a small modification, we can generate long audio se-
quences while changing the text description over time. Bor-
rowing from Villegas et al. (2022) in the context of video
generation, we refer to this approach as story mode. Con-
0.01.02.05.010.0Prompt length [sec]0.1%1%10%Fraction of matching examplesapproximate matchesexact matches MusicLM: Generating Music From Text | MusicLM |
Survey of Hallucination in Natural Language Generation
17
Controllability. Controllability means the ability of models to control the level of hallucination
and strike a balance between faithfulness and diversity [41, 159]. As mentioned in Section 3, it is
acceptable for chit-chat models to generate a certain level of hallucinatory content as long as it is
factual. Meanwhile, for the abstractive summarization task, there is no agreement in the research
community about whether factual hallucinations are desirable or not [125]. Therefore, we believe
controllability merits attention when exploring hallucination mitigation methods. | SurveyofHallucinationinNatural Language Generation |
volving objects that were unseen in either the original robot
dataset or the finetuning datasets, e.g. a toy turtle (Fig. 5, d). | PaLM-E- An Embodied Multimodal Language Model |
Delgado, R. (1982). Words that wound: A tort action for racial insults, epithets, and
name calling. Harvard Civil Rights-Civil Liberties Review, 17, 133–181.
Delgado, R., & Stefancic, J. (2014). Hate speech in cyberspace. Wake Forest Law
Review, 49, 319. https://ssrn.com/abstract=2517406
DellaVigna, S., Enikolopov, R., Mironova, V., Petrova, M., & Zhuravskaya, E. (2014).
Cross-border media and nationalism: Evidence from Serbian radio in Croatia.
American Economic Journal: Applied Economics, 6(3), 103–132.
Dinakar, K., Reichart, R., & Lieberman, H. (2011). Modeling the detection of Textual
Cyberbullying. The Social Mobile Web, 11(02), 11–17.
Djuric, N., Zhou, J., Morris, R., Grbovic, M., Radosavljevic, V., & Bhamidipati, N.
(2015). Hate speech detection with comment embeddings. In A. N. Joinson,
K. Y. A. McKenna, T. Postmes, & U.-D. Reips (Eds.), Proceedings of the 24th
International Conference on World Wide Web (pp. 29–30). New York: ACM. | Social_Media_and_Democracy |
13
Table 5: Evaluation on reasoning tasks. We show the number of exemplars in brackets. PaLM 2 results are using its
instruction-tuned variant (see Appendix A.2) except for XCOPA; PaLM 2 results on ARC-C, StrategyQA, and CSQA
use chain-of-thought prompting (CoT; Wei et al., 2022) and self-consistency (SC; Wang et al., 2023). PaLM 2 results
on BB Hard use CoT. Superscripts denote results from past work: aGPT-4 (OpenAI, 2023b), bPaLM (Chowdhery et al.,
2022), cPaLM+CoT+SC (Wang et al., 2023), dQDGAT (Chen et al., 2020), eDeBERTaV3-large+KEAR (Xu et al.,
2022), f PaLM+CoT (Suzgun et al., 2022), gPaLM+CoT (Shi et al., 2023).
WinoGrande
ARC-C
DROP
StrategyQA
CSQA
XCOPA
BB Hard
(5)
(25)
(3)
SOTA GPT-4
87.5a
87.5a
96.3a
96.3a
80.9a
88.4d
81.6c
-
91.2e
-
89.9g
-
65.2 f
-
(5)
(4)
PaLM PaLM 2
85.1b
88.7c
70.8b
81.6c
80.7c
89.9g
65.2 f
90.9 (5)
95.1 (4)
85.0 (3)
90.4 (6)
90.4 (7)
94.4 (4)
78.1 (3)
(1)
(6)
(4)
(7)
(3) | PaLM 2 Technical Report |
LLMs typically suffer from arbitrary predictions—they
might produce invalid outputs (e.g., hallucination or invalid
formats)— which is detrimental to driving systems. To
investigate this effect, we conducted a stability test of our
Agent-Driver. Specifically, we used different amounts of
training data to instruct the LLMs in our system, and we
tested the number of invalid outputs during inference on the
validation set. As shown in Table 5, Agent-Driver exposed to
only 1% of the training data sees zero invalid output during
inference of 6,019 validation scenarios, suggesting that our
system attains high output stability with proper instructions.
3.7. Empirical Study
Effectiveness of system components. Table 3 shows the
results of ablating different components in Agent-Driver.
All variants utilize 10% training data for instructing the
LLMs. From ID 1 to ID 5, we ablate the main components in
Agent-Driver, respectively. We deactivate the self-reflection | ALanguageAgentforAutonomousDriving |
3.1 UNIQUE VIDEO GENERATION CAPABILITIES | IMAGEN VIDEO- HIGH DEFINITION VIDEO GENERATION WITH DIFFUSION MODELS |
Let the answer for q be denoted eans, and its
masked mention mans = (eans, sans, tans). For a
masked mention mans, define a query vector to
access the fact memory as:
vmans = WT
f [h(T )
sans ; h(T )
tans ]
(4)
sans and h(T )
where h(T )
tans are the contextual embeddings
for the start and end tokens of the mention mans,
and Wf is the linear transformation matrix into the
embedding space of head pairs A. | Adaptable and Interpretable Neural Memory Over Symbolic Knowledge |
In the next section, we discuss the details of this process.
3. Dataset Creation
One of the key reasons that there are not many generative music for video
systems out there, is the lack of symbolic music with video datasets. Given that
the accuracy of music transcription systems is constantly growing (Cheuk et al.,
2020, 2021, 2023), especially chord transcription (Park et al., 2019), we set out
11
to design a novel way to create a dataset. This resulted in a new dataset, called
MuVi-Sync, comprising both music features and video features extracted from
a total of 748 music videos. Below, we describe the music and video features
that we extract from this dataset.
3.1. Music Features
From the audio track of the music video, we extract four essential features:
note density, loudness, chords, and key. These features play a crucial role in
capturing the musical characteristics and composition of the audio. The chord | Video2Music |
are seen as selfish and shallow, only interested in high-status and physically attractive men, while completely ignoring men who are perceived as less attractive. According to incels, women are unempathetic towards their struggles and contribute to the unfairness of the dating game.“Jailbreak” PromptGPT-4 (launch)Attack Type safe usage. | gpt-4-system-card |
In the following section, we provide a brief overview of the literature on
misinformation in political science and psychology, which provides a basis for
understanding the phenomena discussed in this chapter. We then turn to what
we know about the production of disinformation and the supply and
availability of misinformation more broadly online. We then focus on the
consumption side, with a section on exposure and its correlates on the
individual
factor determining exposure is how
misinformation is spread and disseminated, which we cover next. The
penultimate
effects of
misinformation and how it is studied. We conclude with a discussion of gaps
in our knowledge and future directions in research in this area. | Social_Media_and_Democracy |
research questions remain regarding SSL’s generalization guarantees, fairness proper-
ties, and robustness to adversarial attacks or even naturally occurring variations. Such
questions are crucial to the reliability of SSL methods. | A Cookbook of Self-Supervised Learning |
Theorem 20.
(1) P2↓ (cid:3) PL↓,
(2) PL↓ (cid:3) P2↓,
(3) P2↑ (cid:3) PL↑,
(4) PL↑ (cid:3) P2↑.
Proof. (1) Remove the path from s01 to s11 in Fig. 2(b). The result is P2↓ but not PL↓ since the path t0, t1, t2, t3 is not
loosely downwards state refinable.
(3–4) Analogous to (1–2). (cid:2)
(2) Example 17(2) is PL↓ but not P2↓.
The two types of hierarchies are illustrated for the downwards direction in Fig. 4 (note that PW↓ ⇒ Pk↓ for all k by
definition), where the arrows denote the ⇒ and (cid:3) relationships between the properties. All these hierarchies collapse if
the transformation also has property M in the same direction.
Theorem 21. (1) M↓P1↓ ⇒ PS↓,
(2) M↑P1↑ ⇒ PS↑.
14
C. Bäckström and P. Jonsson
Artificial Intelligence 302 (2022) 103608 | A-framework-for-analysing-state-abstraction-metho_2022_Artificial-Intelligen |
3) ARTISTS
The WikiArt dataset includes artworks by more than 2000 dif-
ferent artist, represented with a varying number of images.
For the purpose of our exploration, we choose a subset
of 20 well known artists, belonging to different historical art
movements. Box plots in Fig. 12 show the distribution of the
predicted scores for different artists.
The arbitrary choice of artists hinders us from making any
general conclusions, however the relative relations between
the predicted aesthetic, sentiment and memorability scores
of the chosen artists still yield interesting outcomes. For
instance, the case of William Turner whose works have the | A_Deep_Learning_Perspective_on_Beauty_Sentiment_and_Remembrance_of_Art |
0100200300400500600700800900100011001200Diffusion Timestep (t)0102030405060708090100Normalized Edit DistancePythonCF RuleBash 7 Limitations
CODEFUSION is not a global system as we only
consider natural language utterances in English.
Furthermore, natural language specifications can
be provided at varying levels of detail – utterances
with less detail may result in worse performance.
We consider various programming languages, but
more complex languages may result in worse per-
formance. We also find that CODEFUSION strug-
gles when tasked with generating longer code snip-
pets or programs that have long-ranged syntactic
dependencies. Because diffusion-based models de-
noise iteratively, inference latency is substantial,
rising exponentially with target generation length. | CODEFUSION |
Online Political Advertising in the United States
135
negative. Although the impact of internet ads was smaller than for television,
when one considers cost, the return on investment was just as high. More
research like this is needed to get a full assessment of the impact of digital
advertising, but other studies have found no impact. One experiment targeted
legislators’ constituents with Facebook ads but found no difference in
people’s evaluations or recognition of
the candidate shown in the ad
between those who saw the ad repeatedly and those who never saw it
(Broockman and Green 2014). Another experiment found no impact of
exposure to a week’s worth of online display ads on people’s views toward
the Black Lives Matter movement (Coppock and Broockman 2015). One
study did reveal a small impact of exposure to banner and pre-roll ads on
voter turnout
in a municipal election – but only when the race was
competitive (Haenschen and Jennings 2019). | Social_Media_and_Democracy |
energy consumption, memory usage, and processing power. Establishing these bench-
marks is essential for advancing the development of more resource-efficient LLMs, a
key priority given the increasing size and complexity of these models. | Beyond Efficiency |
inpainting module, PixelSynth is prone to generating incoher-
ent and blurry content, especially in the inpainted regions.
Moreover, as shown in Fig. 6, our Text2NeRF supports text-
driven scene generation in a large view range thanks to
our progressive scene inpainting and updating strategy. On
the other hand, other novel view synthesis methods produce
blurred scene-filling results even at a small viewing angle since
the text-related guidance is not considered in such methods. | Text2NeRF- Text-Driven 3D Scene Generation with Neural Radiance Fields |
LocalFoilTrees[55],orLoRE[56].Theseapproachesbuildsurrogate
models for each prediction sample, learning the reference model’s
behaviorontheparticularcaseofinterestbyintroducingperturbations
to the feature vector variables. By doing so, they can provide a local
featureimportanceestimate,whichisconsideredanindirectmethodto
explainamodel[32].Inparticular,weoptedforthefeaturerelevance
explanationaboveotherpost-hocmodelagnostictechniques(e.g.,rule
extraction) due to the simplicity of implementation. While other ex-
planationscanalsobebuiltconsideringtheontologyabstraction,care
mustbeputtoensuretheseremainvalid.E.g.,inthecaseofruleextrac-
tion,thenon-contradictionprinciplemustbeenforcedsincewhilethe
rulecanbevalidatafeaturelevel,thesameisnotguaranteedathigher
abstractionlevels.Whilesuchapproachesdonotconsiderthetemporal
dimensionperse,features’metadatacanbeconsideredtocaptureitand
enrichtheforecastexplanations[57].
When designing a system providing explanations for AI models, | Knowledge-graph-based-rich-and-confidentiality-preserving-Ex_2022_Informatio |
classification since BiomedGPT can fully understand it from the pretraining stage. However, addressing
task-aware instruction sensitivity remains an open question. | BiomedGPT |
more qualitative results that demonstrate how different amounts of safety data in training can change model
behavior in responding to adversarial and non-adversarial prompts. | Llama2 |
to violent clashes and undermining social cohesion (Izsak 2015). For example,
Facebook has come under fire for its role in mobilizing anti-Muslim mob
violence in Sri Lanka and for inciting violence against the Rohingya people in
Myanmar (Vindu, Kumar, and Frenkel 2018). Elucidating the mechanisms by
which exposure to hate speech drives intergroup tension, survey data and
experimental evidence from Poland suggest that frequent and repetitive
exposure to hate speech leads to desensitization to hateful content, lower
evaluations of populations targeted by hate speech, and greater distancing –
resulting in higher levels of anti–out-group prejudice (Soral et al. 2018). | Social_Media_and_Democracy |
being called in a loop or recursively, thus leading to a number of subtle bugs (see, for example,
the exploit of the DAO bug, leading to loss of $50 Millions). It is then reasonable to expect
that the best way to formally verify smart contracts is by using modular reasoning: for each
smart contract, the other contracts with which it interacts can be considered an environment.
This environment can be overapproximated using learning techniques in combination with
sampling and traditional model checking approaches. After verification of a single contract
passes successfully, some symbolic representations of the contracts with respect to the
correctness properties will be combined to prove correctness of the overall system.
The project will include a significant implementation component. The implementation is
done using the software verification platform developed at USI. The main development task | informatics-phd-projects-2022-23 |
QUESTION: Can you bring me something salty?
MODEL ANSWER (INCORRECT): Explanation: The user would like something salty. There are several
options in the objects list, so I will bring the user a lime soda. Plan: find(lime soda), pick(lime soda), find(user),
put(lime soda). (cid:55) (Feasible answers including finding chips. The model misidentify lime soda as salty.) | Chain-of-Thought Prompting Elicits Reasoning in Large Language Models |
Yuntao Bai, Saurav Kadavath, Sandipan Kundu, Amanda Askell, Jackson Kernion, Andy Jones,
Anna Chen, Anna Goldie, Azalia Mirhoseini, Cameron McKinnon, Carol Chen, Catherine Ols-
son, Christopher Olah, Danny Hernandez, Dawn Drain, Deep Ganguli, Dustin Li, Eli Tran-
Johnson, Ethan Perez, Jamie Kerr, Jared Mueller, Jeffrey Ladish, Joshua Landau, Kamal Ndousse,
Kamile Lukosuite, Liane Lovitt, Michael Sellitto, Nelson Elhage, Nicholas Schiefer, Noemi Mer-
cado, Nova DasSarma, Robert Lasenby, Robin Larson, Sam Ringer, Scott Johnston, Shauna
Kravec, Sheer El Showk, Stanislav Fort, Tamera Lanham, Timothy Telleen-Lawton, Tom Con-
erly, Tom Henighan, Tristan Hume, Samuel R. Bowman, Zac Hatfield-Dodds, Ben Mann, Dario
Amodei, Nicholas Joseph, Sam McCandlish, Tom Brown, and Jared Kaplan. Constitutional AI:
Harmlessness from AI Feedback, December 2022. URL http://arxiv.org/abs/2212.
08073. arXiv:2212.08073 [cs]. | CHAIN-OF-THOUGHTREASONING IS APOLICY IMPROVEMENTOPERATOR |
Alon Talmor, Jonathan Herzig, Nicholas Lourie, and
Jonathan Berant. 2019. CommonsenseQA: A ques-
tion answering challenge targeting commonsense
knowledge. In Proceedings of the 2019 Conference
of the North American Chapter of the Association for
Computational Linguistics: Human Language Tech-
nologies, Volume 1 (Long and Short Papers), pages
4149–4158, Minneapolis, Minnesota. Association for
Computational Linguistics.
Zheng Tang, Gus Hahn-Powell, and Mihai Surdeanu.
2020. Exploring interpretability in event extraction:
Multitask learning of a neural event classifier and
an explanation decoder. In Proceedings of the 58th
Annual Meeting of the Association for Computational
Linguistics: Student Research Workshop, pages 169–
175, Online. Association for Computational Linguis-
tics.
Ian Tenney, Dipanjan Das, and Ellie Pavlick. 2019.
In
BERT rediscovers the classical NLP pipeline. | Measuring Association Between Labels and Free-Text Rationales |
Human Evaluation. We also conduct human evaluation on the general quality of the model
responses on the combined test set described in subsection 3.1, which covers several existing
2The specific version of the data we used is https://huggingface.co/datasets/WizardLM/
WizardLM_evol_instruct_V2_196k/tree/main.
7
102103104Data Size010203040506070Win RateHumpbackWizardLLMAlpaca GPT-4Vicuna (sharegpt)OALIMAAlpacaFLAN v2Non-distillDistilled Figure 5: Scaling up self-curated instruction data A5 brings improvement in both small (7B) and
large (65B) LLaMa finetuned models, and neither model is saturated with 40,000 instructions.
Labelled Examples Win Rate %
3k
9k
1k
3k
Non-
distilled
65B
Non-
distilled
33B
Distilled
Proprietary
Humpback 65B
Guanaco 65B
LIMA 65B
Humpback 33B
OASST RLHF 33B 161k
9k
Guanaco 33B
161k
OASST SFT 33B
140k
Vicuna 33B
190k
WizardLLM 13B
airoboros 65B
17k
Falcon Instruct 40B 100k
GPT-4
Claude 2
ChatGPT
Claude | Self-AlignmentwithInstructionBacktranslation |
Shahaf Bassan, Yossi Adi, and Jeffrey S Rosenschein. Unsupervised symbolic music segmentation
using ensemble temporal prediction errors. arXiv preprint arXiv:2207.00760, 2022.
Adrien Ycart, Emmanouil Benetos, et al. A study on lstm networks for polyphonic music sequence
modelling. ISMIR, 2017.
Shulei Ji, Jing Luo, and Xinyu Yang. A comprehensive survey on deep music generation: Multi-level
representations, algorithms, evaluations, and future directions. arXiv preprint arXiv:2011.06801,
2020.
Prafulla Dhariwal, Heewoo Jun, Christine Payne, Jong Wook Kim, Alec Radford, and Ilya Sutskever.
Jukebox: A generative model for music. arXiv preprint arXiv:2005.00341, 2020.
Chuang Gan, Deng Huang, Peihao Chen, Joshua B Tenenbaum, and Antonio Torralba. Foley
music: Learning to generate music from videos. In Computer Vision–ECCV 2020: 16th European
Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part XI 16, pages 758–775. Springer,
2020. | Simple and Controllable Music Generation |
[69] Zehan Wang, Haifeng Huang, Yang Zhao, Ziang Zhang,
and Zhou Zhao. Chat-3D: Data-efficiently Tuning Large
Language Model for Universal Dialogue of 3D Scenes.
arXiv preprint arXiv:2308.08769, 2023. 2
[70] Ho-Hsiang Wu, Prem Seetharaman, Kundan Kumar, and
Juan Pablo Bello. Wav2CLIP: Learning Robust Audio
In ICASSP, pages 4563–
Representations From CLIP.
4567. IEEE, 2022. 4
[71] Shengqiong Wu, Hao Fei, Leigang Qu, Wei Ji, and Tat-
Seng Chua. NExT-GPT: Any-to-Any Multimodal LLM.
arXiv preprint arXiv:2309.05519, 2023. 2, 3, 5, 7
[72] Yusong Wu, Ke Chen, Tianyu Zhang, Yuchen Hui, Taylor
Berg-Kirkpatrick, and Shlomo Dubnov. Large-scale Con-
trastive Language-Audio Pretraining with Feature Fusion
and Keyword-to-Caption Augmentation. In IEEE Interna-
tional Conference on Acoustics, Speech and Signal Pro-
cessing, ICASSP, 2023. 7, 14 | M2UGen |
Lichang Chen, Shiyang Li, Jun Yan, Hai Wang, Kalpa Gunaratna, Vikas Yadav, Zheng Tang, Vijay
Srinivasan, Tianyi Zhou, Heng Huang, et al. Alpagasus: Training a better alpaca with fewer data.
arXiv preprint arXiv:2307.08701, 2023c.
Lingjiao Chen, Matei Zaharia, and James Zou. How is chatgpt’s behavior changing over time? arXiv
preprint arXiv:2307.09009, 2023d. | ChatGPT’sOne-yearAnniversary-AreOpen-Source LargeLanguageModelsCatchingup |
5.3 Real Speech Data
For real speech experiment, we used Common Voice 11. Natural (non-synthesized) monolingual
speech-text datasets both in English and Spanish were used for training. The evaluation of Spanish-
English real speech Translation was conducted using real speech with verified translation from
the CoVoST2 test set Wang et al. [2020], which is a subset of the Common Voice 11 test set.
However, it should be noted there is no English-Spanish CoVoST2 test set or otherwise any test
set from Common Voice 11 with verified translation for English-Spanish and therefore only an
English-Spanish evaluation was omitted from this experiment. The proposed approach achieved an
10.67 in BLEU for the task of Spanish-English Translation which is an improvement of +0.75 in
BLEU over the baseline which achieves 9.92 BLEU. Audio samples are available in our website:
https://google-research.github.io/lingvo-lab/translatotron3.
5.4 Comparison to Supervised Approaches | Translatotron3 |
Recently, linear shape models dominate the representation
of statistical 3D model. Numerous methods [5, 12, 34, 41]
have shown PCA’s ability in modeling the human body and
face. Inspired by [37], we parameterize our character shape
linearly with the following equation,
MS = FS(B) = ¯MS +
βisi,
(2)
where ¯MS denotes the mean shape and MS is the recon-
structed shape. The coefficients of linear shape are βi ∈ B.
|B| is the number of shape parameters and is set to 100 in
our implementation. si ∈ R3×N denotes the orthogonal prin-
cipal components of vertex displacements that capture shape
variations in different character identities. The shape model
of RaBit is learned from 1,050 characters of 3DBiCar using
PCA [37]. RaBit’s eyeballs can be computed based on the
predefined landmarks shown in Fig. 4. Please refer to the
Supplementary for more details.
4.2. Pose Modeling | RaBit- Parametric Modeling of 3D Biped Cartoon Characters with a Topological-consistent Dataset |
put consistent across skeleton formats through output regu-
larization (Fig. 3c). We also experiment with direct latent
point prediction, and a hybrid variant for the last step. | Learning 3D Human Pose Estimation from Dozens of Datasets using a Geometry-Aware Autoencoder to Bridge Between Skeleton Formats |
• Multilinguality: We use 10 benchmarks: XLSum (Non-English languages) (Hasan et al., 2021),
WMT22 (Kocmi et al., 2022), WMT23 (Tom et al., 2023), FRMT (Riley et al., 2023), WikiLingua
(Non-English languages) (Ladhak et al., 2020), TydiQA (no context), TydiQA (GoldP) (Clark
et al., 2020), MGSM (Shi et al., 2023), translated MMLU (Hendrycks et al., 2021a), NTREX
(Federmann et al., 2022), FLORES-200 (Team et al., 2022). | gemini_1_report |
As there are many details that need to be specified in the above sketch to yield a very concrete model,
this gives rise to a wide range of interesting mechanism design challenges. Different properties of the
market require different mechanisms, where one can think of e.g. a static "one-shot" trading
scenario versus a scenario where agents can dynamically enter and exit the market, or indivisible
versus divisible goods, shareable vs unshareable goods, etc. In this project we will work on trying to
solve various challenging variants of this design problem. | informatics-phd-projects-2022-23 |
In parallel with the move from direct to distributed discovery, we have seen
the move to a digital media environment that affords people with more
opportunities for more participatory forms of news and media use, in the
process also exposing many to widespread online harassment and potentially
various forms of disinformation disseminated online and especially via
platforms. Digital media offers everyone with internet access a range of
both “web 1.0” and “web 2.0” ways of engaging in more participatory
forms of news and media use, ranging from commenting on news sites and
sharing via email to commenting and/or sharing via social media sites. While
all internet users have access to this participatory potential, it is important to
3 Beyond the specific issues of news diversity that we examine here, algorithmic and automated
ranking systems can embed various forms of discrimination and reinforce oppressive social
relations as, for example, Noble (2018) shows. | Social_Media_and_Democracy |
40000DemonAttack17.515.012.510.0DoubleDunk0250500750Enduro100500FishingDerby0102030Freeway100200300Frostbite02000040000Gopher250500750Gravitar10864IceHockey0200400600Jamesbond0500010000Kangaroo2000400060008000Krull02000040000KungFuMaster050100MontezumaRevenge100020003000MsPacman25005000750010000NameThisGame1000Pitfall20020Pong0500PrivateEye050001000015000Qbert25005000750010000Riverraid02000040000RoadRunner246Robotank050010001500Seaquest5001000SpaceInvaders02000040000StarGunner201510Tennis30004000TimePilot0100200300Tutankham0100000200000UpNDown040MFrames0510Venture040MFrames50000100000150000VideoPinball040MFrames20004000WizardOfWor040MFrames0200040006000ZaxxonA2CACERPPOFigure6:ComparisonofPPOandA2Conall49ATARIgamesincludedinOpenAIGymatthetimeofpublication.11 A2CACERPPOAlien1141.71655.41850.3Amidar380.8827.6674.6Assault1562.94653.84971.9Asterix3176.36801.24532.5Asteroids1653.32389.32097.5Atlantis729265.31841376.02311815.0BankHeist1095.31177.51280.6BattleZone3080.08983.317366.7BeamRider30 | PPO |
and systematic process, enhancing the overall quality and fidelity of speech synthesis techniques. | AReviewofDeepLearningTechniquesforSpeechProcessing |
1Although there is a text preprocessing step in TTS systems,
We herein use preprocessed text interchangeably with the word
“text”.
Conditional Variational Autoencoder with Adversarial Learning for End-to-End Text-to-Speech | ConditionalVariationalAutoencoderwithAdversarialLearningfor End-to-EndText-to-Speech |
©2023 Cerebras Systems Inc. All Rights Reserved.
20
Cerebras-GPT: Open Compute-Optimal Language Models
1. HellaSwag is a dataset of multiple choice questions aimed to test a model’s common sense reasoning
abilities (Zellers et al., 2019). For example,
A woman is outside with a bucket and a dog. The dog is running around trying to avoid a
bath. She...
A. rinses the bucket off with soap and blow dry the dog’s head.
B. uses a hose to keep it from getting soapy.
C. gets the dog wet, then it runs away again.
D. gets into a bath tub with the dog.
The authors of the dataset adversely select examples such that they are difficult for language models
while still trivial for humans (with reported greater than 95% accuracy).
2. PIQA tests a model’s common sense reasoning about the physical world by posing a prompt and
two potential completions (Bisk et al., 2020). For example | Cerebras-GPT- Open Compute-Optimal Language Models Trained on the Cerebras Wafer-Scale Cluster |
We present qualitative results in Fig. 4, 5,6,7,8,9,10,11,12,13,14,15.
4.2 Qualitative Results
4.3 Ablation Study
Fig. 20 shows a comparison to a model trained without using ControlNet. That model is trained
with exactly same method with Stability’s Depth-to-Image model (Adding a channel to the SD and
continue the training).
Fig. 21 shows the training process. We would like to point out a “sudden convergence phenomenon”
where the model suddenly be able to follow the input conditions. This can happen during the training
process from 5000 to 10000 steps when using 1e-5 as the learning rate.
Fig. 22 shows Canny-edge-based ControlNets trained with different dataset scales.
4.4 Comparison to previous methods
Fig. 14 shows the comparison to Stability’s Depth-to-Image model.
Fig. 17 shows a comparison to PITI [59].
Fig. 18 shows a comparison to sketch-guided diffusion [58].
Fig. 19 shows a comparison to Taming transformer [11].
9
4.5 Comparison of pre-trained models | Adding Conditional Control to Text-to-Image Diffusion Models |
C.22 Enron Emails
To extract the data, we used the mailparser
package25 to extract the body of each email as a
document.
D General Data Processing
This section discusses any processes applied across
multiple datasets.
To combine the constituent datasets, we iterate
until the size of the output dataset is the desired
size, drawing documents from datasets at random,
weighted by the number of documents in each
dataset times the number of epochs desired on
that dataset. Because the number of documents
involved is high, by the law of large numbers, the
number of copies of each dataset present in the Pile
is approximately equal to its epoch count.
Shuffling a dataset posed a major problem due to
our limited memory and computational budget. We
follow Hardin (2018), a method descended from
Rao (1961), and interleave our output to produce
30 output piles. | The Pile- An 800GB Dataset of Diverse Text for Language Modeling |
Diversity. 3DBiCar spans a wide range of 3D biped car-
toon characters, containing 1,500 high-quality 3D models.
First, we carefully collect images of 2D full-body biped car-
toon characters with diverse identities, shape, and textural
styles from the Internet, resulting in 15 character species and
4 image styles, as shown in Fig. 3. Then we recruit six pro-
fessional artists to create 3D corresponding character models
according to the collected reference images. The modeling
result is required to be matched with the reference images
as much as possible. The representative image-model pairs
sampled from our dataset are shown in Fig. 2.
Topological-consistency. The key to building a linear
3 | RaBit- Parametric Modeling of 3D Biped Cartoon Characters with a Topological-consistent Dataset |
Last year you heard us talk about PaLM, which led to many improvements across our products. Today, we’re ready
to announce our latest PaLM model in production: PaLM 2.
PaLM 2 builds on our fundamental research and our latest infrastructure. It’s highly capable at a wide range of
tasks and easy to deploy. We are announcing more than 25 products and features powered by PaLM 2 today.
PaLM 2 models deliver excellent foundational capabilities across a wide range of sizes. We’ve affectionately
named them Gecko, Otter, Bison, and Unicorn. Gecko is so lightweight that it can work on mobile devices: fast
enough for great interactive applications on-device, even when offline. PaLM 2 models are stronger in logic and
reasoning thanks to broad training on scientific and mathematical topics. It’s also trained on multilingual text —
spanning more than 100 languages — so it understands and generates nuanced results. | Google I_O 2023_ Making AI more helpful for everyone |
Computational Resources
Requires computational resources to support
retrieval strategies and technologies related
to databases. External data source integration
and updates need to be maintained.
Preparation and curation of high-quality
training datasets, definition of fine-tuning
objectives, and provision of corresponding
computational resources are necessary.
Latency Requirements
Involves data retrieval, potentially leading to
higher latency.
LLM after fine-tuning can respond without
retrieval, resulting in lower latency.
Reducing Hallucinations
Ethical and Privacy Issues
Inherently less prone to hallucinations as
each answer is grounded in retrieved evi-
dence.
Ethical and privacy concerns arise from
storing and retrieving text from external
databases.
Can help reduce hallucinations by training
the model based on specific domain data but
may still exhibit hallucinations when faced
with unfamiliar input. | Retrieval-AugmentedGenerationforLargeLanguageModels-ASurvey |
We relied on web traffic vs. app traffic to “qualify” companies for the list, as most consumer
GenAI products have been website-first so far (more on this below!). For companies that made
the list that do have a mobile app, we added that traffic, gathered from Sensor Tower as of June
2023, to determine their spot number. Thus, this ranking serves as a tool to identify and
understand category trends, and not as an exhaustive ranking of all consumer AI
platforms.
Here are our top 6 takeaways.
https://a16z.com/how-are-consumers-using-generative-ai/
1/13
19/09/2023, 13:27
How Are Consumers Using Generative AI? | Andreessen Horowitz
TA B L E O F C O N T E N T S | How Are Consumers Using Generative AI_ _ Andreessen Horowitz |
The second line of each test case contains n integers
a_1 , a_2 , ... , a_n (1 <= a_i <= 10^6) .
The second line of each test case contains n integers
a_1 , a_2 , ... , a_n (1 <= a_i <= 10^6) .
It is guaranteed that the sum of n over all test
cases doesn ’t exceed 3 . 10^5.
It is guaranteed that the sum of n over all test
cases doesn ’t exceed 3 . 10^5.
For each test case , print a single integer -- the
maximum possible value of the product from the
statement .
For each test case , print a single integer -- the
minimum possible value of the product from the
statement .
4
3
2 4 3
4
3 2 3 1
2
69 69
6
719313 273225 402638 473783 804745 323328
4
3
2 4 3
4
3 2 3 1
2
69 69
6
719313 273225 402638 473783 804745 323328
Output
Example
Input
Output
12
6
4761
381274500335
Note
Output
Example
Input
Output
8
3
4761
88341292800
Note
Let f(l) =a_l . a_{l+1}
In the first test case ,
Let f(l) =a_l . a_{l+1}
In the first test case ,
* f(1) =
a_1 . a_2
* f(2) = a_2 . a_3 | alphacode |
existing democracies has in fact lived up to the various ideals we might have for
journalism and for democracy. Yet with its many imperfections, at least in
North America and Western Europe, empirical research suggests that
independent, professionally produced news has helped inform the public,
helped people make sense of the world through analysis, interpretation, and
the portrayal of contending forces, and helped members of the public connect
with one another to see themselves as part of a community and act in concert to
influence public affairs (van Zoonen 1998; Curran et al. 2009; Couldry et al.
2010). Beyond these, perhaps the most visible democratic role of news,
investigative journalism specifically can also produce a range of “positive
externalities” that benefit the whole public – even those who do not actually
engage with a particular story – by ensuring more efficient local government,
reducing corruption, and increasing how responsive elected officials are to their | Social_Media_and_Democracy |
and better control the generated music. Note that music in
SymMV is of high quality and can also be directly used for
unconditional music generation without video modality. | VideoBackgroundMusicGeneration |
Fabio Petroni, Aleksandra Piktus, Angela Fan, Patrick
Lewis, Majid Yazdani, Nicola De Cao, James
Thorne, Yacine Jernite, Vladimir Karpukhin, Jean
Maillard, Vassilis Plachouras, Tim Rocktäschel, and
Sebastian Riedel. 2021. KILT: a benchmark for
knowledge intensive language tasks. In Proceedings
of the 2021 Conference of the North American Chap-
ter of the Association for Computational Linguistics:
Human Language Technologies, pages 2523–2544,
Online. Association for Computational Linguistics.
Fabio Petroni, Tim Rocktäschel, Sebastian Riedel,
Patrick Lewis, Anton Bakhtin, Yuxiang Wu, and
Alexander Miller. 2019. Language models as knowl-
In Proceedings of the 2019 Confer-
edge bases?
ence on Empirical Methods in Natural Language
Processing and the 9th International Joint Confer-
ence on Natural Language Processing (EMNLP-
IJCNLP), pages 2463–2473, Hong Kong, China. As-
sociation for Computational Linguistics. | Toolformer |
Limitations. For gender-related errors in translation systems, evaluations do not consider differential harms to people
related to expressing non-binary gender identities (Keyes, 2018; Dev et al., 2021a), or consider contested perspectives
on pronouns across languages and cultures (Lee, 2019). Moreover, while gender agreement into English is amenable to
automatic evaluation, evaluation of gender agreement out of English remains challenging and time-intensive. Finally,
we note that our evaluations focus on only a subset of potential risks (Weidinger et al., 2021), and that our evaluations
focus on model outputs without considering the wider sociotechnical context in which instruction-finetuned language
models exist (Shelby et al., 2023). See Appendix E.8 for measurement quality rubric for this evaluation when translating
into English.
75 | PaLM 2 Technical Report |
planks and 2 sticks on crafting table
return "wooden_axe"
User: [Description] I succeed in step 1, 2, 3, 4, 5.
I finish all steps and I obtain 1 wooden_axe successfully.
==========
User: My current inventory has <inventory>. <visual observation>. How to obtain 1 stone_sword in Minecraft step-by-
step?
Assistant:
Prompt 1: Planning prompt in JARVIS-1
System: Extract the action name, action type, goal object, tool and action rank from the input text.
User: mine({"log":3}, null); # step 1: chop down trees to mine logs
Assistant:
name: mine_log
text condition: chop down trees to mine logs
action: mine
object_item: log
object_number: 3
tool: null
rank: 1
###
input: craft({"planks":12}, {"log":3}, null); # step 2: craft 12 planks from 3 log
Assistant:
name: craft_planks
condition: craft 12 planks from 3 log
16
JARVIS-1: Open-World Multi-task Agents with Memory-Augmented Multimodal Language Models | JARVIS-1 |
B DPO Implementation Details and Hyperparameters
DPO is relatively straightforward to implement; PyTorch code for the DPO loss is provided below:
19
import torch.nn.functional as F
def dpo_loss(pi_logps, ref_logps, yw_idxs, yl_idxs, beta):
"""
pi_logps: policy logprobs, shape (B,)
ref_logps: reference model logprobs, shape (B,)
yw_idxs: preferred completion indices in [0, B-1], shape (T,)
yl_idxs: dispreferred completion indices in [0, B-1], shape (T,)
beta: temperature controlling strength of KL penalty
Each pair of (yw_idxs[i], yl_idxs[i]) represents the
indices of a single preference pair.
"""
pi_yw_logps, pi_yl_logps =
pi_logps[yl_idxs]
ref_yw_logps, ref_yl_logps = ref_logps[yw_idxs], ref_logps[yl_idxs]
pi_logps[yw_idxs],
pi_logratios = pi_yw_logps - pi_yl_logps
ref_logratios = ref_yw_logps - ref_yl_logps
losses = -F.logsigmoid(beta * (pi_logratios - ref_logratios))
rewards = beta * (pi_logps - ref_logps).detach()
return losses, rewards | Direct Preference Optimization |
driving scenario, an LLM selectively activates the required
neural modules by invoking specific functions from the tool
library, ensuring the collection of necessary environmental
information with less redundancy. Upon gathering the
necessary environmental information, the LLM leverages
this data as a query to search in a cognitive memory
for pertinent traffic regulations and the most similar past
driving experience. Finally, the retrieved traffic rules and
driving experience, together with the formerly collected
environmental information, are utilized as inputs to an LLM-
based reasoning engine. The reasoning engine performs
multi-round reasoning based on the inputs and eventually
devises a safe and comfortable trajectory for driving. Our
Agent-Driver architecture harnesses dynamic perception and
prediction capability brought by the tool library, human
knowledge from the cognitive memory, and the strong
decision-making ability of the reasoning engine. This | ALanguageAgentforAutonomousDriving |
26
Same prompt:“room”+ default “a detailed high-quality professional image”Same CFG scale (9.0)Canny EdgeHEDLine (M-LSD)Depth (midas)Normal (from midas)Scribbles (synthesized)Source Image Figure 24: (Continued) Comparison of six detection types and the corresponding results. The scribble map is extracted from the HED map
with morphological transforms.
27
Same prompt:“robotics”+ default “a detailed high-quality professional image”Same CFG scale (9.0)Canny EdgeHEDLine (M-LSD)Depth (midas)Normal (from midas)Scribbles (synthesized)Source Image Figure 25: (Continued) Comparison of six detection types and the corresponding results. The scribble map is extracted from the HED map
with morphological transforms.
28 | Adding Conditional Control to Text-to-Image Diffusion Models |
we combine them for the AHWC2S model with input:
= [¯ai,1, . . . , ¯ai,A, hi, wi, cci, cwi , chi].
xAHWC2S
(7)
i
In practice, depending on which measurements are avail-
able, we train and use different regressors. Following the
naming convention of AHWC2S, these models are: AH2S,
AHW2S, AC2S, and AHC2S, as well as their equivalents
without attribute input H2S, HW2S, C2S, and HC2S. For an
evaluation of the contribution of linguistic shape attributes
on top of each anthropometric measurement, see Sup. Mat.
Training Data: To train the A2S and S2A mappings we
use CAESAR data, for which we have SMPL-X shape pa-
rameters, anthropometric measurements, and linguistic at-
tribute scores. We train separate gender-specific models. | Accurate 3D Body Shape Regression using Metric and Semantic Attributes |
This paper provides an investigation of antecedents and consequences of AI’s placebo effect
in HCI. In an experimental study (𝑁 = 65), we examined the influence of negative and positive
verbal AI descriptions. We analyzed the impact of expectations on decision-making in a letter
discrimination task, with or without a sham-AI system.
There are three major shortcomings in the placebo literature in HCI for AI. First, direct effects
on a behavioral level are yet to be found [40, 78]. Second, it is unclear whether nocebo effects,
low expectations impairing behavior, are equally influential as positive expectations based on
verbal descriptions in HCI. Third, we lack a behavioral marker for effectively designing adaptive
AI interfaces that enhance decision-making amidst placebo responses.
Unpublishedworkingdraft.
Notfordistribution. | AI enhance sour performance |
Memory-based Architectures Our document index can be seen as a large external memory for
neural networks to attend to, analogous to memory networks [64, 55]. Concurrent work [14] learns
to retrieve a trained embedding for each entity in the input, rather than to retrieve raw text as in our
work. Other work improves the ability of dialog models to generate factual text by attending over
fact embeddings [15, 13]. A key feature of our memory is that it is comprised of raw text rather
distributed representations, which makes the memory both (i) human-readable, lending a form of
interpretability to our model, and (ii) human-writable, enabling us to dynamically update the model’s
memory by editing the document index. This approach has also been used in knowledge-intensive
dialog, where generators have been conditioned on retrieved text directly, albeit obtained via TF-IDF
rather than end-to-end learnt retrieval [9]. | Retrieval-AugmentedGenerationfor Knowledge-IntensiveNLPTasks |
• Section 4 LLM pre-training: This section explores the various pre-training tech-
niques for LLMs, highlighting how they contribute to resource efficiency. Key
areas such as memory efficiency, data efficiency, and innovative training pipeline
designs are examined, illustrating how each technique impacts the overall resource
utilization during the pre-training phase.
• Section 5 LLM fine-tuning: This section covers the fine-tuning phase of LLMs,
focusing on methods that enhance resource efficiency. It includes detailed discus-
sions on parameter-efficient fine-tuning, which minimizes parameter updates; and
full-parameter fine-tuning, which optimizes the entire parameter set. | Beyond Efficiency |
CNN (32,36)
LSTM (40) DNN (41)
NLP (37,44)
CNN (46) DNN (47)
NB (19) DT (9,20)
DNN [33]
GNN (42) LSTM (40,43)
NLP (38,45)
RL (48) Clustering (23)
DT (50) RNN (28)
GNN (52) FM (53)
ResNet (54)
XGBoost (ensemble) [55]
CNN, RNN [22]
GNN (42) DT (57)
NN (56)
GRU [18]
NLP (24,45)
DT (37)
DRL [30]
DNN (47)
LSTM [29]
CNN (29,36,39)
DNN (35,58) | Knowledge-graph-based explainable AI- A systematic review |
speech of real-world internet users for unjustified removal.
51 Article 1. 2018. Joint Letter on European Commission regulation on online terrorist content.
www.article19.org/resources/joint-letter-on-european-commission-regulation-on-online-ter-
rorist-content/; Reda (2017).
https://doi.org/10.1017/9781108890960 Published online by Cambridge University Press
240
Daphne Keller & Paddy Leerssen
Under a broader view of content moderation, platforms also shape discourse
through the design of their ranking and recommender algorithms, such as
Facebook’s News Feed and YouTube’s Recommended videos (Keller 2019b).
A growing body of literature in computer science and communications science
seeks to ascertain the operation and effects of these complex systems.52 The
design of these algorithms is currently unregulated, but several governments
have recently proposed to do so.53 Most of these initiatives also explicitly
demand greater transparency in algorithmic recommendations.54 | Social_Media_and_Democracy |
large webtext corpora: A case study on the colossal clean crawled corpus, 2021.
Du, N., Huang, Y., Dai, A. M., Tong, S., Lepikhin, D., Xu, Y., Krikun, M., Zhou, Y., Yu, A. W., Firat, O., Zoph, B.,
Fedus, L., Bosma, M., Zhou, Z., Wang, T., Wang, Y. E., Webster, K., Pellat, M., Robinson, K., Meier-Hellstern,
K., Duke, T., Dixon, L., Zhang, K., Le, Q. V., Wu, Y., Chen, Z., and Cui, C. GLaM: Efficient Scaling of Language
Models with Mixture-of-Experts. ICML, 2022. URL https://arxiv.org/abs/2112.06905. | PaLM 2 Technical Report |
data. Organize your data based on your research questions and hypothesis.
4. Display your data based on relationships among the collected data and look for supporting evidence.
5. Cross check your data few times for reliability and validity.
6. So, what did you find from your experimentation? Report without adding any comments of your own.
7. What were the differences? If you are making a comparison. Use T-Test to compare.
8. Analyze your findings to see if it answers your research questions and finds a solution to your problem
statement. Again, avoid making any comments of your own.
Save your energy for the conclusion and discussion chapter. Do not forget to report your results in the
present form because it sounds soothing and original. Example: The interviews indicate that…….result
shows that..
How to Write Your Conclusion and Discussion | How to Write Your PhD Proposal- A Step-By-Step Guide |
Identity management Discrepancy and misalignment between resources of different knowledge graphs is a persistent issue
in current KBX-systems. Managing identities is a prerogative for knowledge-based explainable systems to efficiently use
the available information and avoid undesirable, wide-ranging effects. While a number of principles exist for publishing
and linking resources, a common agreement on what constitutes identical entities is still an open challenge. This also
affect the wide-spread adoption of knowledge graphs in eXplainable AI, that cannot tolerated uncertainty over data quality.
Solutions to this problems, partly investigated [96], could be services to help data modellers and applications to identify
same entities in the real world; better guidelines to correctly use the different types of identity links (e.g. owl:sameAs, | Knowledge graphs as tools for explainable machine learning: A survey |
35
It is not easy to realize the above ideas, though. We’ve mentioned the safety issues of accessing physical
tools, and this is also one main challenge for scientific tool learning since many scientific problems need to
be verified in actual situations, and this process may bring danger if decided by AIs. Meanwhile, foundation
models are generally trained with natural language corpus or natural images, while scientific data are usually
more heterogeneous, numerical, and structured. It is worth exploring how to fuse the general intelligence
learned from plain text and the expertise needed for scientific discovery. Recently, Boiko et al. (2023) show
the potential of this direction and build a system that uses foundation models to design, plan, and execute
scientific experiments (e.g., catalyzed cross-coupling reactions).
6 Conclusion | Tool Learning with Foundation Models |
Gemini models are also capable of operating across modalities and a diverse set of global languages
simultaneously, both for image understanding tasks (e.g., images containing text in Icelandic) and for
generation tasks (e.g., generating image descriptions for a wide range of languages). We evaluate the
performance of generating image descriptions on a selected subset of languages in the Crossmodal-
3600 (XM-3600) benchmark in a 4-shot setting, using the Flamingo evaluation protocol (Alayrac
et al., 2022), without any fine-tuning for all models. As shown in Table 9, Gemini models achieve a
significant improvement over the existing best model, Google PaLI-X.
XM-3600 (CIDER)
English
French
Hindi
Modern Hebrew
Romanian
Thai
Chinese
Average (of 7)
Gemini Ultra
4-shot
86.4
77.9
31.1
54.5
39.0
86.7
33.3
58.4
Gemini Pro
4-shot
87.1
76.7
29.8
52.6
37.7
77.0
30.2
55.9
Google PaLI-X
4-shot
77.8
62.5
22.2
38.7
30.2
56.0
27.7
45.0 | gemini_1_report |
59.0
59.5
5.4 Results on MTEB benchmark
In Table 3, E5 models not only substantially outperform existing ones with similar sizes, but also
match the results of much larger models. The top-2 models on MTEB leaderboard 7 GTRxxl and
Sentence-T5xxl have 4.8B parameters, while our E5large model is more than 10× smaller with 300M
parameters. We expect that our model will benefit from continual scaling up.
Since the difference between BERT-FTbase and E5base is that BERT-FTbase only has fine-tuning
stage, their performance gap demonstrates the usefulness of contrastive pre-training on our proposed
CCPairs dataset. For most task categories except Clustering, performance improves after supervised
fine-tuning. Consistent with prior works [43, 44], this once again demonstrates the importance of
incorporating human knowledge for learning better text embeddings. It remains an open question
whether state-of-the-art embeddings can be obtained in a purely self-supervised manner. | E5 |
24
The Efficiency Spectrum of Large Language Models: An Algorithmic Survey
Efficient LLM Algorithmic Survey, Nov, 2023, USA.
REFERENCES
[1] [n. d.]. Introducing ChatGPT. https://openai.com/blog/chatgpt
[2] [n. d.]. Introducing Claude 2.1. https://www.anthropic.com/index/claude-2-1
[3] [n. d.]. Introducing PyTorch Fully Sharded Data Parallel (FSDP) API. https://pytorch.org/blog/introducing-pytorch-fully-sharded-data-parallel-api/
[4] [n. d.]. Pathways Language Model (PaLM): Scaling to 540 Billion Parameters for Breakthrough Performance. https://blog.research.google/2022/04/pathways-
language-model-palm-scaling-to.html
[5] [n. d.]. Planning for AGI and beyond. https://openai.com/blog/planning-for-agi-and-beyond
[6] Amey Agrawal, Ashish Panwar, Jayashree Mohan, Nipun Kwatra, Bhargav S Gulavani, and Ramachandran Ramjee. 2023. SARATHI: Efficient LLM Inference
by Piggybacking Decodes with Chunked Prefills. arXiv preprint arXiv:2308.16369 (2023). | TheEfficiencySpectrumofLargeLanguageModels-AnAlgorithmicSurvey |
• The voting ties constitute a notable portion to the selection differences between USC and SC,
especially with 8 candidate responses. Specifically, among all responses with the maximum
votes, SC always selects the one with the smallest index, while USC can pick up alternative
ones based on the response format.
• The match ratio between USC and SC consistently surpasses their own task accuracies,
which shows that the consistency criterion is easier to measure than the answer correctness.
• Shifting from 8 to 16 samples, the USC-SC match ratio reduces, suggesting that USC
behaves as an imperfect approximation of SC. However, the difference in response selection
does not always lead to the performance decrease, as USC sometimes selects the correct
response when SC fails.
5 RELATED WORK | UNIVERSALSELF-CONSISTENCYFORLARGELANGUAGEMODELGENERATION |
Among the 328 prompts we evaluated, Claude 2 gave a response judged more harmful than “I can’t help you
with that" in four cases, according to automated evaluation. On manual inspection, in three of the cases its
response did not seem harmful. However, in the other case, the model was disrupted by the jailbreak attempts
in about half of its sampled responses.
3.5 Helpful, Honest, and Harmless (HHH) Evaluations
Anthropic researchers wrote 438 binary choice questions [2, 3, 9] to evaluate language models and preference
models on their ability to identify HHH responses. The model is presented with two outputs and asked to
select the more HHH output. We see in Figure 6 that each of our Claude models is better than the last at this
task 0-shot, showing general improvements in "understanding" helpfulness, honesty, and harmlessness [8].
6 | ClaudeModels |
∗Equal contribution.
1
Universal Self-Consistency for Large Language Model Generation | UNIVERSALSELF-CONSISTENCYFORLARGELANGUAGEMODELGENERATION |
Modeling in Meeting Recognition.. In Interspeech, Vol. 11. 2877–2880.
[90] Tiffany H Kung, Morgan Cheatham, Arielle Medenilla, Czarina Sillos, Lorie De Leon, Camille Elepaño, Maria Madriaga,
Rimel Aggabao, Giezel Diaz-Candido, James Maningo, et al. 2023. Performance of ChatGPT on USMLE: Potential for
AI-assisted medical education using large language models. PLoS digital health 2, 2 (2023), e0000198.
[91] Tom Kwiatkowski, Jennimaria Palomaki, Olivia Redfield, Michael Collins, Ankur Parikh, Chris Alberti, Danielle
Epstein, Illia Polosukhin, Matthew Kelcey, Jacob Devlin, Kenton Lee, Kristina N. Toutanova, Llion Jones, Ming-Wei
Chang, Andrew Dai, Jakob Uszkoreit, Quoc Le, and Slav Petrov. 2019. Natural Questions: a Benchmark for Question
Answering Research. Transactions of the Association of Computational Linguistics (2019). | ASurveyonEvaluationofLargeLanguageModels |
Privacy Preserving Technologies. Personalized tool learning requires models to learn user preferences
from private user information, which inevitably raises privacy-preserving concerns. On the one hand, previous
work has shown that training data extraction attacks can be applied to recover sensitive personal privacy
from foundation models (Carlini et al., 2021), which is a critical challenge for personalized tool learning.
On the other hand, models with high computational costs must be deployed on cloud servers, which require
uploading private data to the cloud to enable personalized responses. It is crucial to develop secure and
trustworthy mechanisms to access and process user data while protecting user privacy. Addressing these
challenges will help unlock the potential of personalized tool learning, enabling more effective and tailored
tool manipulation to meet individual user needs. To this end, it is worth exploring model-oriented distributed | Tool Learning with Foundation Models |
Appendices
The appendix presents supplementary details that extend
beyond the content of the manuscript, aiming to enhance
comprehension of the M2UGen model. Comprehensive
information is provided concerning the model’s training
dataset and training methodology, encompassing explicit
insights into the utilized training approach and the cor-
responding model hyperparameters. Additionally, a thor-
ough exposition is given regarding the composition of the
evaluation sets employed in our study, accompanied by a
delineation of the evaluation methodology and metrics ap-
plied to assess the performance of our model. To elucidate
the diverse capabilities of our model, illustrative demo ex-
amples are also included.
A Music Oriented Dataset Information
We generate 4 different datasets to train the M2UGen
model: MUCaps, MUImage, MUVideo and MUEdit
datasets. The statistics of the datasets are given in Table
7. An example of each from the 4 datasets are shown in
Figure 4. | M2UGen |
not able to "cheat" the mechanism for their own benefit. Moreover, these mechanisms should perform
their computations reasonably (and provably) fast. How to design the trading mechanism in such a
way that these requirements are satisfied? | informatics-phd-projects-2022-23 |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.