Datasets:

yxzwayne
/

abs

Dataset card Data Studio Files Files and versions Community

Split (1)

train · 587 rows

filename stringlengths 9 127	text stringlengths 133 11k
2310.15154.pdf	Pre-publication draft LINEAR REPRESENTATIONS OF SENTIMENT INLARGE LANGUAGE MODELS Curt Tigges♣, Oskar John Hollinsworth♡, Atticus Geiger♠⋆, Neel Nanda♢ ♣EleutherAI Institute,♡SERI MATS,♠Stanford University,⋆Pr(Ai)2R Group,♢Independent *Equal primary authors (order random) ABSTRACT Sentiment is a pervasive feature in natural language text, yet it is an open question how sentiment is represented within Large Language Models (LLMs). In this study, we reveal that across a range of models, sentiment is represented linearly: a single direction in activation space mostly captures the feature across a range of tasks with one extreme for positive and the other for negative. Through causal interventions, we isolate this direction and show it is causally relevant in both toy tasks and real world datasets such as Stanford Sentiment Treebank. Through this case study we model a thorough investigation of what a single direction means on a broad data distribution. We further uncover the mechanisms that involve this direction, highlighting the roles of a small subset of attention heads and neurons. Finally, we discover a phenomenon which we term the summarization motif: sentiment is not solely represented on emotionally charged words, but is additionally summarised at in- termediate positions without inherent sentiment, such as punctuation and names. We show that in Stanford Sentiment Treebank zero-shot classification, 76% of above-chance classification accuracy is lost when ablating the sentiment direction, nearly half of which (36%) is due to ablating the summarized sentiment direction exclusively at comma positions. 1 I NTRODUCTION Large language models (LLMs) have displayed increasingly impressive capabilities (Brown et al., 2020; Radford et al., 2019; Bubeck et al., 2023), but their internal workings remain poorly un- derstood. Nevertheless, recent evidence (Li et al., 2023) has suggested that LLMs are capable of forming models of the world, i.e., inferring hidden variables of the data generation process rather than simply modeling surface word co-occurrence statistics. There is significant interest (Christiano et al. (2021), Burns et al. (2022)) in deciphering the latent structure of such representations. In this work, we investigate how LLMs represent sentiment, a variable in the data generation process that is relevant and interesting across a wide variety of language tasks (Cui et al., 2023). Approach- ing our investigations through the frame of causal mediation analysis (Vig et al., 2020; Pearl, 2022; Geiger et al., 2023a), we show that these sentiment features are represented linearly by the models, are causally significant, and are utilized by human-interpretable circuits (Olah et al., 2020; Elhage et al., 2021a). We find the existence of a single direction scientifically interesting as further evidence for the linear representation hypothesis (Mikolov et al., 2013; Elhage et al., 2022)– that models tend to extract properties of the input and internally represent them as directions in activation space. Understanding the structure of internal representations is crucial to begin to decode them, and linear representations are particularly amenable to detailed reverse-engineering (Nanda et al., 2023b). We show evidence of a phenomenon we have labeled the “summarization motif”, where rather than sentiment being directly moved from valenced tokens to the final token, it is first aggregated on intermediate summarization tokens without inherent valence such as commas, periods and particular nouns.1This summarization structure for next token prediction can be seen as a naturally emerging 1Our use of the term “summarization” is distinct from typical NLP summarization tasks 1arXiv:2310.15154v1 [cs.LG] 23 Oct 2023
2212.10559.pdf	Why Can GPT Learn In-Context? Language Models Secretly Perform Gradient Descent as Meta-Optimizers Damai Dai†∗, Yutao Sun∥∗, Li Dong‡, Yaru Hao‡, Zhifang Sui†, Furu Wei‡ †Peking University∥Tsinghua University ‡Microsoft Research https://github.com/microsoft/LMOps Abstract Large pretrained language models have shown surprising In-Context Learning (ICL) ability. With a few demonstration input-label pairs, they can predict the label for an unseen input without additional parameter updates. Despite the great success in performance, the work- ing mechanism of ICL still remains an open problem. In order to better understand how ICL works, this paper explains language mod- els as meta-optimizers and understands ICL as a kind of implicit ﬁnetuning. Theoreti- cally, we ﬁgure out that the Transformer at- tention has a dual form of gradient descent based optimization. On top of it, we under- stand ICL as follows: GPT ﬁrst produces meta- gradients according to the demonstration ex- amples, and then these meta-gradients are ap- plied to the original GPT to build an ICL model. Experimentally, we comprehensively compare the behavior of ICL and explicit ﬁne- tuning based on real tasks to provide empiri- cal evidence that supports our understanding. The results prove that ICL behaves similarly to explicit ﬁnetuning at the prediction level, the representation level, and the attention be- havior level. Further, inspired by our under- standing of meta-optimization, we design a momentum-based attention by analogy with the momentum-based gradient descent algo- rithm. Its consistently better performance over vanilla attention supports our understanding again from another aspect, and more impor- tantly, it shows the potential to utilize our un- derstanding for future model designing. 1 Introduction In recent years, large pretrained language models, especially in Transformer-based architectures (e.g., GPT; Brown et al. 2020), have shown strong emer- gent In-Context Learning (ICL) ability. Different from ﬁnetuning which needs additional parame- ter updates, ICL just needs several demonstration ∗Contribution during internship at Microsoft Research. (Sentence, ?)… Demonstration Examples Query ExampleFeed -Forward Network Self-Attention (Sentence1, Answer1) (Sentence2, Answer2)Meta -GradientsAnswer … GPTIn-Context LearningFinetuning GPT (Sentence1, Answer1) GPT (Sentence2, Answer2) Back-Propagation Forward Computation𝚫𝚫𝑾𝑾𝐅𝐅𝐅𝐅 𝚫𝚫𝑾𝑾𝐈𝐈𝐈𝐈𝐈𝐈Gradients Dual ViewFigure 1: According to the demonstration examples, GPT produces meta-gradients for In-Context Learn- ing (ICL) through forward computation. ICL works by applying these meta-gradients to the model through at- tention. The meta-optimization process of ICL shares a dual view with ﬁnetuning that explicitly updates the model parameters with back-propagated gradients. examples prepended before the original input, and then the model can predict the label for even unseen inputs. On numerous downstream tasks, a large GPT model can achieve a quite great performance, which even exceeds some smaller models with su- pervised ﬁnetuning. However, although ICL has achieved great success in performance, the work- ing mechanism of it is still an open problem to be investigated. In this paper, we explain ICL as a process of meta-optimization and attempt to build connections between GPT-based ICL and ﬁnetuning. Concen- trating on the attention modules, we ﬁgure out that the Transformer attention has a dual form of gra- dient descent based optimization. On top of it, wearXiv:2212.10559v2 [cs.CL] 21 Dec 2022
2306.00297.pdf	Transformers learn to implement preconditioned gradient descent for in-context learning Kwangjun Ahn1,3,, Xiang Cheng1,3,, Hadi Daneshmand2,3,, and Suvrit Sra1,3 1Department of Electrical Engineering and Computer Science, MIT 2Foundations of Data Science Institute (FODSI) 3Laboratory for Information and Decision Systems (LIDS), MIT June 2, 2023 Abstract Motivated by the striking ability of transformers for in-context learning, several works demonstrate that transformers can implement algorithms like gradient descent. By a careful construction of weights, these works show that multiple layers of transformers are expressive enough to simulate gradient descent iterations. Going beyond the question of expressivity, we ask: Can transformers learn to implement such algorithms by training over random problem instances? To our knowledge, we make the first theoretical progress toward this question via analysis of the loss landscape for linear transformers trained over random instances of linear regression. For a single attention layer, we prove the global minimum of the training objective implements a single iteration of preconditioned gradient descent. Notably, the preconditioning matrix not only adapts to the input distribution but also to the variance induced by data inadequacy. For a transformer with kattention layers, we prove certain critical points of the training objective implement kiterations of preconditioned gradient descent. Our results call for future theoretical studies on learning algorithms by training transformers. 1 Introduction In-context learning (ICL) is the striking capability of large language models to adapt to structure in inputs: given a prompt containing examples and a query, the transformer produces the correct output based on the context provided by the examples, without needing to update its parameters [Brown et al., 2020, Lieber et al., 2021, Rae et al., 2021, Black et al., 2022]. This property has become the focus of body of recent research that aims to shed light on the underlying mechanism of large language models [Garg et al., 2022, Akyürek et al., 2022, von Oswald et al., 2022, Li and Malik, 2017, Min et al., 2021, Xie et al., 2021, Elhage et al., 2021, Olsson et al., 2022]. A line of research studies ICL via the expressive power of transformers. Transformer architectures are powerful Turing machines, capable of implementing various algorithms [Pérez et al., 2021, Wei et al., 2022]. Given an in-context prompt, Edelman et al. [2022], Olsson et al. [2022] argue that transformers are able to implement algorithms through the recurrence of multi-head attentions to extract coarse information from raw input prompts. Akyürek et al. [2022], von Oswald et al. [2022] assert that transformers can implement gradient descent on linear regression encoded in a given input prompt. Equal contribution, alphabetical order. 1arXiv:2306.00297v1 [cs.LG] 1 Jun 2023
2105.14368.pdf	Fit without fear: remarkable mathematical phenomena of deep learning through the prism of interpolation Mikhail Belkin Halicio˘ glu Data Science Institute, University of California San Diego La Jolla, USA In memory of Partha Niyogi, a thinker, a teacher, and a dear friend. Abstract In the past decade the mathematical theory of machine learning has lagged far behind the triumphs of deep neural networks on practical chal- lenges. However, the gap between theory and practice is gradually starting to close. In this paper I will attempt to assemble some pieces of the remark- able and still incomplete mathematical mosaic emerging from the eﬀorts to understand the foundations of deep learning. The two key themes will be interpolation, and its sibling, over-parameterization. Interpolation cor- responds to ﬁtting data, even noisy data, exactly. Over-parameterization enables interpolation and provides ﬂexibility to select a right interpolating model. As we will see, just as a physical prism separates colors mixed within a ray of light, the ﬁgurative prism of interpolation helps to disentangle generalization and optimization properties within the complex picture of modern Machine Learning. This article is written with belief and hope that clearer understanding of these issues brings us a step closer toward a general theory of deep learning and machine learning. Contents 1 Preface 2 2 Introduction 3 1arXiv:2105.14368v1 [stat.ML] 29 May 2021
2306.09927.pdf	arXiv:2306.09927v1 [stat.ML] 16 Jun 2023Trained Transformers Learn Linear Models In-Context Ruiqi Zhang UC Berkeley [email protected] Frei UC Berkeley [email protected] Peter L. Bartlett UC Berkeley and Google DeepMind [email protected] June 19, 2023 Abstract Attention-based neural networks such as transformers have demonstrated a remarkable ability to ex- hibit in-context learning (ICL): Given a short prompt seque nce of tokens from an unseen task, they can formulate relevant per-token and next-token predictions w ithout any parameter updates. By embedding a sequence of labeled training data and unlabeled test data as a prompt, this allows for transformers to be- have like supervised learning algorithms. Indeed, recent w ork has shown that when training transformer architectures over random instances of linear regression p roblems, these models’ predictions mimic those of ordinary least squares. Towards understanding the mechanisms underlying this phen omenon, we investigate the dynamics of ICL in transformers with a single linear self-attention lay er trained by gradient ﬂow on linear regression tasks. We show that despite non-convexity, gradient ﬂow wit h a suitable random initialization ﬁnds a global minimum of the objective function. At this global min imum, when given a test prompt of labeled examples from a new prediction task, the transformer achiev es prediction error competitive with the best linear predictor over the test prompt distribution. We additionally characterize the robustness of the trained transformer to a variety of distribution shifts and show that although a number of shifts are tolerated, shifts in the covariate distribution of the prom pts are not. Motivated by this, we consider a generalized ICL setting where the covariate distribution s can vary across prompts. We show that although gradient ﬂow succeeds at ﬁnding a global minimum in this setting, the trained transformer is still brittle under mild covariate shifts. 1 Introduction Transformer-based neural networks have quickly become the default machine learning model for problems in natural language processing, forming the basis of chatbo ts like ChatGPT [ Ope23 ], and are increasingly popular in computer vision [ Dos+21 ]. These models can take as input sequences of tokens and retu rn relevant next-token predictions. When trained on sufﬁcien tly large and diverse datasets, these models are often able to perform in-context learning (ICL): when given a short sequence of input-output pairs (ca lled a prompt ) from a particular task as input, the model can formulate pre dictions on test examples without having to make any updates to the parameters in the model. 1
2310.15418.pdf	Fractal Landscapes in Policy Optimization Tao Wang UC San Diego [email protected] Herbert UC San Diego [email protected] Gao UC San Diego [email protected] Abstract Policy gradient lies at the core of deep reinforcement learning (RL) in continuous domains. Despite much success, it is often observed in practice that RL training with policy gradient can fail for many reasons, even on standard control problems with known solutions. We propose a framework for understanding one inherent limitation of the policy gradient approach: the optimization landscape in the policy space can be extremely non-smooth or fractal for certain classes of MDPs, such that there does not exist gradient to be estimated in the first place. We draw on techniques from chaos theory and non-smooth analysis, and analyze the maximal Lyapunov exponents and Hölder exponents of the policy optimization objectives. Moreover, we develop a practical method that can estimate the local smoothness of objective function from samples to identify when the training process has encountered fractal landscapes. We show experiments to illustrate how some failure cases of policy optimization can be explained by such fractal landscapes. 1 Introduction Deep reinforcement learning has achieved much success in various applications [ 23,30,38], but they also often fail, especially in continuous spaces, on control problems that other methods can readily solve. The understanding of such failure cases is still limited. For instance, the training process of reinforcement learning is unstable and the learning curve can fluctuate during training in ways that are hard to predict. The probability of obtaining satisfactory policies can also be inherently low in reward-sparse or highly nonlinear control tasks. Existing analysis of the failures focuses on limitations of the sampling and optimization algorithms, such as function approximation errors [35,39], difficulty in data collection [ 34], and aggressive updates in the policy space [ 28]. There has not been much study of potentially deeper causes of failures that may be inherent in the formulation of policy optimization problems. Motivated by the common observation that small updates in the policy parameters can significantly change the performance, we analyze the smoothness of the optimization landscapes in policy opti- mization. Drawing on chaos theory, we introduce the concept of maximal Lyapunov exponent (MLE) [17] to the RL setting to measure the exponential rate of trajectory divergence in MDP. It seems contradictory that a trajectory in chaotic systems can be both exponentially divergent and uniformly bounded at the same time, and we will show that these two conflicting facts combine to yield the fractal structure in the optimization landscape. Intuitively, the objective function is non-differentiable when the rate of trajectory divergence exceeds the decay rate of discount factor. Furthermore, this finding indicates that the fluctuations observed in the loss curve are not just due to the numerical or sampling error but rather reflect the intrinsic properties of the corresponding MDP. We should emphasize that the fractal landscapes that we will demonstrate are stronger than various existing results on the non-smoothness [ 2,7]. Most nonsmooth objectives that have been studied still assume is local Lipschitz continuity or piecewise smoothness that implies differentiability almost everywhere (such as f(x) =\|x\|). Instead, by showing that the loss landscape of policy optimization can be fractal, we demonstrate the absence of descent directions, which causes the failure of first-order 37th Conference on Neural Information Processing Systems (NeurIPS 2023).arXiv:2310.15418v1 [cs.LG] 24 Oct 2023
2205.14135.pdf	FlashAttention : Fast and Memory-Eﬃcient Exact Attention with IO-Awareness Tri Daoy, Daniel Y. Fuy, Stefano Ermony, Atri Rudraz, and Christopher Réy yDepartment of Computer Science, Stanford University zDepartment of Computer Science and Engineering, University at Buﬀalo, SUNY {trid,danfu}@cs.stanford.edu ,[email protected] ,[email protected] , [email protected] June 24, 2022 Abstract Transformers are slow and memory-hungry on long sequences, since the time and memory complexity of self-attention are quadratic in sequence length. Approximate attention methods have attempted to address this problem by trading oﬀ model quality to reduce the compute complexity, but often do not achieve wall-clock speedup. We argue that a missing principle is making attention algorithms IO- aware—accounting for reads and writes between levels of GPU memory. We propose FlashAttention , an IO-aware exact attention algorithm that uses tiling to reduce the number of memory reads/writes between GPU high bandwidth memory (HBM) and GPU on-chip SRAM. We analyze the IO complexity ofFlashAttention , showing that it requires fewer HBM accesses than standard attention, and is optimal for a range of SRAM sizes. We also extend FlashAttention to block-sparse attention, yielding an approximate attention algorithm that is faster than any existing approximate attention method. FlashAttention trains Transformers faster than existing baselines: 15% end-to-end wall-clock speedup on BERT-large (seq. length 512) compared to the MLPerf 1.1 training speed record, 3 speedup on GPT-2 (seq. length 1K), and 2.4 speedup on long-range arena (seq. length 1K-4K). FlashAttention and block-sparse FlashAttention enable longer context in Transformers, yielding higher quality models (0.7 better perplexity on GPT-2 and 6.4 points of lift on long-document classiﬁcation) and entirely new capabilities: the ﬁrst Transformers to achieve better-than-chance performance on the Path-X challenge (seq. length 16K, 61.4% accuracy) and Path-256 (seq. length 64K, 63.1% accuracy). 1 Introduction Transformer models [ 82] have emerged as the most widely used architecture in applications such as natural language processing and image classiﬁcation. Transformers have grown larger [ 5] and deeper [ 83], but equipping them with longer context remains diﬃcult [ 80], since the self-attention module at their heart has time and memory complexity quadratic in sequence length. An important question is whether making attention faster and more memory-eﬃcient can help Transformer models address their runtime and memory challenges for long sequences. Many approximate attention methods have aimed to reduce the compute and memory requirements of attention. These methods range from sparse-approximation [ 51,74] to low-rank approximation [ 12,50,84], and their combinations [ 3,9,92]. Although these methods reduce the compute requirements to linear or near-linear in sequence length, many of them do not display wall-clock speedup against standard attention and have not gained wide adoption. One main reason is that they focus on FLOP reduction (which may not correlate with wall-clock speed) and tend to ignore overheads from memory access (IO). In this paper, we argue that a missing principle is making attention algorithms IO-aware [1]—that is, carefully accounting for reads and writes to diﬀerent levels of fast and slow memory (e.g., between fast GPU on-chip SRAM and relatively slow GPU high bandwidth memory, or HBM [ 45], Figure 1 left). On modern 1arXiv:2205.14135v2 [cs.LG] 23 Jun 2022
bayesian-interactive-optimization.pdf	Eurographics/ ACM SIGGRAPH Symposium on Computer Animation (2010) M. Otaduy and Z. Popovic (Editors) A Bayesian Interactive Optimization Approach to Procedural Animation Design Eric Brochu Tyson Brochu Nando de Freitas University of British Columbia Abstract The computer graphics and animation ﬁelds are ﬁlled with applications that require the setting of tricky param-eters. In many cases, the models are complex and the parameters unintuitive for non-experts. In this paper, wepresent an optimization method for setting parameters of a procedural ﬂuid animation system by showing the userexamples of different parametrized animations and asking for feedback. Our method employs the Bayesian tech-nique of bringing in “prior” belief based on previous runs of the system and/or expert knowledge, to assist users inﬁnding good parameter settings in as few steps as possible. To do this, we introduce novel extensions to Bayesianoptimization, which permit effective learning for parameter-based procedural animation applications. We showthat even when users are trying to ﬁnd a variety of different target animations, the system can learn and improve.We demonstrate the effectiveness of our method compared to related active learning methods. We also present aworking application for assisting animators in the challenging task of designing curl-based velocity ﬁelds, evenwith minimal domain knowledge other than identifying when a simulation “looks right”. Categories and Subject Descriptors (according to ACM CCS) : Learning [I.2.6]: Parameter Learning.—User Inter- faces [H.5.2]: Interaction Styles.—Three-Dimensional Graphics and Realism [I.3.7]: Animation.— 1 Introduction Procedural methods for generating animation have long been used by visual effects and games studios due to their efﬁ-ciency and artist controllability. However, this control comeswith a cost: a set of often unintuitive parameters confrontsthe user of a procedural animation system. The desired endresult is often identiﬁable by the user, but these parametersmust be tuned in a tedious trial-and-error process. For example, realistic animation of smoke can be achieved by driving a particle system through a simple combinationof vortex rings and curl noise [BHN07]. However even thesetwo relatively simple procedural methods are inﬂuenced byseveral parameters: The velocity, radius and magnitude ofthe vortex rings, and the length scale and magnitude of thecurl noise. Adding more procedural “ﬂow primitives”, suchas uniform and vortical ﬂows, sources and sinks [WH91],turbulent wind [SF93], vortex particles [SRF05], and vortexﬁlaments [AN05] can produce a wider variety of animations,but each of these primitives carries its own set of associatedparameters. These parameters can interact in subtle and non-intuitive ways, and small adjustments to certain settings mayresult in non-uniform changes in the appearance. Brochu et al. [BGdF07, BdFG07] propose a Bayesian op- timization technique to assist artists with parameter tuningfor bidirectional reﬂectance distribution functions (BRDF s). In their iterative scheme, the algorithm selects two sets ofparameters and generates example images from them. Theuser selects the preferred image and the algorithm incorpo-rates this feedback to learn a model of the user’s valuation function over the domain of parameter values. Given thisvaluation function, the algorithm is able to select parame-ters to generate simulations that are likely to be closer to theones wanted by the artist. The process is repeated until theuser is satisﬁed with the results. During the development of a procedural smoke anima- tion system, we found ourselves with a parameterized sys-tem with 12 continuous parameters. Setting these was a chal-lenge for the developers, let alone other users, so we lookedto adapt [BdFG07]. In the process, though, we found that themodel as presented was unsuitable for our procedural anima-tion. In particular, we identiﬁed several limitations: c⃝The Eurographics Association 2010. DOI: 10.2312/SCA/SCA10/103-112
Introduction to Probabilistic Topic Models.pdf	Introduction to Probabilistic Topic Models David M. Blei Princeton University Abstract Probabilistic topic models are a suite of algorithms whose aim is to discover the hidden thematic structure in large archives of documents. In this article, we review the main ideas of this ﬁeld, survey the current state-of-the-art, and describe some promising future directions. We ﬁrst describe latent Dirichlet allocation (LDA) [ 8], which is the simplest kind of topic model. We discuss its connections to probabilistic modeling, and describe two kinds of algorithms for topic discovery. We then survey the growing body of research that extends and applies topic models in interesting ways. These extensions have been developed by relaxing some of the statistical assumptions of LDA, incorporating meta-data into the analysis of the documents, and using similar kinds of models on a diversity of data types such as social networks, images and genetics. Finally, we give our thoughts as to some of the important unexplored directions for topic modeling. These include rigorous methods for checking models built for data exploration, new approaches to visualizing text and other high dimensional data, and moving beyond traditional information engineering applications towards using topic models for more scientiﬁc ends. 1I n t r o d u c t i o n As our collective knowledge continues to be digitized and stored—in the form of news, blogs, web pages, scientiﬁc articles, books, images, sound, video, and social networks—it becomes more dicult to ﬁnd and discover what we are looking for. We need new computational tools to help organize, search and understand these vast amounts of information. Right now, we work with online information using two main tools—search and links. We type keywords into a search engine and ﬁnd a set of documents related to them. We look at the documents in that set, possibly navigating to other linked documents. This is a powerful way of interacting with our online archive, but something is missing. Imagine searching and exploring documents based on the themes that run through them. We might “zoom in” and “zoom out” to ﬁnd speciﬁc or broader themes; we might look at how those themes changed through time or how they are connected to each other. Rather than 1
GPT-2.pdf	Language Models are Unsupervised Multitask Learners Alec Radford1Jeffrey Wu1Rewon Child1David Luan1Dario Amodei1Ilya Sutskever1 Abstract Natural language processing tasks, such as ques- tion answering, machine translation, reading com- prehension, and summarization, are typically approached with supervised learning on task- speciﬁc datasets. We demonstrate that language models begin to learn these tasks without any ex- plicit supervision when trained on a new dataset of millions of webpages called WebText. When conditioned on a document plus questions, the an- swers generated by the language model reach 55 F1 on the CoQA dataset - matching or exceeding the performance of 3 out of 4 baseline systems without using the 127,000+ training examples. The capacity of the language model is essential to the success of zero-shot task transfer and in- creasing it improves performance in a log-linear fashion across tasks. Our largest model, GPT-2, is a 1.5B parameter Transformer that achieves state of the art results on 7 out of 8 tested lan- guage modeling datasets in a zero-shot setting but still underﬁts WebText. Samples from the model reﬂect these improvements and contain co- herent paragraphs of text. These ﬁndings suggest a promising path towards building language pro- cessing systems which learn to perform tasks from their naturally occurring demonstrations. 1. Introduction Machine learning systems now excel (in expectation) at tasks they are trained for by using a combination of large datasets, high-capacity models, and supervised learning (Krizhevsky et al., 2012) (Sutskever et al., 2014) (Amodei et al., 2016). Yet these systems are brittle and sensitive to slight changes in the data distribution (Recht et al., 2018) and task speciﬁcation (Kirkpatrick et al., 2017). Current sys- tems are better characterized as narrow experts rather than , *Equal contribution1OpenAI, San Francisco, Califor- nia, United States. Correspondence to: Alec Radford <[email protected] >.competent generalists. We would like to move towards more general systems which can perform many tasks – eventually without the need to manually create and label a training dataset for each one. The dominant approach to creating ML systems is to col- lect a dataset of training examples demonstrating correct behavior for a desired task, train a system to imitate these behaviors, and then test its performance on independent and identically distributed (IID) held-out examples. This has served well to make progress on narrow experts. But the often erratic behavior of captioning models (Lake et al., 2017), reading comprehension systems (Jia & Liang, 2017), and image classiﬁers (Alcorn et al., 2018) on the diversity and variety of possible inputs highlights some of the short- comings of this approach. Our suspicion is that the prevalence of single task training on single domain datasets is a major contributor to the lack of generalization observed in current systems. Progress towards robust systems with current architectures is likely to require training and measuring performance on a wide range of domains and tasks. Recently, several benchmarks have been proposed such as GLUE (Wang et al., 2018) and decaNLP (McCann et al., 2018) to begin studying this. Multitask learning (Caruana, 1997) is a promising frame- work for improving general performance. However, mul- titask training in NLP is still nascent. Recent work re- ports modest performance improvements (Yogatama et al., 2019) and the two most ambitious efforts to date have trained on a total of 10 and 17 (dataset, objective) pairs respectively (McCann et al., 2018) (Bowman et al., 2018). From a meta-learning perspective, each (dataset, objective) pair is a single training example sampled from the distribution of datasets and objectives. Current ML systems need hundreds to thousands of examples to induce functions which generalize well. This suggests that multitask training many need just as many effective training pairs to realize its promise with current approaches. It will be very difﬁcult to continue to scale the creation of datasets and the design of objectives to the degree that may be re- quired to brute force our way there with current techniques. This motivates exploring additional setups for performing multitask learning. The current best performing systems on language tasks
2647-elbo-ing-stein-mixtures.pdf	Under review as a conference paper at ICLR 2023 ELBO- INGSTEIN MIXTURES Anonymous authors Paper under double-blind review ABSTRACT Stein variational gradient descent (SVGD) (Liu & Wang, 2016) is a particle-based technique for Bayesian inference. SVGD has recently gained popularity because it combines the ability of variational inference to handle tall data with the modeling power of non-parametric inference. Unfortunately, variance estimation scales inversely with the dimensionality of the model leading to underestimation, meaning more particles are required to represent high-dimensional models adequately. Stein mixtures (Nalisnick & Smyth, 2017) alleviate the exponential growth in particles by letting each particle parameterize a distribution. However, the inference algorithm proposed by Nalisnick & Smyth (2017) can be numerically unstable. We show that their algorithm corresponds to inference with the Rényi α-divergence for α= 0and that using other values for αcan lead to a more stable inference. We empirically study the performance of Stein mixtures inferred with different αvalues on various real-world problems, demonstrating significantly and consistently improved results when using α= 1, which corresponds to using the evidence lower bound (ELBO). We call this instance of our algorithm ELBO-within-Stein. An easy-to-use version of the inference algorithm (for arbitrary α∈R) is available in the deep probabilistic programming language NumPyro (Phan et al., 2019). 1 I NTRODUCTION The ability of Bayesian deep learning to quantify the uncertainty of predictions by deep models is causing a surge of interest in using these techniques (Izmailov et al., 2021). Bayesian inference aims to describe i.i.d. data D={xi}n i=1using a model with latent a variable z. Bayesian inference does this by computing a posterior distribution p(z\|D)over the latent variable given a model describing the joint distribution p(z,D) =p(D\|z)p(z). We obtain the posterior by following Bayes’ theorem, p(z\|D) =Qn i=1p(xi\|z)p(z)/p(D), where p(D) =R zQn i=1p(xi\|z)p(z)dzis the normalization constant. For most practical models, the normalization constant lacks an analytic solution or poses a computability problem, complicating the Bayesian inference problem. Stein variational gradient descent (SVGD) (Liu & Wang, 2016) is a recent technique for Bayesian inference that uses a set of particles Z={zi}N i=1to approximate the posterior p(z\|D). The idea behind SVGD is to iteratively transport Zaccording to a force field SZ, called the Stein force. The Stein force is given by SZ(zi) =Ezj∼qZ[k(zi,zj)∇zjlogp(zj\|D) +∇zjk(zi,zj)], (1) where k(·,·)is a reproducing kernel (Berlinet & Thomas-Agnan, 2011), qZ=N−1P iδziis the empirical measure on the set of particles Z,δx(y)represents the Dirac delta measure, which is equal to1ifx=yand0otherwise, and ∇zjlogp(zj\|D)is the gradient of the posterior with respect to thej-th particle. The technique is scalable to tall data (i.e. datasets with many data points) and offers the flexibility and scope of techniques such as Markov chain Monte Carlo (MCMC). SVGD is good at capturing multi-modality (Liu & Wang, 2016; Wang & Liu, 2019), and has useful theoretical interpretations such as a set of particles following a gradient flow (Liu, 2017) or in terms of the properties of kernels (Liu & Wang, 2018). The main problem is that SVGD suffers from the curse of dimensionality: variance estimation scales inversely with dimensionality (Ba et al., 2021). Nalisnick & Smyth (2017) suggest resolving this by 1
1711.00165.pdf	Published as a conference paper at ICLR 2018 DEEPNEURAL NETWORKS AS GAUSSIAN PROCESSES Jaehoon Lee∗†, Yasaman Bahri∗†, Roman Novak , Samuel S. Schoenholz, Jeffrey Pennington, Jascha Sohl-Dickstein Google Brain {jaehlee, yasamanb, romann, schsam, jpennin, jaschasd }@google.com ABSTRACT It has long been known that a single-layer fully-connected neural network with an i.i.d. prior over its parameters is equivalent to a Gaussian process (GP), in the limit of inﬁnite network width. This correspondence enables exact Bayesian inference for inﬁnite width neural networks on regression tasks by means of evaluating the corresponding GP. Recently, kernel functions which mimic multi-layer random neural networks have been developed, but only outside of a Bayesian framework. As such, previous work has not identiﬁed that these kernels can be used as co- variance functions for GPs and allow fully Bayesian prediction with a deep neural network. In this work, we derive the exact equivalence between inﬁnitely wide deep net- works and GPs. We further develop a computationally efﬁcient pipeline to com- pute the covariance function for these GPs. We then use the resulting GPs to per- form Bayesian inference for wide deep neural networks on MNIST and CIFAR- 10. We observe that trained neural network accuracy approaches that of the corre- sponding GP with increasing layer width, and that the GP uncertainty is strongly correlated with trained network prediction error. We further ﬁnd that test perfor- mance increases as ﬁnite-width trained networks are made wider and more similar to a GP, and thus that GP predictions typically outperform those of ﬁnite-width networks. Finally we connect the performance of these GPs to the recent theory of signal propagation in random neural networks. 1 I NTRODUCTION Deep neural networks have emerged in recent years as ﬂexible parametric models which can ﬁt complex patterns in data. As a contrasting approach, Gaussian processes have long served as a traditional nonparametric tool for modeling. An equivalence between these two approaches was derived in Neal (1994a), for the case of one layer networks in the limit of inﬁnite width. Neal (1994a) further suggested that a similar correspondence might hold for deeper networks. Consider a deep fully-connected neural network with i.i.d. random parameters. Each scalar output of the network, an afﬁne transformation of the ﬁnal hidden layer, will be a sum of i.i.d. terms. As we will discuss in detail below, in the limit of inﬁnite width the Central Limit Theorem1implies that the function computed by the neural network (NN) isa function drawn from a Gaussian process (GP). In the case of single hidden-layer networks, the form of the kernel of this GP is well known (Neal (1994a); Williams (1997)). This correspondence implies that if we choose the hypothesis space to be the class of inﬁnitely wide neural networks, an i.i.d. prior over weights and biases can be replaced with a corresponding GP prior over functions. As noted by (Williams, 1997), this substitution enables exact Bayesian inference for regression using neural networks. The computation requires building the necessary covariance matrices over the training and test sets and straightforward linear algebra computations. ∗Both authors contributed equally to this work. †Work done as a member of the Google AI Residency program (g.co/airesidency). 1Throughout this paper, we assume the conditions on the parameter distributions and nonlinearities are such that the Central Limit Theorem will hold; for instance, that the weight variance is scaled inversely proportional to the layer width. 1arXiv:1711.00165v3 [stat.ML] 3 Mar 2018
2210.03370.pdf	GNM: A General Navigation Model to Drive Any Robot Dhruv Shah†β, Ajay Sridhar†β, Arjun Bhorkarβ, Noriaki Hiroseβτ, Sergey Levineβ 𝜏 ot og GNM Training Large Heterogeneous Datasets Fig. 1: A general navigation model to drive any robot. By training on diverse, heterogeneous datasets, a single “omnipolicy” can control a variety of robots in challenging environments, including new robots, without any robot-speciﬁc data collection. Abstract — Learning provides a powerful tool for vision-based navigation, but the capabilities of learning-based policies are constrained by limited training data. If we could combine data from all available sources, including multiple kinds of robots, we could train more powerful navigation models. In this paper, we study how a general goal-conditioned model for vision-based navigation can be trained on data obtained from many distinct but structurally similar robots, and enable broad generalization across environments and embodiments. We analyze the necessary design decisions for effective data sharing across robots, including the use of temporal context and standardized action spaces, and demonstrate that an omnipolicy trained from heterogeneous datasets outperforms policies trained on any single dataset. We curate 60 hours of navigation trajectories from 6 distinct robots, and deploy the trained GNM on a range of new robots, including an underactuated quadrotor. We ﬁnd that training on diverse data leads to robustness against degradation in sensing and actuation. Using a pre-trained navigation model with broad generalization capabilities can bootstrap applications on novel robots going forward, and we hope that the GNM represents a step in that direction. For more information on the datasets, code, and videos, please check out our project page1. I. I NTRODUCTION Machine learning methods have enabled broad general- ization with real-world applicability in natural language pro- cessing [1], visual perception [2–4], and other domains [5, 6] by leveraging Internet-scale data. Such generalization typi- cally requires learning general patterns from diverse datasets, which are usually collected once and then reused for various purposes. Such large-scale models also support the ability to be adapted for new tasks by reusing the representations learned from broader, larger, and more general datasets, for example by or zero-shot transfer [7–9], or ﬁne-tuning on target-domain data. Although this paradigm has been very successful, it is difﬁcult to apply in robotics due to the sheer diversity of environments and platforms across researchers. Control policies learned end-to-end usually require sepa- rate data collection for each robotic platform, leading to “fragmentation” in progress, where every researcher works with their own robot-speciﬁc dataset and policies, making it infeasible to accumulate large enough datasets. Can we 1sites.google.com/view/drive-any-robot †Equal Contribution.βUC Berkeley,τToyota Motor North America.overcome this challenge by training models on more general and reusable cross-robot datasets? We study this question in the context of visual navigation, where heterogeneity between robots might include different camera hardware, viewpoints, dynamics, and more broadly, embodiments, but where the over-arching navigation objec- tive looks similar irrespective of these differences. A wheeled robot, quadruped, or a drone all have the same abstract objectives: to explore the environment, plan a path to the goal, and avoid collisions. Leveraging this shared abstraction across robots and training a general navigational omnipolicy from large-scale data could enable broad generalization to novel environments, unseen sensor parameters (e.g., camera intrinsics and extrinsics), and new robot conﬁgurations. In this paper, we propose to take a step towards this kind of data sharing by training an embodiment-agnostic general navigation model (GNM) from an aggregated multi- robot dataset. The primary contribution of our work is a framework for training a general omnipolicy from multi- robot datasets, with empirical evidence that such an om- nipolicy can effectively learn from heterogeneous datasets and generalize to novel robot platforms. To facilitate this, we aggregate a large heterogeneous dataset of navigation trajectories collected across 6 robots, spanning 60 hours of interactions in challenging indoor and outdoor environments. We train the GNM on this dataset and deploy it on 4 distinct robot platforms, including 2 new robots. We show that a single learned policy can be used across multiple robots to perform goal-reaching in challenging indoor and outdoor environments, outperforming policies trained with any single dataset. We also report robustness to degradation in camera parameters, tire damage, and other gradual changes that the robot may experience over its lifetime. We have publicly released the trained GNM policy, code used to train and deploy our models on various popular robot platforms, as well as the dataset used to train these models at our project page. We hope that this represents a step to- wards both general-purpose multi-robot datasets and general- purpose visual navigational models that can be deployed on a wide range of robots — similar to how practitioners currently use pre-trained models in vision and language, such models could constitute pre-trained backbones for visual navigation.arXiv:2210.03370v2 [cs.RO] 22 May 2023
2203.03466.pdf	Tensor Programs V: Tuning Large Neural Networks via Zero-Shot Hyperparameter Transfer Greg Yang∗×Edward J. Hu∗×†Igor Babuschkin◦Szymon Sidor◦Xiaodong Liu× David Farhi◦Nick Ryder◦Jakub Pachocki◦Weizhu Chen×Jianfeng Gao× ×Microsoft Corporation◦OpenAI Abstract Hyperparameter (HP) tuning in deep learning is an expensive process, prohibitively so for neural networks (NNs) with billions of parameters. We show that, in the recently discovered Maximal Update Parametrization ( µP), many optimal HPs remain stable even as model size changes. This leads to a new HP tuning paradigm we callµTransfer : parametrize the target model in µP, tune the HP indirectly on a smaller model, and zero-shot transfer them to the full-sized model, i.e., without directly tuning the latter at all. We verify µTransfer on Transformer and ResNet. For example, 1) by transferring pretraining HPs from a model of 13M parameters, we outperform published numbers of BERT-large (350M parameters), with a total tuning cost equivalent to pretraining BERT-large once; 2) by transferring from 40M parameters, we outperform published numbers of the 6.7B GPT-3 model, with tuning cost only 7% of total pretraining cost. A Pytorch implementation of our technique can be found at github.com/microsoft/mup and installable via pip install mup . 1 Introduction 20 18 16 14 12 10 log2LearningRate 3.5 4.0 4.5 5.0 5.5 6.0 6.5 7.0Training Loss optimum shiftsStandard Practice Width 128 256 512 1024 2048 4096 8192 20 18 16 14 12 10 log2LearningRate optimum stableOur Work Figure 1: Training loss against learning rate on Transformers of varying dmodel trained with Adam. Conventionally and in contrast with our technique, different widths do not share the same optimal hy- perparameter; wider networks do not always per- form better than narrower ones; in fact they under- perform the same-width networks in our technique even after tuning learning rate (see dashed line). See Sections 3 and 4 for experimental setup.Hyperparameter (HP) tuning is critical to deep learning. Poorly chosen HPs result in subpar performance and training instability. Many pub- lished baselines are hard to compare to one another due to varying degrees of HP tuning. These issues are exacerbated when training ex- tremely large deep learning models, since state- of-the-art networks with billions of parameters become prohibitively expensive to tune. Recently, [ 57] showed that different neural net- work parametrizations induce different inﬁnite- width limits and proposed the Maximal Update Parametrization (abbreviated µP)(summarized in Table 3) that enables “maximal” feature learn- ing in the limit. Intuitively, it ensures that each layer is updated on the same order during train- ingregardless of width .2In contrast, while the standard parametrization (SP) ensures activations are of unit order at initialization, it actually causes them to blow up in wide models during training [ 57] essentially due to an imbalance of per-layer †Work done partly during Microsoft AI Residency Program. ∗Equal contribution. Order is random. Correspondence to {gregyang, edwardhu}@microsoft.com 2i.e., the updates’ effect on activations becomes roughly independent of width in the large width limit.arXiv:2203.03466v2 [cs.LG] 28 Mar 2022
1705.01509.pdf	Neural Models for Information Retrieval Bhaskar Mitra Microsoft, UCL∗ Cambridge, UK [email protected] Craswell Microsoft Bellevue, USA [email protected] Abstract Neural ranking models for information retrieval (IR) use shallow or deep neural networks to rank search results in response to a query. Traditional learning to rank models employ machine learning techniques over hand-crafted IR features. By contrast, neural models learn representations of language from raw text that can bridge the gap between query and document vocabulary. Unlike classical IR models, these new machine learning based approaches are data-hungry, requiring large scale training data before they can be deployed. This tutorial introduces basic concepts and intuitions behind neural IR models, and places them in the context of traditional retrieval models. We begin by introducing fundamental concepts of IR and different neural and non-neural approaches to learning vector representations of text. We then review shallow neural IR methods that employ pre-trained neural term embeddings without learning the IR task end-to-end. We introduce deep neural networks next, discussing popular deep architectures. Finally, we review the current DNN models for information retrieval. We conclude with a discussion on potential future directions for neural IR. 1 Introduction Since the turn of the decade, there have been dramatic improvements in performance in computer vision, speech recognition, and machine translation tasks, witnessed in research and in real-world applications [ 112]. These breakthroughs were largely fuelled by recent advances in neural network models, usually with multiple hidden layers, known as deep architectures [ 8,49,81,103,112]. Exciting novel applications, such as conversational agents [ 185,203], have also emerged, as well as game-playing agents with human-level performance [ 147,180]. Work has now begun in the information retrieval (IR) community to apply these neural methods, leading to the possibility of advancing the state of the art or even achieving breakthrough performance as in these other ﬁelds. Retrieval of information can take many forms. Users can express their information need in the form of a text query—by typing on a keyboard, by selecting a query suggestion, or by voice recognition—or the query can be in the form of an image, or in some cases the need can even be implicit. Retrieval can involve ranking existing pieces of content, such as documents or short-text answers, or composing new responses incorporating retrieved information. Both the information need and the retrieved results may use the same modality (e.g., retrieving text documents in response to keyword queries), or different ones (e.g., image search using text queries). Retrieval systems may consider user history, physical location, temporal changes in information, or other context when ranking results. They may also help users formulate their intent (e.g., via query auto-completion or query suggestion) and/or extract succinct summaries of results for easier inspection. Neural IR refers to the application of shallow or deep neural networks to these retrieval tasks. This tutorial serves as an introduction to neural methods for ranking documents in response to a query, an ∗The author is a part-time PhD student at University College London. DRAFT. Copyright is held by the author(s). May, 2017.arXiv:1705.01509v1 [cs.IR] 3 May 2017
2312.12456.pdf	arXiv:2312.12456v1 [cs.LG] 16 Dec 2023PowerInfer: Fast Large Language Model Serving with a Consum er-grade GPU Yixin Song, Zeyu Mi∗, Haotong Xie and Haibo Chen Institute of Parallel and Distributed Systems (IPADS), Sha nghai Jiao Tong University Abstract This paper introduces PowerInfer, a high-speed Large Lan- guage Model (LLM) inference engine on a personal com- puter (PC) equipped with a single consumer-grade GPU. The key underlying the design of PowerInfer is exploiting the high locality inherent in LLM inference, characterized by a power-law distribution in neuron activation. This distrib ution indicates that a small subset of neurons, termed hot neurons , are consistently activated across inputs, while the majori ty, cold neurons , vary based on speciﬁc inputs. PowerInfer ex- ploits such an insight to design a GPU-CPU hybrid inference engine: hot-activated neurons are preloaded onto the GPU for fast access, while cold-activated neurons are computed on the CPU, thus signiﬁcantly reducing GPU memory demands and CPU-GPU data transfers. PowerInfer further integrates adaptive predictors and neuron-aware sparse operators, op ti- mizing the efﬁciency of neuron activation and computationa l sparsity. Evaluation shows that PowerInfer attains an aver age token generation rate of 13.20 tokens/s, with a peak of 29.08 tokens/s, across various LLMs (including OPT-175B) on a single NVIDIA RTX 4090 GPU, only 18% lower than that achieved by a top-tier server-grade A100 GPU. This signiﬁ- cantly outperforms llama.cpp by up to 11.69 ×while retain- ing model accuracy. 1 Introduction Generative large language models (LLMs) have garnered attention for their remarkable capabilities in creative wr it- ing, advanced code generation, and sophisticated natural l an- guage processing tasks [ 5,42,49]. These models, widely de- ployed in data centers equipped with high-end and expensive server-grade GPUs, have signiﬁcantly inﬂuenced our daily lives and work practices. Meanwhile, there is an emerging trend of running LLMs on more accessible local platforms, particularly personal computers (PCs) with consumer-grad e GPUs. This evolution is driven by the need for enhanced data privacy [ 25], model customization [ 22], and reduced in- ference costs [ 42]. In contrast to data-center deployments, which prioritize high throughput [ 18,37,47], local deploy- ments focus on low latency in processing small batches. Nonetheless, deploying LLMs on consumer-grade GPUs presents signiﬁcant challenges due to their substantial me m- ∗Zeyu Mi (yzmizeyu@sjtu .edu.cn) is the corresponding author.ory requirements. LLMs, typically functioning as autoregr es- sive Transformers, sequentially generate text token-by-t oken, each needing to access the entire model consisting of hun- dreds of billions of parameters. Therefore, the inference p ro- cess is fundamentally constrained by the GPU’s memory ca- pacity. This limitation is particularly acute in local depl oy- ments where the processing of individual requests (often ju st one at a time) [ 6] leaves minimal opportunity for parallel pro- cessing. Existing approaches to such memory issues include model compression and ofﬂoading. Compression techniques like quantization [ 12,46], distillation [ 48], and pruning [ 23] re- duce the model size. However, even deeply compressed models remain too large for consumer-grade GPUs. For in- stance, an OPT-66B model with 4-bit precision demands approximately 40GB of memory just to load its parame- ters [ 20], exceeding the capacity of even high-end GPUs like the NVIDIA RTX 4090. Model ofﬂoading, which partitions the model between GPU and CPU at the Transformer layer level [ 3,14,37]. State-of-the-art systems like llama.cpp [ 14] distribute layers between CPU and GPU memories, leverag- ing both for inference, thus reducing the GPU resources re- quired. However, this method is hindered by the slow PCIe interconnect and the CPUs’ limited computational capabili - ties, resulting in high inference latency. In this paper, we argue that the key reason for memory issues in LLM inference is the locality mismatch between hardware architecture and the characteristics of LLM in- ference. Current hardware architectures are designed with a memory hierarchy optimized for data locality. Ideally, a small, frequently accessed working set should be stored in the GPU, which offers higher memory bandwidth but limited capacity. In contrast, larger, less frequently accessed da ta are better suited for CPUs, which provide more extensive mem- ory capacity but lower bandwidth. Nevertheless, the vast vo l- ume of parameters required for each LLM inference iteration leads to a working set that is too large for a single GPU, thus impeding efﬁcient locality exploitation. We have observed that LLM inference inherently exhibits high locality. Speciﬁcally, during each inference iterati on, a limited number of neurons1are activated, signiﬁcantly inﬂu- 1This paper deﬁnes a neuron as a speciﬁc row/column in a weight matrix. 1
1802.09568.pdf	arXiv:1802.09568v2 [cs.LG] 2 Mar 2018Shampoo: Preconditioned Stochastic Tensor Optimization Vineet Gupta6Tomer Koren6Yoram Singer‹ March 5, 2018 Abstract Preconditioned gradient methods are among the most general and powerful tools in optimization. However, preconditioning requires storing and manipulating prohibitively large matrices. We describe and analyze a new structure-awa re preconditioning algorithm, called Shampoo, for stochastic optimization over tensor sp aces. Shampoo maintains a set of preconditioning matrices, each of which operates on a singl e dimension, contracting over the remaining dimensions. We establish convergence guarantee s in the stochastic convex setting, the proof of which builds upon matrix trace inequalities. Ou r experiments with state-of- the-art deep learning models show that Shampoo is capable of converging considerably faster than commonly used optimizers. Although it involves a more complex update rule, Shampoo’s runtime per step is comparable to that of simple gr adient methods such as SGD, AdaGrad, and Adam. 1 Introduction Over the last decade, stochastic ﬁrst-order optimization m ethods have emerged as the canonical tools for training large-scale machine learning models. Th ese methods are particularly appealing due to their wide applicability and their low runtime and mem ory costs. A potentially more powerful family of algorithms consists o fpreconditioned gradient methods. Preconditioning methods maintain a matrix, termed a precon ditioner, which is used to transform (i.e., premultiply) the gradient vector before it is used to take a step. Classic algorithms in this family include Newton’s method, which employs the local Hes sian as a preconditioner, as well as a plethora of quasi-Newton methods (e.g., [ 8,15,19]) that can be used whenever second-order information is unavailable or too expensive to compute. New er additions to this family are preconditioned online algorithms, most notably AdaGrad [ 6], that use the covariance matrix of the accumulated gradients to form a preconditioner. While preconditioned methods often lead to improved conver gence properties, the dimen- sionality of typical problems in machine learning prohibit s out-of-the-box use of full-matrix preconditioning. To mitigate this issue, specialized vari ants have been devised in which the full preconditioner is replaced with a diagonal approximation [ 6,14], a sketched version [ 9,20], or various estimations thereof [ 7,2,23]. While the diagonal methods are heavily used in practice thanks to their favorable scaling with the dimension, the ot her approaches are seldom practical at large scale as one typically requires a ﬁne approximation (or estimate) of the preconditioner that often demands super-linear memory and computation. In this paper, we take an alternative approach to preconditi oning and describe an eﬃcient and practical apparatus that exploits the structure of the p arameter space. Our approach is motivated by the observation that in numerous machine learn ing applications, the parameter space entertains a more complex structure than a monolithic vector in Euclidean space. In 6Google Brain. Email: {vineet,tkoren}@google.com ‹Princeton University and Google Brain. Email: [email protected] 1
Variational Inference.pdf	Variational Inference David M. Blei 1 Set up •As usual, we will assume that x=x1:nare observations and z=z1:mare hidden variables. We assume additional parameters αthat are ﬁxed. •Note we are general—the hidden variables might include the “parameters,” e.g., in a traditional inference setting. (In that case, αare the hyperparameters.) •We are interested in the posterior distribution , p(z\|x,α) =p(z,x\|α)∫ zp(z,x\|α). (1) •As we saw earlier, the posterior links the data and a model. It is used in all downstream analyses, such as for the predictive distribution. •(Note: The problem of computing the posterior is an instance of a more general problem that variational inference solves.) 2 Motivation •We can’t compute the posterior for many interesting models. •Consider the Bayesian mixture of Gaussians, 1. Drawµk∼N(0,τ2) fork= 1...K . 2. Fori= 1...n: (a) Drawzi∼Mult(π); 1
2208.02813.pdf	Towards Understanding Mixture of Experts in Deep Learning Zixiang Chen∗and Yihe Deng†and Yue Wu‡and Quanquan Gu§and Yuanzhi Li¶ Abstract The Mixture-of-Experts (MoE) layer, a sparsely-activated model controlled by a router, has achieved great success in deep learning. However, the understanding of such architecture remains elusive. In this paper, we formally study how the MoE layer improves the performance of neural network learning and why the mixture model will not collapse into a single model. Our empirical results suggest that the cluster structure of the underlying problem and the non-linearity of the expert are pivotal to the success of MoE. To further understand this, we consider a challenging classiﬁcation problem with intrinsic cluster structures, which is hard to learn using a single expert. Yet with the MoE layer, by choosing the experts as two-layer nonlinear convolutional neural networks (CNNs), we show that the problem can be learned successfully. Furthermore, our theory shows that the router can learn the cluster-center features, which helps divide the input complex problem into simpler linear classiﬁcation sub-problems that individual experts can conquer. To our knowledge, this is the ﬁrst result towards formally understanding the mechanism of the MoE layer for deep learning. 1 Introduction The Mixture-of-Expert (MoE) structure (Jacobs et al., 1991; Jordan and Jacobs, 1994) is a classic design that substantially scales up the model capacity and only introduces small computation overhead. In recent years, the MoE layer (Eigen et al., 2013; Shazeer et al., 2017), which is an extension of the MoE model to deep neural networks, has achieved remarkable success in deep learning. Generally speaking, an MoE layer contains many experts that share the same network architecture and are trained by the same algorithm, with a gating (or routing) function that routes individual inputs to a few experts among all the candidates. Through the sparse gating function, the router in the MoE layer can route each input to the top- K(K≥2) best experts (Shazeer et al., 2017), or the single ( K= 1) best expert (Fedus et al., 2021). This routing scheme only costs the computation of Kexperts for a new input, which enjoys fast inference time. ∗Department of Computer Science, University of California, Los Angeles, CA 90095, USA; e-mail: [email protected] †Department of Computer Science, University of California, Los Angeles, CA 90095, USA; e-mail: [email protected] ‡Department of Computer Science, University of California, Los Angeles, CA 90095, USA; e-mail: : [email protected] §Department of Computer Science, University of California, Los Angeles, CA 90095, USA; e-mail: [email protected] ¶Machine Learning Department, Carnegie Mellon University, Pittsburgh, PA, USA; email: [email protected] 1arXiv:2208.02813v1 [cs.LG] 4 Aug 2022
wenzel20a.pdf	How Good is the Bayes Posterior in Deep Neural Networks Really? Florian Wenzel* 1Kevin Roth* + 2Bastiaan S. Veeling* + 3 1Jakub ´Swi ˛ atkowski4 +Linh Tran5 + Stephan Mandt6 +Jasper Snoek1Tim Salimans1Rodolphe Jenatton1Sebastian Nowozin7 + Abstract During the past ﬁve years the Bayesian deep learn- ing community has developed increasingly accu- rate and efﬁcient approximate inference proce- dures that allow for Bayesian inference in deep neural networks. However, despite this algo- rithmic progress and the promise of improved uncertainty quantiﬁcation and sample efﬁciency there are—as of early 2020—no publicized de- ployments of Bayesian neural networks in indus- trial practice. In this work we cast doubt on the current understanding of Bayes posteriors in popular deep neural networks: we demonstrate through careful MCMC sampling that the pos- terior predictive induced by the Bayes posterior yields systematically worse predictions compared to simpler methods including point estimates ob- tained from SGD. Furthermore, we demonstrate that predictive performance is improved signiﬁ- cantly through the use of a “cold posterior” that overcounts evidence. Such cold posteriors sharply deviate from the Bayesian paradigm but are com- monly used as heuristic in Bayesian deep learn- ing papers. We put forward several hypotheses that could explain cold posteriors and evaluate the hypotheses through experiments. Our work questions the goal of accurate posterior approx- imations in Bayesian deep learning: If the true Bayes posterior is poor, what is the use of more accurate approximations? Instead, we argue that it is timely to focus on understanding the origin of the improved performance of cold posteriors. CODE: https://github.com/ google-research/google-research/ tree/master/cold_posterior_bnn *Equal contribution+Work done while at Google1Google Research2ETH Zurich3University of Amsterdam4University of Warsaw5Imperial College London6University of California, Irvine7Microsoft Research. Correspondence to: Florian Wenzel <ﬂ[email protected]>. Proceedings of the 37thInternational Conference on Machine Learning , Online, PMLR 119, 2020. Copyright 2020 by the au- thor(s). 10−410−310−210−1100 Temperature T0.880.900.920.94Test accuracy SG-MCMC Baseline: SGDFigure 1. The “ cold posterior ” effect: for a ResNet-20 on CIFAR- 10 we can improve the generalization performance signiﬁcantly by cooling the posterior with a temperature T≪1, deviating from the Bayes posterior p(θ\|D)∝exp(−U(θ)/T)atT= 1. 1. Introduction In supervised deep learning we use a training dataset D={(xi,yi)}i=1,...,n and a probabilistic model p(y\|x,θ) to minimize the regularized cross-entropy objective, L(θ) :=−1 nn∑ i=1logp(yi\|xi,θ) + Ω( θ), (1) where Ω(θ)is a regularizer over model parameters. We approximately optimize (1) using variants of stochastic gra- dient descent (SGD), (Sutskever et al., 2013). Beside being efﬁcient, the SGD minibatch noise also has generalization beneﬁts (Masters & Luschi, 2018; Mandt et al., 2017). 1.1. Bayesian Deep Learning In Bayesian deep learning we do not optimize for a single likely model but instead want to discover alllikely models. To this end we approximate the posterior distribution over model parameters, p(θ\|D)∝exp(−U(θ)/T), whereU(θ) is the posterior energy function , U(θ) :=−n∑ i=1logp(yi\|xi,θ)−logp(θ), (2) andTis atemperature . Herep(θ)is aproper prior density function, for example a Gaussian density. If we scale U(θ) by1/nand set Ω(θ) =−1 nlogp(θ)we recoverL(θ)in (1). Therefore exp(−U(θ))simply gives high probability to models which have low loss L(θ). Givenp(θ\|D)wepredict on a new instance xby averaging over all likely models, p(y\|x,D) =∫ p(y\|x,θ)p(θ\|D)dθ, (3)
2301.13856.pdf	Simplex Random Features Isaac Reid1Krzysztof Choromanski* 2 3Valerii Likhosherstov1Adrian Weller* 1 4 Abstract We present Simplex Random Features (SimRFs), a new random feature (RF) mechanism for unbi- ased approximation of the softmax and Gaussian kernels by geometrical correlation of random pro- jection vectors. We prove that SimRFs provide the smallest possible mean square error (MSE) on unbiased estimates of these kernels among the class of weight-independent geometrically- coupled positive random feature (PRF) mech- anisms, substantially outperforming the previ- ously most accurate Orthogonal Random Fea- tures (ORFs, Yu et al., 2016) at no observable extra cost. We present a more computationally expensive SimRFs+ variant, which we prove is asymptotically optimal in the broader family of weight-dependent geometrical coupling schemes (which permit correlations between random vec- tor directions and norms). In extensive empirical studies, we show consistent gains provided by SimRFs in settings including pointwise kernel estimation, nonparametric classification and scal- able Transformers (Choromanski et al., 2020).1 1. Introduction Embedding methods, which project feature vectors into a new space, are ubiquitous in machine learning. The canoni- cal example is the Johnson-Lindenstrauss Transform (JLT) (Johnson, 1984; Dasgupta et al., 2010; Kane & Nelson, 2014; Kar & Karnick, 2012), where a collection of high- dimensional points is embedded in a much lower dimen- sional space whilst (approximately) preserving their metric relationships, e.g. distances and dot-products. Another ap- plication is found in kernel approximation (Liu et al., 2022; Yang et al., 2014; Pennington et al., 2015; Li et al., 2010), *Equal senior co-leads1University of Cambridge2Google 3Columbia University4Alan Turing Institute. Correspondence to: Isaac Reid <[email protected] >, Krzysztof Choromanski <[email protected] >. Proceedings of the 40thInternational Conference on Machine Learning , Honolulu, Hawaii, USA. PMLR 202, 2023. Copyright 2023 by the author(s). 1Code is available at https://github.com/isaac-reid/ simplex random features.where the nonlinear similarity measure (kernel) in the origi- nal space is translated to a linear kernel in the latent space. For example, a kernel K(·,·) :Rd×Rd→Rcan be approx- imated using so-called random features (RFs): randomised nonlinear transformations ϕ(·) :Rd→Rd′constructed such that K(x,y) =E[bK(x,y)],wherebK(x,y)def=ϕ(x)⊤ϕ(y). (1) Provided Kis stationary, meaning K(x,y) =K(x−y), we can use Bochner’s theorem to write K(x−y) =Z Rdp(w)eiw⊤(x−y)ddw, (2) where p(w)is the Fourier transform of K. IfKis posi- tive semidefinite, p(w)is non-negative so we can treat it as a probability density. This invites Monte Carlo (MC) sampling, yielding Random Fourier Features (RFFs) of the following form, where vectors wiare sampled from p(w), mis their number and ⊙denotes concatenation (Rahimi & Recht, 2007; 2008): ϕRFF(z)def=r 1 m(⊙m i=1[sin(w⊤ iz),cos(w⊤ iz)])⊤.(3) Furthermore, if Kis a Gaussian kernel, defined by Kgauss(x,y)def= exp( −∥x−y∥2 2 2), (4) random vectors wiare sampled from the multivariate Gaus- sian distribution N(0,Id). Another kernel, of key interest in Transformer architectures (Vaswani et al., 2017; Choro- manski et al., 2020), is the so-called softmax kernel : Ksmax(x,y)def= exp( x⊤y). (5) Since Kgauss(x,y) =Ksmax(x,y) exp(−x2 2−y2 2), RF mechanisms for the Gaussian kernel can be readily con- verted into the corresponding mechanism for softmax and vice versa (Likhosherstov et al., 2022). Our results will hence apply to both settings. For brevity, we will mostly refer to Kgauss . However, as noted in (Choromanski et al., 2020), RFFs lead to unstable training of implicit linear-attention Transformers. 1arXiv:2301.13856v2 [stat.ML] 7 Oct 2023
2024.02.06.579080v1.full.pdf	Direct Coupling Analysis and the Attention Mechanism 1 Francesco Caredda1†and Andrea Pagnani1,2,3† 2 1DISAT, Politecnico di Torino, Corso Duca degli Abruzzi, 24, I-10129, Torino, Italy 3 2Italian Institute for Genomic Medicine, IRCCS Candiolo, SP-142, I-10060, 4 Candiolo, Italy 5 3INFN, Sezione di Torino, Torino, Via Pietro Giuria, 1 10125 Torino Italy 6 Abstract 7 Proteins serve as the foundation for nearly all biological functions within cells, encompassing 8 roles in transport, signaling, enzymatic activity, and more. Their functionalities hinge signifi- 9 cantly on their intricate three-dimensional structures, often posing challenges in terms of diffi- 10 culty, time, and expense for accurate determination. The introduction of AlphaFold 2 marked 11 a groundbreaking solution to the enduring challenge of predicting a protein’s tertiary structure 12 from its amino acid sequence. However, the inherent complexity of AlphaFold’s architecture 13 presents obstacles in deciphering its learning process and understanding the decision-making 14 that ultimately shapes the protein’s final structure. 15 In this study, we introduce a shallow, unsupervised model designed to understand the self- 16 attention layer within the Evoformer block of AlphaFold. We establish a method based on Direct 17 Coupling Analysis (DCA), wherein the interaction tensor undergoes decomposition, leveraging 18 the same structure employed in Transformer architectures. The model’s parameters, notably 19 fewer than those in standard DCA, are interpretable through an examination of the resulting 20 attention matrices. These matrices enable the extraction of contact information, subsequently 21 utilized for constructing the contact map of a protein family. Additionally, the self-attention 22 decomposition in the DCA Hamiltonian form adopted here facilitates the definition of multi- 23 family learning architecture, enabling the inference of parameter sets shared across diverse 24 protein families. Finally, an autoregressive generative version of the model is implemented, 25 capable of efficiently generating new proteins in silico. This generative model reproduces the 26 summary statistics of the original protein family while concurrently inferring direct contacts in 27 the tertiary structure of the protein. The effectiveness of our Attention-Based DCA architecture 28 is evaluated using Multiple Sequence Alignments (MSAs) of varying lengths and depths, with 29 structural data sourced from the Pfam database. 30 1 Introduction 31 Proteins constitute a diverse category of biological compounds constructed from a set of 20 amino 32 acids. Within an organism, they serve various functions, including structural support, mobility, 33 1. CC-BY 4.0 International license available under a(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprint this version posted February 8, 2024. ; https://doi.org/10.1101/2024.02.06.579080doi: bioRxiv preprint
2307.08691.pdf	FlashAttention-2 : Faster Attention with Better Parallelism and Work Partitioning Tri Dao1,2 1Department of Computer Science, Princeton University 2Department of Computer Science, Stanford University [email protected] July 18, 2023 Abstract Scaling Transformers to longer sequence lengths has been a major problem in the last several years, promising to improve performance in language modeling and high-resolution image understanding, as well as to unlock new applications in code, audio, and video generation. The attention layer is the main bottleneck in scaling to longer sequences, as its runtime and memory increase quadratically in the sequence length. FlashAttention [5] exploits the asymmetric GPU memory hierarchy to bring significant memory saving (linear instead of quadratic) and runtime speedup (2-4 ×compared to optimized baselines), with no approximation. However, FlashAttention is still not nearly as fast as optimized matrix-multiply (GEMM) operations, reaching only 25-40% of the theoretical maximum FLOPs/s. We observe that the inefficiency is due to suboptimal work partitioning between different thread blocks and warps on the GPU, causing either low-occupancy or unnecessary shared memory reads/writes. We propose FlashAttention-2 , with better work partitioning to address these issues. In particular, we (1) tweak the algorithm to reduce the number of non-matmul FLOPs (2) parallelize the attention computation, even for a single head, across different thread blocks to increase occupancy, and (3) within each thread block, distribute the work between warps to reduce communication through shared memory. These yield around 2×speedup compared to FlashAttention , reaching 50-73% of the theoretical maximum FLOPs/s on A100 and getting close to the efficiency of GEMM operations. We empirically validate that when used end-to-end to train GPT-style models, FlashAttention-2 reaches training speed of up to 225 TFLOPs/s per A100 GPU (72% model FLOPs utilization).1 1 Introduction Scaling up the context length of Transformers [ 18] is a challenge, since the attention layer at their heart has runtime and memory requirements quadratic in the input sequence length. Ideally, we would like to go beyond the standard 2k sequence length limit to train models to understand books, high resolution images, and long-form videos. Just within the last year, there have been several language models with much longer context than before: GPT-4 [ 12] with context length 32k, MosaicML’s MPT with context length 65k, and Anthropic’s Claude with context length 100k. Emerging use cases such as long document querying and story writing have demonstrated a need for models with such long context. To reduce the computational requirement of attention on such long context, there have been numerous methods proposed to approximate attention [ 2,3,4,8,9,14,19,20]. Though these methods have seen some use cases, as far as we know, most large-scale training runs still use standard attention. Motivated by this, Dao et al. [5]proposed to reorder the attention computation and leverages classical techniques (tiling, recomputation) to significantly speed it up and reduce memory usage from quadratic to linear in sequence length. This yields 2-4 ×wall-clock time speedup over optimized baselines, up to 10-20 ×memory saving, 1FlashAttention-2 is available at https://github.com/Dao-AILab/flash-attention 1arXiv:2307.08691v1 [cs.LG] 17 Jul 2023
supplementary-gpsa.pdf	Supplementary Information for: Generative Capacity of Probabilistic Protein Sequence Models Francisco McGee Sandro Hauri Quentin Novinger Slobodan Vucetic Ronald M. Levy Vincenzo Carnevale Allan Haldane Supplementary Note 1 - sVAE implementation The standard variational autoencoder (sVAE) is a deep, symmetrical, and undercomplete autoencoder neural network composed of a separate encoder qφ(Z\|S)and decoder pθ(S\|Z)1, which map input sequences Sinto regions of a low-dimensional latent space Zand back (see Fig. S1). It is a probabilistic model, and in our “vanilla”2 implementation we assume sequences will be distributed according to a unit normal distribution in latent space, p(Z) =N[0,1](Z)3. Training of a VAE can be understood as maximization of (the logarithm of) the dataset likelihood L=∏Spθ(S) =∑S∫pθ(S\|Z)p(Z)dZwith the addition of a Kullback-Leibler regularization term DKL[qφ(Z\|S),pθ(Z\|S)], where pθ(Z\|S)is the posterior of the decoder, which allows use of the ﬁtted encoder qφ(Z\|S)to perform efﬁcient estimation of the likelihood and its gradient by Monte-Carlo sampling, for appropriate encoder models. The sVAE architecture is built on the same basic VAE architecture of “EVOVAE”4, which itself appears to be built on the VAE implementation provided by developers for the Keras library5, and this same VAE architecture is used for each protein presented in this work. Similarly to EVOVAE, sVAE’s hyperparameters were tuned using grid search. sVAE is composed of 3symmetrical ELU-activated layers in both the encoder and decoder, each layer with 250dense (fully-connected) nodes. The encoder and decoder are connected by a latent layer of lnodes, and we use l=7in the main text. We provide further justiﬁcation for the selection of l=7elsewhere in the Supplementary Note 3. sVAE’s input layer accepts one-hot encoded sequences, the output layer is sigmoid-activated, and its node output values can be interpreted as a Bernoulli distribution of the same dimensions as a one-hot encoded sequence. The ﬁrst layer of the encoder and the middle layer of the decoder have dropout regularization applied with 30% dropout rate, and the middle layer of the encoder uses batch normalization4,6,7. In all inferences, we hold out 10% of the training sequences as a validation dataset, and perform maximum likelihood optimization using the Keras Adam stochastic gradient optimizer on the remaining 90%8, using mini-batch gradient descent with a batch size of 200. After each training epoch we evaluate the loss function for the training and validation data subsets separately. We have tested using early-stopping regularization to stop inference once the validation loss has not decreased for three epochs in a row, as in previous implementations, but this led to some variability in the model depending on when the early stopping criterion was reached. To avoid this variability, and to make different models more directly comparable, we instead ﬁxed the number of epochs to 32for all models, since in the early stopping tests this led to near minimum training loss and validation loss, and did not lead to signiﬁcant overﬁtting as would be apparent from an increase in the validation loss. sVAE was implemented using Keras, building on previous implementations4,5, however with a modiﬁcation of the loss function relative to both of these, to remove a scaling factor of Lqon the reconstruction loss, which is sometimes used to avoid issues with local minima as described further below. This prefactor leads to a non-unit variance of the latent space distribution of the dataset sequences, violating our deﬁnition that the latent space distribution should be normal with unit variance, p(Z) =N[0,1](Z). In the next section we show that after removing the prefactor the latent space distribution is approximately a unit normal, which more closely follows the original VAE conception3,9. Our implementation is available at https://github.com/ahaldane/MSA_VAE . To generate a sequence from the model we generate a random sample in latent space from the latent distribution N[0,1], and pass this value to the decoder to obtain a Bernoulli distribution, from which we sample once. 1/20
2010.02502.pdf	Published as a conference paper at ICLR 2021 DENOISING DIFFUSION IMPLICIT MODELS Jiaming Song, Chenlin Meng & Stefano Ermon Stanford University {tsong,chenlin,ermon }@cs.stanford.edu ABSTRACT Denoising diffusion probabilistic models (DDPMs) have achieved high qual- ity image generation without adversarial training, yet they require simulating a Markov chain for many steps in order to produce a sample. To accelerate sam- pling, we present denoising diffusion implicit models (DDIMs), a more efﬁcient class of iterative implicit probabilistic models with the same training procedure as DDPMs. In DDPMs, the generative process is deﬁned as the reverse of a particular Markovian diffusion process. We generalize DDPMs via a class of non-Markovian diffusion processes that lead to the same training objective. These non-Markovian processes can correspond to generative processes that are deterministic, giving rise to implicit models that produce high quality samples much faster. We empirically demonstrate that DDIMs can produce high quality samples 10×to50×faster in terms of wall-clock time compared to DDPMs, allow us to trade off computation for sample quality, perform semantically meaningful image interpolation directly in the latent space, and reconstruct observations with very low error. 1 I NTRODUCTION Deep generative models have demonstrated the ability to produce high quality samples in many domains (Karras et al., 2020; van den Oord et al., 2016a). In terms of image generation, genera- tive adversarial networks (GANs, Goodfellow et al. (2014)) currently exhibits higher sample quality than likelihood-based methods such as variational autoencoders (Kingma & Welling, 2013), autore- gressive models (van den Oord et al., 2016b) and normalizing ﬂows (Rezende & Mohamed, 2015; Dinh et al., 2016). However, GANs require very speciﬁc choices in optimization and architectures in order to stabilize training (Arjovsky et al., 2017; Gulrajani et al., 2017; Karras et al., 2018; Brock et al., 2018), and could fail to cover modes of the data distribution (Zhao et al., 2018). Recent works on iterative generative models (Bengio et al., 2014), such as denoising diffusion prob- abilistic models (DDPM, Ho et al. (2020)) and noise conditional score networks (NCSN, Song & Ermon (2019)) have demonstrated the ability to produce samples comparable to that of GANs, with- out having to perform adversarial training. To achieve this, many denoising autoencoding models are trained to denoise samples corrupted by various levels of Gaussian noise. Samples are then produced by a Markov chain which, starting from white noise, progressively denoises it into an im- age. This generative Markov Chain process is either based on Langevin dynamics (Song & Ermon, 2019) or obtained by reversing a forward diffusion process that progressively turns an image into noise (Sohl-Dickstein et al., 2015). A critical drawback of these models is that they require many iterations to produce a high quality sample. For DDPMs, this is because that the generative process (from noise to data) approximates the reverse of the forward diffusion process (from data to noise), which could have thousands of steps; iterating over all the steps is required to produce a single sample, which is much slower compared to GANs, which only needs one pass through a network. For example, it takes around 20 hours to sample 50k images of size 32×32from a DDPM, but less than a minute to do so from a GAN on a Nvidia 2080 Ti GPU. This becomes more problematic for larger images as sampling 50k images of size 256×256could take nearly 1000 hours on the same GPU. To close this efﬁciency gap between DDPMs and GANs, we present denoising diffusion implicit models (DDIMs). DDIMs are implicit probabilistic models (Mohamed & Lakshminarayanan, 2016) and are closely related to DDPMs, in the sense that they are trained with the same objective function. 1arXiv:2010.02502v4 [cs.LG] 5 Oct 2022
1901.09321.pdf	Published as a conference paper at ICLR 2019 FIXUP INITIALIZATION : RESIDUAL LEARNING WITHOUT NORMALIZATION Hongyi Zhang∗ MIT [email protected] N. Dauphin† Google Brain [email protected] Ma‡ Stanford University [email protected] ABSTRACT Normalization layers are a staple in state-of-the-art deep neural network archi- tectures. They are widely believed to stabilize training, enable higher learning rate, accelerate convergence and improve generalization, though the reason for their effectiveness is still an active research topic. In this work, we challenge the commonly-held beliefs by showing that none of the perceived beneﬁts is unique to normalization. Speciﬁcally, we propose ﬁxed-update initialization (Fixup), an initialization motivated by solving the exploding and vanishing gradient problem at the beginning of training via properly rescaling a standard initialization. We ﬁnd training residual networks with Fixup to be as stable as training with nor- malization — even for networks with 10,000 layers. Furthermore, with proper regularization, Fixup enables residual networks without normalization to achieve state-of-the-art performance in image classiﬁcation and machine translation. 1 I NTRODUCTION Artiﬁcial intelligence applications have witnessed major advances in recent years. At the core of this revolution is the development of novel neural network models and their training techniques. For example, since the landmark work of He et al. (2016), most of the state-of-the-art image recognition systems are built upon a deep stack of network blocks consisting of convolutional layers and additive skip connections, with some normalization mechanism (e.g., batch normalization (Ioffe & Szegedy, 2015)) to facilitate training and generalization. Besides image classiﬁcation, various normalization techniques (Ulyanov et al., 2016; Ba et al., 2016; Salimans & Kingma, 2016; Wu & He, 2018) have been found essential to achieving good performance on other tasks, such as machine translation (Vaswani et al., 2017) and generative modeling (Zhu et al., 2017). They are widely believed to have multiple beneﬁts for training very deep neural networks, including stabilizing learning, enabling higher learning rate, accelerating convergence, and improving generalization. Despite the enormous empirical success of training deep networks with normalization, and recent progress on understanding the working of batch normalization (Santurkar et al., 2018), there is currently no general consensus on why these normalization techniques help training residual neural networks. Intrigued by this topic, in this work we study (i)without normalization, can a deep residual network be trained reliably? (And if so,) (ii)without normalization, can a deep residual network be trained with the same learning rate, converge at the same speed, and generalize equally well (or even better)? Perhaps surprisingly, we ﬁnd the answers to both questions are Yes. In particular, we show: •Why normalization helps training. We derive a lower bound for the gradient norm of a residual network at initialization, which explains why with standard initializations, normalization tech- niques are essential for training deep residual networks at maximal learning rate. (Section 2) ∗Work done at Facebook. Equal contribution. †Work done at Facebook. Equal contribution. ‡Work done at Facebook. 1arXiv:1901.09321v2 [cs.LG] 12 Mar 2019
2402.03300.pdf	DeepSeekMath: Pushing the Limits of Mathematical Reasoning in Open Language Models Zhihong Shao1,2∗†, Peiyi Wang1,3∗†, Qihao Zhu1,3∗†, Runxin Xu1, Junxiao Song1 Mingchuan Zhang1, Y.K. Li1, Y. Wu1, Daya Guo1∗ 1DeepSeek-AI,2Tsinghua University,3Peking University {zhihongshao,wangpeiyi,zhuqh,guoday}@deepseek.com https://github.com/deepseek-ai/DeepSeek-Math Abstract Mathematical reasoning poses a significant challenge for language models due to its complex and structured nature. In this paper, we introduce DeepSeekMath 7B, which continues pre- training DeepSeek-Coder-Base-v1.5 7B with 120B math-related tokens sourced from Common Crawl, together with natural language and code data. DeepSeekMath 7B has achieved an impressive score of 51.7% on the competition-level MATH benchmark without relying on external toolkits and voting techniques, approaching the performance level of Gemini-Ultra and GPT-4. Self-consistency over 64 samples from DeepSeekMath 7B achieves 60.9% on MATH. The mathematical reasoning capability of DeepSeekMath is attributed to two key factors: First, we harness the significant potential of publicly available web data through a meticulously engineered data selection pipeline. Second, we introduce Group Relative Policy Optimization (GRPO), a variant of Proximal Policy Optimization (PPO), that enhances mathematical reasoning abilities while concurrently optimizing the memory usage of PPO. Figure 1\|Top1 accuracy of open-source models on the competition-level MATH benchmark (Hendrycks et al., 2021) without the use of external toolkits and voting techniques. ∗Core contributors. †Work done during internship at DeepSeek-AI.arXiv:2402.03300v1 [cs.CL] 5 Feb 2024
2111.02080.pdf	An Explanation of In-context Learning as Implicit Bayesian Inference Sang Michael Xie Stanford University [email protected] Raghunathan Stanford University [email protected] Percy Liang Stanford University [email protected] Ma Stanford University [email protected] Abstract Large language models (LMs) such as GPT-3 have the surprising ability to do in-context learning, where the model learns to do a downstream task simply by conditioning on a prompt consisting of input-output examples. The LM learns from these examples without being explicitly pretrained to learn . Thus, it is unclear what enables in-context learning. In this paper, we study how in-context learning can emerge when pretraining documents have long-range coherence. Here, the LM must infer a latent document-level concept to generate coherent next tokens during pretraining. At test time, in-context learning occurs when the LM also infers a shared latent concept between examples in a prompt. We prove when this occurs despite a distribution mismatch between prompts and pretraining data in a setting where the pretraining distribution is a mixture of HMMs. In contrast to messy large-scale datasets used to train LMs capable of in-context learning, we generate a small-scale synthetic dataset (GINC) where Transformers and LSTMs both exhibit in-context learning1. Beyond the theory, experiments on GINC exhibit large-scale real-world phenomena including improved in-context performance with model scaling (despite the same pretraining loss), sensitivity to example order, and instances where zero-shot is better than few-shot in-context learning. 1 Introduction Large language models (LMs) such as GPT-3 (Brown et al., 2020, Lieber et al., 2021, Radford et al., 2019, Wang and Komatsuzaki, 2021) are pretrained on massive text corpora to predict the next word given previous words. They demonstrate the surprising ability to do in-context learning , where an LM “learns” to do a task simply by conditioning on a prompt containing input-output pairs, achiev- ing SOTA results on LAMBADA (Paperno et al., 2016) and TriviaQA (Joshi et al., 2017) tasks (18% and 3% over previous SOTA (Brown et al., 2020)). For example, consider the task of predicting nationalities from names. A prompt (Figure 1) is constructed by concatenating independent “train- ing” examples (e.g., “Albert Einstein was German”) followed by a “test example” (“Marie Curie was”). Conditioning on this prompt, GPT-3 places the largest probability on the correct output p(“Polish”\|“Albert Einstein was German \n Mahatma Gandhi was Indian \n Marie Curie was” ) 1The code, data, and experiments are located on GitHub and CodaLab. 1arXiv:2111.02080v6 [cs.CL] 21 Jul 2022
2404.16014v1.pdf	2024-4-25 Improving Dictionary Learning with Gated Sparse Autoencoders Senthooran Rajamanoharan, Arthur Conmy, Lewis Smith, Tom Lieberum†, Vikrant Varma†, János Kramár, Rohin Shah and Neel Nanda *: Joint contribution.†: Core infrastructure contributor. Recent work has found that sparse autoencoders (SAEs) are an effective technique for unsupervised discovery of interpretable features in language models’ (LMs) activations, by finding sparse, linear reconstructions of LM activations. We introduce the Gated Sparse Autoencoder (Gated SAE), which achieves a Pareto improvement over training with prevailing methods. In SAEs, the L1 penalty used to encourage sparsity introduces many undesirable biases, such as shrinkage – systematic underestimation of feature activations. The key insight of Gated SAEs is to separate the functionality of (a) determining which directions to use and (b) estimating the magnitudes of those directions: this enables us to apply the L1 penalty only to the former, limiting the scope of undesirable side effects. Through training SAEs on LMs of up to 7B parameters we find that, in typical hyper-parameter ranges, Gated SAEs solve shrinkage, are similarly interpretable, and require half as many firing features to achieve comparable reconstruction fidelity. 1. Introduction Mechanistic interpretability research aims to explain how neural networks produce outputs in terms of the learned algorithms executed during a forward pass (Olah, 2022; Olah et al., 2020). Much work makes use of the fact that many concept representations appear to be linear (Elhage et al., 2021; Gurnee et al., 2023; Olah et al., 2020; Park et al., 2023). However, finding the set of all interpretable directions is a highly non-trivial problem. Classic approaches, like interpreting neurons (i.e. directions in the standard basis) are insufficient, as many are polysemantic and tend to activate for a range of different seemingly unrelated concepts (Bolukbasi et al., 2021; Elhage et al., 2022a,b, Empirical Phenomena). Thesuperpositionhypothesis (Elhageetal.,2022b, DefinitionsandMotivation)positsamechanistic explanation for these observations: in an intermediate representation of dimension 𝑛, a model will encode𝑀≫𝑛concepts as linear directions, where the set of concepts and their directions is fixed across all inputs, but on a given input only a sparse subset of concepts are active, ensuring that there is not much simultaneous interference (Gurnee et al., 2023, Appendix A) between these (non-orthogonal) concepts. Motivated by the superposition hypothesis, Bricken et al. (2023) and Cunningham et al. (2023) recently used sparse autoencoders (SAEs; Ng (2011)) to find sparse decompositions of model activations in terms of an overcomplete basis, or dictionary (Mallat and Zhang, 1993).1 Although SAEs show promise in this regard, the L1 penalty used in the prevailing training method to encourage sparsity also introduces biases that harm the accuracy of SAE reconstructions, as the loss can be decreased by trading-off some reconstruction accuracy for lower L1. We refer to this 1Although motivated by the superposition hypothesis, the utility of line of research is not contingent on this hypothesis being true. If a faithful, sparse and interpretable decomposition can be found, we expect this to be a useful basis in its own right for downstream interpretability tasks, such as understanding or intervening on a model’s representations and circuits, even if some fraction of the model’s computation is e.g. represented non-linearly and not captured. Corresponding authors: [email protected] and [email protected] ©2024 Google DeepMind. All rights reserved arXiv:2404.16014v1 [cs.LG] 24 Apr 2024
1811.07871.pdf	Scalable agent alignment via reward modeling: a research direction Jan Leike DeepMindDavid Krueger∗ DeepMind MilaTom Everitt DeepMindMiljan Martic DeepMindVishal Maini DeepMindShane Legg DeepMind Abstract One obstacle to applying reinforcement learning algorithms to real-world problems is the lack of suitable reward functions. Designing such reward functions is difﬁcult in part because the user only has an implicit understanding of the task objective. This gives rise to the agent alignment problem : how do we create agents that behave in accordance with the user’s intentions? We outline a high-level research direction to solve the agent alignment problem centered around reward modeling : learning a reward function from interaction with the user and optimizing the learned reward function with reinforcement learning. We discuss the key challenges we expect to face when scaling reward modeling to complex and general domains, concrete approaches to mitigate these challenges, and ways to establish trust in the resulting agents. 1 Introduction Games are a useful benchmark for research because progress is easily measurable. Atari games come with a score function that captures how well the agent is playing the game; board games or competitive multiplayer games such as Dota 2 and Starcraft II have a clear winner or loser at the end of the game. This helps us determine empirically which algorithmic and architectural improvements work best. However, the ultimate goal of machine learning (ML) research is to go beyond games and improve human lives. To achieve this we need ML to assist us in real-world domains, ranging from simple tasks like ordering food or answering emails to complex tasks like software engineering or running a business. Yet performance on these and other real-world tasks is not easily measurable, since they do not come readily equipped with a reward function. Instead, the objective of the task is only indirectly available through the intentions of the human user. This requires walking a ﬁne line. On the one hand, we want ML to generate creative and brilliant solutions like AlphaGo’s Move 37 (Metz, 2016)—a move that no human would have recommended, yet it completely turned the game in AlphaGo’s favor. On the other hand, we want to avoid degenerate solutions that lead to undesired behavior like exploiting a bug in the environment simulator (Clark & Amodei, 2016; Lehman et al., 2018). In order to differentiate between these two outcomes, our agent needs to understand its user’s intentions , and robustly achieve these intentions with its behavior. We frame this as the agent alignment problem : How can we create agents that behave in accordance with the user’s intentions? With this paper we outline a research direction to solve the agent alignment problem. We build on taxonomies and problem deﬁnitions from many authors before us, highlighting tractable and neglected problems in the ﬁeld of AI safety (Russell et al., 2015; Soares, 2015; Amodei et al., 2016; ∗Work done during an internship at DeepMind.arXiv:1811.07871v1 [cs.LG] 19 Nov 2018
1803.03635.pdf	Published as a conference paper at ICLR 2019 THELOTTERY TICKET HYPOTHESIS : FINDING SPARSE , TRAINABLE NEURAL NETWORKS Jonathan Frankle MIT CSAIL [email protected] Carbin MIT CSAIL [email protected] ABSTRACT Neural network pruning techniques can reduce the parameter counts of trained net- works by over 90%, decreasing storage requirements and improving computational performance of inference without compromising accuracy. However, contemporary experience is that the sparse architectures produced by pruning are difﬁcult to train from the start, which would similarly improve training performance. We ﬁnd that a standard pruning technique naturally uncovers subnetworks whose initializations made them capable of training effectively. Based on these results, we articulate the lottery ticket hypothesis : dense, randomly-initialized, feed-forward networks contain subnetworks ( winning tickets ) that—when trained in isolation— reach test accuracy comparable to the original network in a similar number of iterations. The winning tickets we ﬁnd have won the initialization lottery: their connections have initial weights that make training particularly effective. We present an algorithm to identify winning tickets and a series of experiments that support the lottery ticket hypothesis and the importance of these fortuitous initializations. We consistently ﬁnd winning tickets that are less than 10-20% of the size of several fully-connected and convolutional feed-forward architectures for MNIST and CIFAR10. Above this size, the winning tickets that we ﬁnd learn faster than the original network and reach higher test accuracy. 1 I NTRODUCTION Techniques for eliminating unnecessary weights from neural networks ( pruning ) (LeCun et al., 1990; Hassibi & Stork, 1993; Han et al., 2015; Li et al., 2016) can reduce parameter-counts by more than 90% without harming accuracy. Doing so decreases the size (Han et al., 2015; Hinton et al., 2015) or energy consumption (Yang et al., 2017; Molchanov et al., 2016; Luo et al., 2017) of the trained networks, making inference more efﬁcient. However, if a network can be reduced in size, why do we not train this smaller architecture instead in the interest of making training more efﬁcient as well? Contemporary experience is that the architectures uncovered by pruning are harder to train from the start, reaching lower accuracy than the original networks.1 Consider an example. In Figure 1, we randomly sample and train subnetworks from a fully-connected network for MNIST and convolutional networks for CIFAR10. Random sampling models the effect of the unstructured pruning used by LeCun et al. (1990) and Han et al. (2015). Across various levels of sparsity, dashed lines trace the iteration of minimum validation loss2and the test accuracy at that iteration. The sparser the network, the slower the learning and the lower the eventual test accuracy. 1“Training a pruned model from scratch performs worse than retraining a pruned model, which may indicate the difﬁculty of training a network with a small capacity.” (Li et al., 2016) “During retraining, it is better to retain the weights from the initial training phase for the connections that survived pruning than it is to re-initialize the pruned layers...gradient descent is able to ﬁnd a good solution when the network is initially trained, but not after re-initializing some layers and retraining them.” (Han et al., 2015) 2As a proxy for the speed at which a network learns, we use the iteration at which an early-stopping criterion would end training. The particular early-stopping criterion we employ throughout this paper is the iteration of minimum validation loss during training. See Appendix C for more details on this choice. 1arXiv:1803.03635v5 [cs.LG] 4 Mar 2019
2002.10957v2.pdf	MINILM: Deep Self-Attention Distillation for Task-Agnostic Compression of Pre-Trained Transformers Wenhui Wang Furu Wei Li Dong Hangbo Bao Nan Yang Ming Zhou Microsoft Research {wenwan,fuwei,lidong1,t-habao,nanya,mingzhou}@microsoft.com Abstract Pre-trained language models (e.g., BERT (Devlin et al., 2018) and its variants) have achieved re- markable success in varieties of NLP tasks. How- ever, these models usually consist of hundreds of millions of parameters which brings challenges for ﬁne-tuning and online serving in real-life ap- plications due to latency and capacity constraints. In this work, we present a simple and effective ap- proach to compress large Transformer (Vaswani et al., 2017) based pre-trained models, termed as deep self-attention distillation. The small model (student) is trained by deeply mimicking the self-attention module, which plays a vital role in Transformer networks, of the large model (teacher). Speciﬁcally, we propose distilling the self-attention module of the last Transformer layer of the teacher, which is effective and ﬂexible for the student. Furthermore, we introduce the scaled dot-product between values in the self-attention module as the new deep self-attention knowledge, in addition to the attention distributions (i.e., the scaled dot-product of queries and keys) that have been used in existing works. Moreover, we show that introducing a teacher assistant (Mirzadeh et al., 2019) also helps the distillation of large pre-trained Transformer models. Experimental results demonstrate that our monolingual model1 outperforms state-of-the-art baselines in different parameter size of student models. In particular, it retains more than 99% accuracy on SQuAD 2.0 and several GLUE benchmark tasks using 50% of the Transformer parameters and computations of the teacher model. We also obtain competitive results in applying deep self-attention distillation to multilingual pre-trained models. Correspondence to: Furu Wei <[email protected]>. 1The code and models are publicly available at https:// aka.ms/minilm .1. Introduction Language model (LM) pre-training has achieved remarkable success for various natural language processing tasks (Pe- ters et al., 2018; Howard & Ruder, 2018; Radford et al., 2018; Devlin et al., 2018; Dong et al., 2019; Yang et al., 2019; Joshi et al., 2019; Liu et al., 2019). The pre-trained language models, such as BERT (Devlin et al., 2018) and its variants, learn contextualized text representations by predict- ing words given their context using large scale text corpora, and can be ﬁne-tuned with additional task-speciﬁc layers to adapt to downstream tasks. However, these models usually contain hundreds of millions of parameters which brings challenges for ﬁne-tuning and online serving in real-life applications for latency and capacity constraints. Knowledge distillation (Hinton et al., 2015; Romero et al., 2015) (KD) has been proven to be a promising way to com- press a large model (called the teacher model) into a small model (called the student model), which uses much fewer parameters and computations while achieving competitive results on downstream tasks. There have been some works that task-speciﬁcally distill pre-trained large LMs into small models (Tang et al., 2019; Turc et al., 2019b; Sun et al., 2019a; Aguilar et al., 2019). They ﬁrst ﬁne-tune the pre- trained LMs on speciﬁc tasks and then perform distillation. Task-speciﬁc distillation is effective, but ﬁne-tuning large pre-trained models is still costly, especially for large datasets. Different from task-speciﬁc distillation, task-agnostic LM distillation mimics the behavior of the original pre-trained LMs and the student model can be directly ﬁne-tuned on downstream tasks (Tsai et al., 2019; Sanh et al., 2019; Jiao et al., 2019; Sun et al., 2019b). Previous works use soft target probabilities for masked lan- guage modeling predictions or intermediate representations of the teacher LM to guide the training of the task-agnostic student. DistilBERT (Sanh et al., 2019) employs a soft-label distillation loss and a cosine embedding loss, and initializes the student from the teacher by taking one layer out of two. But each Transformer layer of the student is required to have the same architecture as its teacher. TinyBERT (Jiao et al., 2019) and MOBILE BERT (Sun et al., 2019b) utilizearXiv:2002.10957v2 [cs.CL] 6 Apr 2020
reka-vibe-eval.pdf	Vibe-Eval: A hard evaluation suite for measuring progress of multimodal language models Piotr Padlewski∗Max Bain∗Matthew Henderson Zhongkai Zhu Nishant Relan Hai Pham Donovan Ong Kaloyan Aleksiev Aitor Ormazabal Samuel Phua Ethan Yeo Eugenie Lamprecht Qi Liu Yuqi Wang Eric Chen Deyu Fu Lei Li Che Zheng Cyprien de Masson d’Autume Dani Yogatama Mikel Artetxe Yi Tay Reka What does this image say? There is a message hidden in the image that says " GOOD VIBES ONLY ". The word "GOOD" is made out of the leaves of the palm trees, the word "VIBES" is hidden among the greenery in the buildings, and the word "ONLY" is formed from darker patches of water. Golden Reference The image features the word "ONLY" written in large letters on the sandy beach. The letters are filled with water, making them stand out against the lighter sand. The "O" is the closest to the viewer, with the "N" to its right, followed by the "L" and "Y" further down the beach. Reka Core The assistant response only identified the word "only" in the sand, while the ground truth response revealed the complete hidden message "GOOD VIBES ONLY" using various elements in the image, such as palm trees, greenery in the buildings, and darker patches of water. Therefore, the assistant response did not correctly match the ground truth. Rating: 2 / 5 Evaluator Figure 1: An example from our Vibe-Eval benchmark. This prompt is from the hard-set, containing only difficultprompts. Theuserpromptandimageisshownontheleft,andontherightisthehuman-written golden reference response, a generation from Reka Core, and output from the automatic evaluator. Abstract We introduce Vibe-Eval: a new open benchmark and framework for evaluating multimodal chat models. Vibe-Eval consists of 269 visual understanding prompts, including 100 of harddifficulty, and complete with gold-standardresponsesauthoredbyexperts. Vibe-Evalisopen-endedandchallengingwithdualobjectives: (i) vibe checking multimodal chat models for day-to-day tasks and (ii) deeply challenging and probing the capabilities of present frontier models. To this end, our hard set contains >50%questions that allfrontier models answerincorrectly. We explorethe nuancesof designing, evaluating,and ranking modelson ultra challenging prompts. We also discuss trade-offs between human and automatic evaluation, and show that automaticmodelevaluationusingRekaCoreroughlycorrelatestohumanjudgment. WehavemadeAPI access free-of-charge for the purpose of lightweight evaluation with the intention to run formal human evaluations for public models that do well on the Vibe-Eval automatic scores. We release the evaluation code and data at github.com/reka-ai/reka-vibe-eval. ∗Equal contribution
10.1016.j.cell.2023.12.012.pdf	Article Human fetal brain self-organizes into long-term expanding organoids Graphical abstract Highlights dHuman fetal brain organoids (FeBOs) display cellular heterogeneity and can be expanded dFeBOs produce a tissue-like ECM niche and enable ECMperturbation studies dDerivation of regional FeBOs allows the study of regionalmorphogen effects dCRISPR-engineered FeBOs are a scalable bottom-up tumormodeling platformAuthors Delilah Hendriks, Anna Pagliaro,Francesco Andreatta, ..., Wei Wu,Hans Clevers, Benedetta Artegiani Correspondence [email protected] (D.H.), [email protected] (H.C.),b.a.artegiani@prinsesmaximacentrum. nl (B.A.) In brief Tissue-derived human fetal brain organoids (FeBOs) are established from different anatomical regions that can beexpanded in culture. FeBOs may helpunderstand developmental and disease-related biology. Hendriks et al., 2024, Cell 187, 712–732 February 1, 2024 ª2023 The Author(s). Published by Elsevier Inc. https://doi.org/10.1016/j.cell.2023.12.012 ll
2009.01325v3.pdf	Learning to summarize from human feedback Nisan Stiennon∗Long Ouyang∗Jeff Wu∗Daniel M. Ziegler∗Ryan Lowe∗ Chelsea Voss∗Alec Radford Dario Amodei Paul Christiano∗ OpenAI Abstract As language models become more powerful, training and evaluation are increas- ingly bottlenecked by the data and metrics used for a particular task. For example, summarization models are often trained to predict human reference summaries and evaluated using ROUGE, but both of these metrics are rough proxies for what we really care about—summary quality. In this work, we show that it is possible to signiﬁcantly improve summary quality by training a model to optimize for human preferences. We collect a large, high-quality dataset of human comparisons be- tween summaries, train a model to predict the human-preferred summary, and use that model as a reward function to ﬁne-tune a summarization policy using reinforce- ment learning. We apply our method to a version of the TL;DR dataset of Reddit posts [ 63] and ﬁnd that our models signiﬁcantly outperform both human reference summaries and much larger models ﬁne-tuned with supervised learning alone. Our models also transfer to CNN/DM news articles [ 22], producing summaries nearly as good as the human reference without any news-speciﬁc ﬁne-tuning.2We con- duct extensive analyses to understand our human feedback dataset and ﬁne-tuned models.3We establish that our reward model generalizes to new datasets, and that optimizing our reward model results in better summaries than optimizing ROUGE according to humans. We hope the evidence from our paper motivates machine learning researchers to pay closer attention to how their training loss affects the model behavior they actually want. 1 Introduction Large-scale language model pretraining has become increasingly prevalent for achieving high per- formance on a variety of natural language processing (NLP) tasks. When applying these models to a speciﬁc task, they are usually ﬁne-tuned using supervised learning, often to maximize the log probability of a set of human demonstrations. While this strategy has led to markedly improved performance, there is still a misalignment between this ﬁne-tuning objective—maximizing the likelihood of human-written text—and what we care about—generating high-quality outputs as determined by humans. This misalignment has several causes: the maximum likelihood objective has no distinction between important errors (e.g. making up facts [ 41]) and unimportant errors (e.g. selecting the precise word from a set of synonyms); models ∗This was a joint project of the OpenAI Reﬂection team. Author order was randomized amongst {LO, JW, DZ, NS}; CV and RL were full-time contributors for most of the duration. PC is the team lead. 2Samples from all of our models can be viewed on our website. 3We provide inference code for our 1.3B models and baselines, as well as a model card and our human feedback dataset with over 64k summary comparisons, here. 34th Conference on Neural Information Processing Systems (NeurIPS 2020), Vancouver, Canada.arXiv:2009.01325v3 [cs.CL] 15 Feb 2022
2306.04751.pdf	How Far Can Camels Go? Exploring the State of Instruction Tuning on Open Resources Yizhong Wang∗♣♠Hamish Ivison∗♣Pradeep Dasigi♣Jack Hessel♣ Tushar Khot♣Khyathi Raghavi Chandu♣David Wadden♣Kelsey MacMillan♣ Noah A. Smith♣♠Iz Beltagy♣Hannaneh Hajishirzi♣♠ ♣Allen Institute for AI♠University of Washington {yizhongw,hamishi}@allenai.org Abstract In this work we explore recent advances in instruction-tuning language models on a range of open instruction-following datasets. Despite recent claims that open modelscanbeonparwithstate-of-the-artproprietarymodels,theseclaimsareoften accompaniedbylimitedevaluation,makingitdifficulttocomparemodelsacross the board and determine the utility of various resources. We provide a large set of instruction-tunedmodelsfrom6.7Bto65Bparametersinsize,trainedon12instruc- tion datasets ranging from manually curated (e.g., OpenAssistant) to synthetic and distilled(e.g.,Alpaca)andsystematicallyevaluatethemontheirfactualknowledge, reasoning, multilinguality, coding, safety, and open-ended instruction following abilities through a collection of automatic, model-based, and human-based metrics. We further introduce TÜLU ,our bestperforminginstruction-tuned modelsuite finetuned on a combination of high-quality open resources. Our experiments show that different instruction-tuning datasets can uncover or enhance specific skills, while no single dataset (or combination) provides the best performanceacrossallevaluations. Interestingly,wefindthatmodelandhuman preference-basedevaluationsfailtoreflectdifferencesinmodelcapabilitiesexposed by benchmark-based evaluations, suggesting the need for the type of systemic evaluation performed in this work. Our evaluations show that the best model in anygivenevaluationreachesonaverage87%ofChatGPTperformance,and73% ofGPT-4performance,suggestingthatfurtherinvestmentin buildingbetterbase models and instruction-tuning data is required to close the gap. We release our instruction-tuned models, including a fully finetuned 65B TÜLU , along with our code, data, and evaluation framework to facilitate future research.2 1 Introduction Thelatestgenerationoflargelanguagemodelshasbroughtunprecedentedattentiontothepotentialof language technologies. To support imperative user requests and a chat interface, these models often undergoan instruction-tuning stepwhich involves trainingonsupervised input/outputpairs. Recent instructiontuningcorporaareoftengatheredviacrowdsourcing(Dolly[ 12],OpenAssistant[ 26]) or via distillation from another model (Alpaca [ 43], Vicuna [ 8]). However, while some public, instruction-tuned models are advertised as comparable to powerful closed-source proprietary models such as ChatGPT, most experiments that support such claims only cover a small set of tasks, and mostlyrely onmodel-based evaluationmetrics [ 8,56]. We contendthatthe evaluationsetup should ∗Equal contribution. 2https://github.com/allenai/open-instruct 37th Conference on Neural Information Processing Systems (NeurIPS 2023) Track on Datasets and Benchmarks.arXiv:2306.04751v2 [cs.CL] 30 Oct 2023
2403.19887.pdf	Jamba: A Hybrid Transformer-Mamba Language Model Opher Lieber∗Barak Lenz∗Hofit Bata Gal Cohen Jhonathan Osin Itay Dalmedigos Erez Safahi Shaked Meirom Yonatan Belinkov Shai Shalev-Shwartz Omri Abend Raz Alon Tomer Asida Amir Bergman Roman Glozman Michael Gokhman Avashalom Manevich Nir Ratner Noam Rozen Erez Shwartz Mor Zusman Yoav Shoham Abstract We present Jamba, a new base large language model based on a novel hybrid Transformer-Mamba mixture-of-experts (MoE) architecture. Specifically, Jamba interleaves blocks of Transformer and Mamba layers, enjoying the benefits of both model families. MoE is added in some of these layers to increase model capacity while keeping active parameter usage manageable. This flexible architecture allows resource- and objective-specific configurations. In the particular configuration we have implemented, we end up with a powerful model that fits in a single 80GB GPU. Built at large scale, Jamba provides high throughput and small memory footprint compared to vanilla Transformers, and at the same time state-of-the-art performance on standard language model benchmarks and long-context evaluations. Remarkably, the model presents strong results for up to 256K tokens context length. We study various architectural decisions, such as how to combine Transformer and Mamba layers, and how to mix experts, and show that some of them are crucial in large scale modeling. We also describe several interesting properties of these architectures which the training and evaluation of Jamba have revealed, and plan to release checkpoints from various ablation runs, to encourage further exploration of this novel architecture. We make the weights of our implementation of Jamba publicly available under a permissive license. Model: https://huggingface.co/ai21labs/Jamba-v0.1 1 Introduction We introduce Jamba, a new publicly available large language model. Jamba is based on a novel hybrid architecture, which combines Transformer layers [ 46] with Mamba layers [ 16], a recent state-space model [ 17,18], as well as a mixture-of-experts (MoE) component [ 13,41]. Jamba thus combines two orthogonal architectural designs that together give it improved performance and higher throughput, while maintaining a manageable memory footprint. The 7B-based Jamba model (12B active parameters, 52B total available parameters) we are releasing was designed to fit in a single 80GB GPU, but the Jamba architecture supports other design choices, depending on one’s hardware and performance requirements. ∗Equal contribution.arXiv:2403.19887v1 [cs.CL] 28 Mar 2024
20-302.pdf	Journal of Machine Learning Research 22 (2021) 1-35 Submitted 3/20; Revised 10/20; Published 3/21 Attention is Turing Complete Jorge P´ erez [email protected] Department of Computer Science Universidad de Chile IMFD Chile Pablo Barcel´ o [email protected] Institute for Mathematical and Computational Engineering School of Engineering, Faculty of Mathematics Universidad Cat´ olica de Chile IMFD Chile Javier Marinkovic [email protected] Department of Computer Science Universidad de Chile IMFD Chile Editor: Luc de Raedt Abstract Alternatives to recurrent neural networks, in particular, architectures based on self-attention , are gaining momentum for processing input sequences. In spite of their relevance, the com- putational properties of such networks have not yet been fully explored. We study the computational power of the Transformer , one of the most paradigmatic architectures ex- emplifying self-attention. We show that the Transformer with hard-attention is Turing complete exclusively based on their capacity to compute and access internal dense repre- sentations of the data. Our study also reveals some minimal sets of elements needed to obtain this completeness result. Keywords: Transformers, Turing completeness, self-Attention, neural networks, arbi- trary precision 1. Introduction There is an increasing interest in designing neural network architectures capable of learning algorithms from examples (Graves et al., 2014; Grefenstette et al., 2015; Joulin and Mikolov, 2015; Kaiser and Sutskever, 2016; Kurach et al., 2016; Dehghani et al., 2018). A key requirement for any such an architecture is thus to have the capacity of implementing arbitrary algorithms, that is, to be Turing complete . Most of the networks proposed for learning algorithms are Turing complete simply by deﬁnition, as they can be seen as a control unit with access to an unbounded memory; as such, they are capable of simulating any Turing machine. On the other hand, the work by Siegelmann and Sontag (1995) has established a dif- ferent way of looking at the Turing completeness of neural networks. In particular, their work establishes that recurrent neural networks (RNNs) are Turing complete even if only a bounded number of resources (i.e., neurons and weights) is allowed. This is based on two c⃝2021 Jorge P´ erez, Pablo Barcel´ o, Javier Marinkovic. License: CC-BY 4.0, see https://creativecommons.org/licenses/by/4.0/ . Attribution requirements are provided athttp://jmlr.org/papers/v22/20-302.html .
Pretrained Transformers for Text Ranking: BERT and Beyond.pdf	Pretrained Transformers for Text Ranking: BERT and Beyond Jimmy Lin,1Rodrigo Nogueira,1and Andrew Yates2,3 1David R. Cheriton School of Computer Science, University of Waterloo 2University of Amsterdam 3Max Planck Institute for Informatics Version 0.99 — August 20, 2021 Abstract The goal of text ranking is to generate an ordered list of texts retrieved from a corpus in response to a query for a particular task. Although the most common formulation of text ranking is search, instances of the task can also be found in many text processing applications. This survey provides an overview of text ranking with neural network architectures known as transformers, of which BERT is the best-known example. The combination of transformers and self-supervised pretraining has been responsible for a paradigm shift in natural language processing (NLP), information retrieval (IR), and beyond. For text ranking, transformer-based models produce high quality results across many domains, tasks, and settings. This survey provides a synthesis of existing work as a single point of entry for practitioners who wish to deploy transformers for text ranking and researchers who wish to pursue work in this area. We cover a wide range of techniques, grouped into two categories: transformer models that perform reranking in multi-stage architectures and dense retrieval techniques that perform ranking directly. Examples in the ﬁrst category include approaches based on relevance classiﬁcation, evidence aggregation from multiple segments of text, and document and query expansion. The second category involves using transformers to learn dense representations of texts, where ranking is formulated as comparisons between query and document representations that take advantage of nearest neighbor search. At a high level, there are two themes that pervade our survey: techniques for handling long documents, beyond typical sentence-by-sentence processing in NLP, and techniques for addressing the tradeoff between effectiveness (i.e., result quality) and efﬁciency (e.g., query latency, model and index size). Much effort has been devoted to developing ranking models that address the mismatch between document lengths and the length limitations of existing transformers. The computational costs of inference with transformers has led to alternatives and variants that aim for different tradeoffs, both within multi-stage architectures as well as with dense learned representations. Although transformer architectures and pretraining techniques are recent innova- tions, many aspects of how they are applied to text ranking are relatively well understood and represent mature techniques. However, there remain many open research questions, and thus in addition to laying out the foundations of pretrained transformers for text ranking, this survey also attempts to prognosticate where the ﬁeld is heading.arXiv:2010.06467v3 [cs.IR] 19 Aug 2021
2212.10560.pdf	SELF-INSTRUCT : Aligning Language Model with Self Generated Instructions Yizhong Wang♣Yeganeh Kordi♢Swaroop Mishra♡Alisa Liu♣ Noah A. Smith♣+Daniel Khashabi♠Hannaneh Hajishirzi♣+ ♣University of Washington♢Tehran Polytechnic♡Arizona State University ♠Johns Hopkins University+Allen Institute for AI [email protected] Abstract Large “instruction-tuned” language models (ﬁnetuned to respond to instructions) have demonstrated a remarkable ability to gener- alize zero-shot to new tasks. Nevertheless, theydependheavilyonhuman-writteninstruc- tion data that is limited in quantity, diver- sity, and creativity, therefore hindering the generality of the tuned model. We intro- duce SELF-INSTRUCT , a framework for im- proving the instruction-following capabilities of pretrained language models by bootstrap- ping oﬀ its own generations. Our pipeline generates instruction, input, and output sam- ples from a language model, then prunes them before using them to ﬁnetune the orig- inal model. Applying our method to vanilla GPT3, we demonstrate a 33% absolute im- provement over the original model on S UPER- NATURALINSTRUCTIONS , on par with the per- formanceofInstructGPT0011,whichistrained with private user data and human annotations. For further evaluation, we curate a set of expert-writteninstructionsfornoveltasks,and show through human evaluation that tuning GPT3withS ELF-INSTRUCT outperformsusing existing public instruction datasets by a large margin,leavingonlya5%absolutegapbehind InstructGPT001. SELF-INSTRUCT provides an almostannotation-freemethodforaligningpre- trainedlanguagemodelswithinstructions,and we release our large synthetic dataset to facili- tate future studies on instruction tuning2. 1 Introduction TherecentNLPliteraturehaswitnessedatremen- dous amount of activity in building models that 1Unless otherwise speciﬁed, our comparisons are with the text-davinci-001 engine. We focus on this engine since it istheclosesttoourexperimentalsetup: supervisedﬁne-tuning with human demonstrations. The newer engines are more powerful,thoughtheyusemoredata(e.g.,codecompletionor latest user queries) oralgorithms(e.g., PPO) that are diﬃcult to compare with. 2Code and data will be available at https://github. com/yizhongw/self-instruct .can follow natural language instructions (Mishra et al., 2022; Wei et al., 2022; Sanh et al., 2022; Wangetal.,2022;Ouyangetal.,2022;Chungetal., 2022, i.a.). These developments are powered by two key components: large pre-trained language models(LM)andhuman-writteninstructiondata. PROMPTSOURCE(Bach et al., 2022) and SUPER- NATURALINSTRUCTIONS (Wang et al., 2022) are two notable recent datasets that use extensive man- ual annotation for collecting instructions to con- structT0(Bach et al., 2022; Sanh et al., 2022) and Tk-INSTRUCT (Wangetal.,2022). However,this process is costly and often suﬀers limited diver- sitygiventhatmosthumangenerationstendtobe popular NLP tasks, falling short of covering a true varietyoftasksanddiﬀerentwaystodescribethem. Given these limitations, continuing to improve the qualityofinstruction-tunedmodelsnecessitatesthe developmentofalternativeapproachesforsupervis- ing instruction-tuned models. In this work, we introduce SELF-INSTRUCT , a semi-automated process for instruction-tuning a pretrained LM using instructional signals from the model itself. The overall process is an iterative bootstrappingalgorithm(seeFigure1),whichstarts oﬀwithalimited(e.g.,175inourstudy)seedset of manually-written instructions that are used to guide the overallgeneration. In the ﬁrst phase, the model is prompted to generate instructions for new tasks. This step leverages the existing collection of instructions to create more broad-coverage instruc- tionsthatdeﬁne(oftennew)tasks. Giventhenewly- generated set of instructions, the framework also creates input-output instances for them, which can be later used for supervising the instruction tuning. Finally, various measures are used to prune low- quality and repeated instructions, before adding themtothetaskpool. Thisprocesscanberepeated formanyinteractionsuntilreachingalargenumber of tasks. Toevaluate SELF-INSTRUCT empirically,werunarXiv:2212.10560v1 [cs.CL] 20 Dec 2022
2310.18313.pdf	FP8-LM: Training FP8 Large Language Models Houwen Peng∗Kan Wu∗Yixuan Wei∗ Guoshuai Zhao Yuxiang Yang Ze Liu Yifan Xiong Ziyue Yang Bolin Ni Jingcheng Hu Ruihang Li Miaosen Zhang Chen Li Jia Ning Ruizhe Wang Zheng Zhang Shuguang Liu Joe Chau Han Hu†Peng Cheng† Microsoft Azure and Microsoft Research Abstract In this paper, we explore FP8 low-bit data formats for efficient training of large language models (LLMs). Our key insight is that most variables, such as gradients and optimizer states,inLLMtrainingcanemploylow-precisiondataformatswithoutcompromisingmodel accuracy and requiring no changes to hyper-parameters. Specifically, we propose a new FP8 automatic mixed-precision framework for training LLMs. This framework offers three levels ofFP8utilizationto streamlinemixed-precisionanddistributedparalleltrainingfor LLMs. It gradually incorporates 8-bit gradients, optimizer states, and distributed learning in an incremental manner. Experiment results show that, during the training of GPT-175B model onH100GPUplatform,ourFP8mixed-precisiontrainingframeworknotonlyachieveda remarkable 42% reduction in real memory usage but also ran 64% faster than the widely adopted BF16 framework (i.e., Megatron-LM), surpassing the speed of Nvidia Transformer Engine by 17%. This largely reduces the training costs for large foundation models. Fur- thermore,ourFP8mixed-precisiontrainingmethodologyisgeneric. Itcanbeseamlessly applied to other tasks such as LLM instruction tuning and reinforcement learning with human feedback, offering savings in fine-tuning expenses. Our FP8 low-precision training framework is open-sourced at aka.ms/MS.AMP. 1 Introduction Largelanguagemodels(LLMs)(Brownetal.,2020;Smithetal.,2022;Chowdheryetal.,2022; Zhangetal., 2022) have demonstrated unprecedented capabilities in language comprehension and generation, leading to breakthroughs in reasoning, math, science, and many other tasks (OpenAI, 2023; Anil et al., 2023). However, training LLMs is extremely costly. For example, PaLM takes 6,144 TPUv4 chips to train a 540B model, while GPT-3175Bconsumesseveralthousandpetaflop/s-daysofcomputeforpre-training(Chowdheryetal.,2022; Brownetal.,2020). ThismotivatestheneedsofreducingthetrainingcostsofLLMs,especiallyforthescaling of next-generation super-intelligent models. Low-precision training is one of the most promising directions to reduce the costs, as it can provide high speed, small memory footprint, and low communication overhead. Most existing training systems, e.g., Megatron-LM(Shoeybietal.,2019),MetaSeq(Zhangetal.,2022),andColossal-AI(Lietal.,2023a),train LLMswitheitherFP32full-precisionorFP16/BF16mixed-precisionbydefault. Thisisnotessential,however, Contributions for all the authors can be found in Section 5. * equal work†contact: {hanhu \| pengc}@microsoft.comarXiv:2310.18313v1 [cs.LG] 27 Oct 2023
2307.10169.pdf	Challenges and Applications of Large Language Models Jean Kaddourα,†,∗, Joshua Harrisβ,∗, Maximilian Mozesα, Herbie Bradleyγ,δ,ϵ, Roberta Raileanuζ, and Robert McHardyη,∗ αUniversity College LondonβUK Health Security AgencyγEleutherAI δUniversity of CambridgeϵStability AIζMeta AI ResearchηInstaDeep Abstract Large Language Models (LLMs) went from non-existent to ubiquitous in the machine learn- ing discourse within a few years. Due to the fast pace of the field, it is difficult to identify the remaining challenges and already fruitful application areas. In this paper, we aim to es- tablish a systematic set of open problems and application successes so that ML researchers can comprehend the field’s current state more quickly and become productive. Contents 1 Introduction 1 2 Challenges 2 2.1 Unfathomable Datasets . . . . . . 2 2.2 Tokenizer-Reliance . . . . . . . . 4 2.3 High Pre-Training Costs . . . . . 6 2.4 Fine-Tuning Overhead . . . . . . 10 2.5 High Inference Latency . . . . . . 11 2.6 Limited Context Length . . . . . . 14 2.7 Prompt Brittleness . . . . . . . . 17 2.8 Hallucinations . . . . . . . . . . . 19 2.9 Misaligned Behavior . . . . . . . 22 2.10 Outdated Knowledge . . . . . . . 27 2.11 Brittle Evaluations . . . . . . . . 27 2.12 Evaluations Based on Static, Human-Written Ground Truth . . 28 2.13 Indistinguishability between Gen- erated and Human-Written Text . 29 2.14 Tasks Not Solvable By Scale . . . 30 2.15 Lacking Experimental Designs . . 31 2.16 Lack of Reproducibility . . . . . . 33 3 Applications 34 3.1 Chatbots . . . . . . . . . . . . . . 34 3.2 Computational Biology . . . . . . 36 3.3 Computer Programming . . . . . 37 *Equal contribution. †{jean.kaddour,robert.mchardy}[email protected], [email protected] Design Unfathomable Datasets, Tokenizer-Reliance, Fine-Tuning Overhead Science Evaluations Based on Static Human-Written Ground Truth, Lacking Experimental Designs, Lack of Reproducibility Behavior Prompt Brittleness, Misaligned Behavior, Outdated Knowledge Detecting Generated Texts, Brittle Evaluations High Pre-Training Costs High Inference Latency, Limited Context Length, Hallucinations Tasks Not Solvable By ScaleFigure 1: Overview of LLM Challenges. Design ing LLMs relates to decisions taken before deployment. Be- havior ial challenges occur during deployment. Science challenges hinder academic progress. 3.4 Creative Work . . . . . . . . . . . 39 3.5 Knowledge Work . . . . . . . . . 40 3.6 Law . . . . . . . . . . . . . . . . 42 3.7 Medicine . . . . . . . . . . . . . 43 3.8 Reasoning . . . . . . . . . . . . . 44 3.9 Robotics and Embodied Agents . . 45 3.10 Social Sciences & Psychology . . 46 3.11 Synthetic Data Generation . . . . 48 4 Related Work 49 5 Conclusion 49 1 Introduction Given the quickly growing plethora of LLM re- search papers, we aim to address two questions: (1) Challenges : What problems remain unresolved? and (2) Applications : Where are LLMs currently being applied, and how are the challenges con- straining them? For (1), we group the challenges 1arXiv:2307.10169v1 [cs.CL] 19 Jul 2023
2401.01325.pdf	LLM Maybe LongLM: Self-Extend LLM Context Window Without Tuning Hongye Jin1 Xiaotian Han1 Jingfeng Yang2Zhimeng Jiang1Zirui Liu3Chia-Yuan Chang1 Huiyuan Chen4Xia Hu3 Abstract This work elicits LLMs’ inherent ability to handle long contexts without fine-tuning. The limited length of the training sequence during training may limit the application of Large Language Mod- els (LLMs) on long input sequences for inference. In this work, we argue that existing LLMs them- selves have inherent capabilities for handling long contexts. Based on this argument, we suggest ex- tending LLMs’ context window by themselves to fully utilize the inherent ability.We propose Self- Extend to stimulate LLMs’ long context handling potential. The basic idea is to construct bi-level attention information: the group level and the neighbor level. The two levels are computed by the original model’s self-attention, which means the proposed does not require any training. With only four lines of code modification, the proposed method can effortlessly extend existing LLMs’ context window without any fine-tuning. We con- duct comprehensive experiments and the results show that the proposed method can effectively extend existing LLMs’ context window’s length. 1. Introduction The context window length of most existing LLMs is lim- ited since they are trained with a fixed length of training sequences. It’s determined by the context window length during the pretraining stage. Once the length of the input texts exceeds the pretraining context window during the inference, the behavior of LLMs will be unpredictable and suffer from severe performance degradation, which is shown on the perplexity (PPL) metric, the PPL of the model will explode with the long input sequence (Xiao et al., 2023; Peng et al., 2023b; Han et al., 2023; Chen et al., 2023b). *Equal contribution1Texas A&M University2Amazon, the views expressed or the conclusions reached are his own and do not represent the view of Amazon3Rice University4Case Western Reserve University. Correspondence to: Hongye Jin <[email protected] >. Working in Progress.Recently, a variety of content window extension methods have been developed to tackle the challenge of extending the context window size of pretrained LLMs. A common and straightforward approach is to fine-tune these models on enough extensive texts (Xiong et al., 2023). Besides this, some methods seek to achieve extension either without the need for fine-tuning or with only minimal fine-tuning, offer- ing a more efficient alternative. Among these contemporary methods, some notable techniques include ’PI’ (Chen et al., 2023b), ’CLEX’ (Chen et al., 2023a) and ’Yarn’ (Peng et al., 2023b). However, they typically necessitate a few fine- tuning steps to achieve extension, which could be resource- intensive and time-consuming. And these methods aim to extend the content window based on the assumption that pretrained LLMs lack the ability to handle long content . Thus, limited finetuning could make LLMs overfit to specific long sequences, which lacks generalizability over out-of- distribution long sequences and loses performance on short sequences obtained during pretraining. On the other hand, some methods (Xiao et al., 2023; Han et al., 2023; Ding et al., 2023) aim to avoid fine-tuning. These fine-tuning- free approaches rely predominantly on local information in the sequence. However, these methods may fall short in effectively expanding the context window, as it is limited to using only local tokens rather than expanding the overall context-handling capacity of the LLMs. Consequently, they may not fully realize the potential of extending the context window in LLMs and have inferior performance. Instead of extending the content window, in this paper, we believe LLMs should have inherent capabilities to handle long contexts . Our belief stems from the fact that when we, as human beings, are children, we are taught how to read and write using relatively short texts, such as articles spanning several pages. We rarely use extremely long texts like entire books or complete documents as learning materials. Yet, we are still able to understand long texts effectively. With this strong motivation, the poor performance of LLMs while facing long text out of the pretraining context window size is not due to the lack of long context understanding capabilities. We suppose that there should be a way to elicit LLMs’ inherent long context capability. In our analysis, we observe that the key challenge prevent- ing LLMs from effectively managing extensive contexts is 1arXiv:2401.01325v1 [cs.CL] 2 Jan 2024
2304.11082.pdf	Preprint. Under review. FUNDAMENTAL LIMITATIONS OF ALIGNMENT INLARGE LANGUAGE MODELS Yotam Wolf∗ The Hebrew University [email protected] Wies∗ The Hebrew University [email protected] Yoav Levine AI21 Labs [email protected] Shashua The Hebrew University [email protected] ABSTRACT An important aspect in developing language models that interact with humans is aligning their behavior to be useful and unharmful for their human users. This is usually achieved by tuning the model in a way that enhances desired behaviors and inhibits undesired ones, a process referred to as alignment . In this paper, we propose a theoretical approach called Behavior Expectation Bounds (BEB) which allows us to formally investigate several inherent characteristics and limitations of alignment in large language models. Importantly, we prove that for any behavior that has a ﬁnite probability of being exhibited by the model, there exist prompts that can trigger the model into outputting this behavior, with probability that increases with the length of the prompt. This implies that any alignment process that attenuates undesired behavior but does not remove it altogether, is not safe against adversarial prompting attacks. Furthermore, our framework hints at the mechanism by which leading alignment approaches such as reinforcement learning from human feedback increase the LLM’s proneness to being prompted into the undesired behaviors. Moreover, we include the notion of personas in our BEB framework, and ﬁnd that behaviors which are generally very unlikely to be exhibited by the model can be brought to the front by prompting the model to behave as speciﬁc persona. This theoretical result is being experimentally demonstrated in large scale by the so called contemporary “chatGPT jailbreaks", where adversarial users trick the LLM into breaking its alignment guardrails by triggering it into acting as a malicious persona. Our results expose fundamental limitations in alignment of LLMs and bring to the forefront the need to devise reliable mechanisms for ensuring AI safety. 1 I NTRODUCTION Training large language models (LLMs) over vast corpora has revolutionized natural language processing, giving LLMs the ability to mimic human-like interactions and serve as general purpose assistants in a wide variety of tasks, such as wide-scoped question answering, writing assistance, teaching, and more (Radford et al., 2019; Devlin et al., 2019; Brown et al., 2020; Schulman et al., 2023; OpenAI, 2023; Bubeck et al., 2023; Nori et al., 2023; West, 2023; Park et al., 2023). A growing concern due to the increasing reliance on LLMs for such purposes is the harm they can cause their users, such as feeding fake information (Lin et al., 2022; Weidinger et al., 2022), behaving offensively and feeding social biases (Hutchinson et al., 2020; Venkit et al., 2022; Weidinger et al., 2022), or encouraging problematic behaviors by users (even by psychologically manipulating them Roose 2023; Atillah 2023). Indeed, evidently, the unsupervised textual data used for pretraining modern LLMs includes enough demonstrations of the above undesired behaviors for them to be present in the resulting models (Bender et al., 2021). The act of removing these undesired behaviors is often called alignment (Yudkowsky, 2001; Taylor et al., 2016; Amodei et al., 2016; Shalev-Shwartz et al., 2020; Hendrycks et al., 2021; Pan et al., 2022; Ngo, 2022). ∗Equal contribution 1arXiv:2304.11082v1 [cs.CL] 19 Apr 2023
2302.04065.pdf	Monge, Bregman and Occam: Interpretable Optimal Transport in High-Dimensions with Feature-Sparse Maps Marco Cuturi1Michal Klein1Pierre Ablin1 Abstract Optimal transport (OT) theory focuses, among all mapsT:Rd→Rdthat can morph a prob- ability measure onto another, on those that are the “thriftiest”, i.e. such that the averaged cost c(x,T(x))between xand its image T(x)be as small as possible. Many computational ap- proaches have been proposed to estimate such Monge maps when cis theℓ2 2distance, e.g., us- ing entropic maps (Pooladian and Niles-Weed, 2021), or neural networks (Makkuva et al., 2020; Korotin et al., 2020). We propose a new model for transport maps, built on a family of transla- tion invariant costs c(x,y) :=h(x−y), where h:=1 2∥·∥2 2+τandτis a regularizer. We pro- pose a generalization of the entropic map suitable forh, and highlight a surprising link tying it with theBregman centroids of the divergence Dhgen- erated byh, and the proximal operator of τ. We show that choosing a sparsity-inducing norm for τ results in maps that apply Occam ’s razor to trans- port, in the sense that the displacement vectors ∆(x) :=T(x)−xthey induce are sparse, with a sparsity pattern that varies depending on x. We showcase the ability of our method to estimate meaningful OT maps for high-dimensional single- cell transcription data, in the 34000 -dspace of gene counts for cells, without using dimensional- ity reduction, thus retaining the ability to interpret all displacements at the gene level. 1. Introduction A fundamental task in machine learning is learning how to transfer observations from a source to a target probability measure. For such problems, optimal transport (OT) (San- tambrogio, 2015) has emerged as a powerful toolbox that can improve performance and guide theory in various set- tings. For instance, the computational approaches advo- cated in OT have been used to transfer knowledge across 1Apple.{cuturi,michalk,p ablin}@apple.comdatasets in domain adaptation tasks (Courty et al., 2016; 2017), train generative models (Montavon et al., 2016; Ar- jovsky et al., 2017; Genevay et al., 2018; Salimans et al., 2018), and realign datasets in natural sciences (Janati et al., 2019; Schiebinger et al., 2019). High-dimensional Transport. OT ﬁnds its most straight- forward and intuitive use-cases in low-dimensional geomet- ric domains (grids and meshes, graphs, etc...). This work focuses on the more challenging problem of using it on distributions in Rd, withd≫1. InRd, the ground cost c(x,y)between observations x,yis often theℓ2metric or its squareℓ2 2. However, when used on large- ddata samples, that choice is rarely meaningful. This is due to the curse-of- dimensionality associated with OT estimation (Dudley et al., 1966; Weed and Bach, 2019) and the fact that the Euclidean distance loses its discriminative power as dimension grows. To mitigate this, practitioners rely on dimensionality reduc- tion, either in two steps , before running OT solvers, using, e.g., PCA, a V AE, or a sliced-Wasserstein approach (Rabin et al., 2012; Bonneel et al., 2015); or jointly , by estimating both a projection and transport, e.g., on hyperplanes (Niles- Weed and Rigollet, 2022; Paty and Cuturi, 2019; Lin et al., 2020; Huang et al., 2021; Lin et al., 2021), lines (Deshpande et al., 2019; Kolouri et al., 2019), trees (Le et al., 2019) or more advanced featurizers (Salimans et al., 2018). However, an obvious drawback of these approaches is that transport maps estimated in reduced dimensions are hard to interpret in the original space (Muzellec and Cuturi, 2019). Contributions. To target high dregimes, we introduce a radically different approach. We use the sparsity tool- box (Hastie et al., 2015; Bach et al., 2012) to build OT maps that are, adaptively to input x, drastically simpler: •We introduce a generalized entropic map (Pooladian and Niles-Weed, 2021) for translation invariant costs c(x,y) :=h(x−y), wherehis strongly convex. That entropic map Th,εis deﬁned almost everywhere (a.e.), and we show that it induces displacements ∆(x) :=T(x)−x that can be cast as Bregman centroids, relative to the Bregman divergence generated by h. •Whenhis an elastic-type regularizer, the sum of a strongly-convex term ℓ2 2and a sparsifying norm τ, we show that such centroids are obtained using the proximalarXiv:2302.04065v1 [stat.ML] 8 Feb 2023
2106.09685.pdf	LORA: L OW-RANK ADAPTATION OF LARGE LAN- GUAGE MODELS Edward Hu∗Yelong Shen∗Phillip Wallis Zeyuan Allen-Zhu Yuanzhi Li Shean Wang Lu Wang Weizhu Chen Microsoft Corporation {edwardhu, yeshe, phwallis, zeyuana, yuanzhil, swang, luw, wzchen }@microsoft.com [email protected] (Version 2) ABSTRACT An important paradigm of natural language processing consists of large-scale pre- training on general domain data and adaptation to particular tasks or domains. As we pre-train larger models, full ﬁne-tuning, which retrains all model parameters, becomes less feasible. Using GPT-3 175B as an example – deploying indepen- dent instances of ﬁne-tuned models, each with 175B parameters, is prohibitively expensive. We propose Low-RankAdaptation, or LoRA, which freezes the pre- trained model weights and injects trainable rank decomposition matrices into each layer of the Transformer architecture, greatly reducing the number of trainable pa- rameters for downstream tasks. Compared to GPT-3 175B ﬁne-tuned with Adam, LoRA can reduce the number of trainable parameters by 10,000 times and the GPU memory requirement by 3 times. LoRA performs on-par or better than ﬁne- tuning in model quality on RoBERTa, DeBERTa, GPT-2, and GPT-3, despite hav- ing fewer trainable parameters, a higher training throughput, and, unlike adapters, no additional inference latency . We also provide an empirical investigation into rank-deﬁciency in language model adaptation, which sheds light on the efﬁcacy of LoRA. We release a package that facilitates the integration of LoRA with PyTorch models and provide our implementations and model checkpoints for RoBERTa, DeBERTa, and GPT-2 at https://github.com/microsoft/LoRA . 1 I NTRODUCTION Pretrained Weights 𝑊∈ℝ𝑑×𝑑 xh 𝐵=0 𝐴=𝒩(0,𝜎2) 𝑑𝑟Pretrained Weights 𝑊∈ℝ𝑑×𝑑 xf(x) 𝑑 Figure 1: Our reparametriza- tion. We only train AandB.Many applications in natural language processing rely on adapt- ingonelarge-scale, pre-trained language model to multiple down- stream applications. Such adaptation is usually done via ﬁne-tuning , which updates all the parameters of the pre-trained model. The ma- jor downside of ﬁne-tuning is that the new model contains as many parameters as in the original model. As larger models are trained every few months, this changes from a mere “inconvenience” for GPT-2 (Radford et al., b) or RoBERTa large (Liu et al., 2019) to a critical deployment challenge for GPT-3 (Brown et al., 2020) with 175 billion trainable parameters.1 Many sought to mitigate this by adapting only some parameters or learning external modules for new tasks. This way, we only need to store and load a small number of task-speciﬁc parameters in ad- dition to the pre-trained model for each task, greatly boosting the operational efﬁciency when deployed. However, existing techniques ∗Equal contribution. 0Compared to V1, this draft includes better baselines, experiments on GLUE, and more on adapter latency. 1While GPT-3 175B achieves non-trivial performance with few-shot learning, ﬁne-tuning boosts its perfor- mance signiﬁcantly as shown in Appendix A. 1arXiv:2106.09685v2 [cs.CL] 16 Oct 2021
10.2307.2334029.pdf	A note on DPO with noisy preferences & relationship to IPO Eric Mitchell November 25, 2023 (v1.1) ‘OG’ RLHF aims for reward maximization with a KL constraint to reference model πref(inputs xomitted): π∗= argmax πEy∼π r(y)−βlogπ(y) πref(y) (1) DPO [3] derives a loss on the current policy πθ(where our dataset says ywis preferred to yl, oryw≻yl): LDPO(θ, yw, yl) =−logσ βlogπθ(yw) πref(yw)−βlogπθ(yl) πref(yl) , (2) i.e., the binary cross entropy with ˆ pθ(yw≻yl) =σ βlogπθ(yw) πref(yw)−βlogπθ(yl) πref(yl) and target p(yw≻yl) = 1. What if preference labels are noisy? Say the labels have been flipped with some small probability ϵ∈(0,0.5). We can use a conservative target distribution instead, p(yw≻yl) = 1−ϵ, giving BCE loss: Lϵ DPO(θ, yw, yl) =−(1−ϵ) log ˆpθ(yw≻yl)−ϵlog(1−ˆpθ(yw≻yl)) (3) = (1−ϵ)LDPO(θ, yw, yl) +ϵLDPO(θ, yl, yw) (4) The gradient of Lϵ DPO(θ, yw, yl) is simply the weighted sum of gradients (1 −ϵ)∇θL(θ, yw, yl)+ϵ∇θL(θ, yl, yw), which reduces to the simplified form (ignoring constants; see [3] for the gradient of the original DPO loss): ∇θLϵ DPO(θ, yw, yl) =− (1−ϵ)(1−ˆpθ)−ϵˆpθh ∇θlogπθ(yw)\|{z} upweight yw−∇ θlogπθ(yl)\|{z} downweight yli (5) = ˆpθ−(1−ϵ)h ∇θlogπθ(yw)− ∇ θlogπθ(yl)i . (6) The gradient is zero when ˆ pθ(yw≻yl) = (1 −ϵ), i.e., our (implicit) reward assigns the desired confidence level in this training example under the Bradley-Terry model [2]. For normal DPO, the gradient is never zero! Using the shorthand hyw,ylπθ= logπθ(yw) πref(yw)−logπθ(yl) πref(yl), let’s compare the conservative DPO (cDPO?) and IPO [1] loss gradient, where the IPO loss is given in Eq. 17 of [1] as LIPO(θ, yw, yl) = hyw,ylπ−1 2β2 : ∇θLIPO(θ, yw, yl) = hyw,ylπθ−1 2βh ∇θlogπθ(yw)− ∇ θlogπθ(yl)i (7) ∇θLϵ DPO(θ, yw, yl) = σ(βhyw,ylπθ)−(1−ϵ)h ∇θlogπθ(yw)− ∇ θlogπθ(yl)i (8) TL;DR: conservative DPO trains the model until a desired improvement in the implicit probability assigned by the model to the observed preferences1is met; IPO trains the model until a desired improvement in implicit reward is met. The ability for cDPO and IPO to optimize only to a fixed delta from the reference model and then stop (or even reverse!) likely makes these more stable than the original DPO loss after lots of training. [1] Mohammad Gheshlaghi Azar, Mark Rowland, Bilal Piot, Daniel Guo, Daniele Calandriello, Michal Valko, and R´ emi Munos. A General Theoretical Paradigm to Understand Learning from Human Preferences . 2023. arXiv: 2310.12036 [cs.AI] . [2] Ralph Allan Bradley and Milton E. Terry. “Rank Analysis of Incomplete Block Designs: I. The Method of Paired Comparisons”. In:Biometrika 39.3/4 (1952), pp. 324–345. doi:https://doi.org/10.2307/2334029 . [3] Rafael Rafailov, Archit Sharma, Eric Mitchell, Christopher D Manning, Stefano Ermon, and Chelsea Finn. “Direct Preference Optimization: Your Language Model is Secretly a Reward Model”. In: Neural Information Processing Systems . 2023. 1The Bradley-Terry model of human preferences [2] converts the β-scaled reward gap hyw,ylπθto a probability assigned by the model to the observed preference bit using the sigmoid of the scaled reward gap.
10.1016.j.cell.2024.01.003.pdf	Leading Edge Commentary Structure is beauty, but not always truth James S. Fraser1,and Mark A. Murcko2, 1Department of Bioengineering and Therapeutic Sciences, University of California San Francisco, San Francisco, CA, USA 2Disruptive Biomedical LLC, Holliston, MA, USA *Correspondence: [email protected] (J.S.F.), [email protected] (M.A.M.) https://doi.org/10.1016/j.cell.2024.01.003 Structural biology, as powerful as it is, can be misleading. We highlight four fundamental challenges: inter- preting raw experimental data; accounting for motion; addressing the misleading nature of in vitro structures; and unraveling interactions between drugs and ‘‘anti-targets.’’ Overcoming these challenges will amplify theimpact of structural biology on drug discovery. Introduction Amidst the many uncertainties that compli- cate drug discovery, structural biology an-chors the process in beautiful and con- crete images of drugs interacting with receptors. Structure can enable the tack-ling of many of the key challenges of drug design. Atomistic models emerging from cryo-electron microscopy (cryo-EM),X-ray crystallography, and nuclear mag- netic resonance (NMR) provide strong starting points for thinking broadly andcreatively about how to modulate proteinfunction by identifying binding pockets and potential allosteric sites. Ligand- bound structures greatly focus the searchof chemical space to molecules that main- tain key interactions with the receptor. By offering a ‘‘ground truth,’’ structural biology is clear, quantiﬁable, and inter- pretable. In the best cases, the precise location of every atom is clearly deﬁned.For example, measuring the distances between atoms on the ligand and recep- tor allow us to infer that ‘‘this hydrogenbond is better than that one.’’ By compar-ison, the biology of the target is often quite complex and difﬁcult to model quantita- tively. Cellular assays and animal modelsare approximations that do not fully reca- pitulate the human disease process or the potential of small molecules to inducetoxicities. In addition, medicinal chemistry is full of uncertainty. Even with a structure, in the course of inhibitor optimization, it isdifﬁcult to know what molecules to makenext or how to synthesize them. Finally, we struggle to understand why some mol- ecules are more potent or more bioavail-able than others, making it next to impos- sible to optimize the pharmacokinetic and safety proﬁle of a drug candidate.However, the ‘‘truth’’ of structural biology raises a legitimate concern: does the availability of structural information,such as a protein crystal structure, irrepa- rably constrain the creative process? This is a potential risk for those who fail torecognize the inherent limitations in the structures and the new predictions (e.g., AlphaFold2) 1trained on the corpus of the Protein DataBank (PDB). We suggest it is useful to consider four kinds of limita- tions and offer ways that the ﬁeld canaddress each of them to optimize thevalue we derive from structural biology and further improve the quality of predic- tive modeling. Four harsh truths about structural biology and drug discovery (1) A structure is a model, not experi- mental reality. (2) Representing wiggling and jiggling is hard. (3)In vitro can be deceiving. (4) Drugs mingle with many different receptors. A structure is a model, not experimental realityUndoubtedly, AlphaFold2 1shook up the ﬁeld of structural biology by ‘‘solving’’ the protein structure prediction problem.‘‘Solving’’ means that the predictedmodels are highly similar to ‘‘ground truth’’ experimentally determined struc- tures by the metrics used by the CASP(Critical Assessment of Structure Pre- diction) competition, a community-wide experiment to determine and advancethe state of the art in modeling protein structure from amino acid sequence. It isimportant to note that ground truth struc- tures contain inaccuracies beyond the signal-to-noise of the experiments thatgenerate them. For example, in X-ray crystallography, the experimental data is measured very precisely (usually to lessthan 5% error), but the structures reﬁned against that data have large residual errors compared to the experimentaldata (generally > /C2420%). 2Adding in prior knowledge, such as geometry restraints, is especially important as the resolutionof the experimental data gets worse. 2 ‘‘Truth’’ therefore may not lie only in com- parison to the atomic coordinates, espe- cially when the structure is based onlow-resolution data. Rather, comparisons to density maps (or even raw diffraction images or micrographs) may reveal adeeper form of truth ( Figure 1 ). Recent work has started the important task of comparing AlphaFold2 modelsdirectly to experimental crystallographic density maps. 3In many cases, predic- tions closely matched experimentalmaps. Reﬁnement of the AlphaFold2models against experimental data can resolve some global scale distortion and issues of domain orientation. Reﬁnementalso improves local backbone and side- chain conformations. However, most very high-conﬁdence predictions differedfrom experimental maps to a greater extent than independently determined experimental structures. Beyond suggesting that direct agree- ment with experimental data, not ‘‘struc- tures,’’ could be a new benchmark of ‘‘ground truth,’’ these ﬁndings promptus to ask how we can maximize the utility of computationally predicted models in drug discovery. Some differences ll Cell 187, February 1, 2024 ª2024 Elsevier Inc. 517
2307.13304.pdf	QuIP: 2-Bit Quantization of Large Language Models With Guarantees Jerry Chee Department of Computer Science Cornell University [email protected] Cai Department of Electrical and Computer Engineering Cornell University [email protected] Volodymyr Kuleshov Department of Computer Science Cornell University [email protected] De Sa Department of Computer Science Cornell University [email protected] Abstract This work studies post-training parameter quantization in large language models (LLMs). We introduce quantization with incoherence processing (QuIP), a new method based on the insight that quantization benefits from incoherent weight and Hessian matrices, i.e., from the weights and the directions in which it is important to round them accurately being unaligned with the coordinate axes. QuIP consists of two steps: (1) an adaptive rounding procedure minimizing a quadratic proxy objective; (2) efficient pre- and post-processing that ensures weight and Hessian incoherence via multiplication by random orthogonal matrices. We complement QuIP with the first theoretical analysis for an LLM-scale quantization algorithm, and show that our theory also applies to an existing method, OPTQ. Empirically, we find that our incoherence preprocessing improves several existing quantization algorithms and yields the first LLM quantization methods that produce viable results using only two bits per weight. Our code can be found on GitHub. 1 Introduction Large language models (LLMs) have enabled advances in text generation, few-shot learning, reason- ing, protein sequence modeling, and other tasks [ 2,29,34]. The massive size of these models—often reaching into hundreds of billions of parameters—requires sophisticated deployment methods and motivates research into efficient inference algorithms. This work studies the post-training quantization of LLM parameters as a way to improve their runtime efficiency [ 4,8,22,30,32,33]. Our key insight is that quantization can be most effective when weight and proxy Hessian matrices are incoherent —intuitively, both the weights themselves and the directions in which it is important to have good rounding accuracy are not too large in any one coordinate, which makes it easier to adaptively round the weights to a finite set of compressed values. We use this intuition to develop theoretically sound two-bit quantization algorithms that scale to LLM-sized models. Specifically, we introduce quantization with incoherence processing (QuIP), a new method motivated by the above insight. QuIP consists of two steps: (1) an adaptive rounding [ 20] procedure, which minimizes a quadratic proxy objective ℓ(ˆW) = tr(( ˆW−W)H(ˆW−W)T)of the error between the original weights Wand the quantized weights ˆWusing an estimate of the Hessian H; (2) efficient Preprint. Under review.arXiv:2307.13304v1 [cs.LG] 25 Jul 2023
2203.02155.pdf	Training language models to follow instructions with human feedback Long Ouyang∗Jeff Wu∗Xu Jiang∗Diogo Almeida∗Carroll L. Wainwright∗ Pamela Mishkin∗Chong Zhang Sandhini Agarwal Katarina Slama Alex Ray John Schulman Jacob Hilton Fraser Kelton Luke Miller Maddie Simens Amanda Askell†Peter Welinder Paul Christiano∗† Jan Leike∗Ryan Lowe∗ OpenAI Abstract Making language models bigger does not inherently make them better at following a user’s intent. For example, large language models can generate outputs that are untruthful, toxic, or simply not helpful to the user. In other words, these models are not aligned with their users. In this paper, we show an avenue for aligning language models with user intent on a wide range of tasks by ﬁne-tuning with human feedback. Starting with a set of labeler-written prompts and prompts submitted through the OpenAI API, we collect a dataset of labeler demonstrations of the desired model behavior, which we use to ﬁne-tune GPT-3 using supervised learning. We then collect a dataset of rankings of model outputs, which we use to further ﬁne-tune this supervised model using reinforcement learning from human feedback. We call the resulting models InstructGPT . In human evaluations on our prompt distribution, outputs from the 1.3B parameter InstructGPT model are preferred to outputs from the 175B GPT-3, despite having 100x fewer parameters. Moreover, InstructGPT models show improvements in truthfulness and reductions in toxic output generation while having minimal performance regressions on public NLP datasets. Even though InstructGPT still makes simple mistakes, our results show that ﬁne-tuning with human feedback is a promising direction for aligning language models with human intent. 1 Introduction Large language models (LMs) can be “prompted” to perform a range of natural language process- ing (NLP) tasks, given some examples of the task as input. However, these models often express unintended behaviors such as making up facts, generating biased or toxic text, or simply not following user instructions (Bender et al., 2021; Bommasani et al., 2021; Kenton et al., 2021; Weidinger et al., 2021; Tamkin et al., 2021; Gehman et al., 2020). This is because the language modeling objective ∗Primary authors. This was a joint project of the OpenAI Alignment team. RL and JL are the team leads. Corresponding author: [email protected] . †Work done while at OpenAI. Current afﬁliations: AA: Anthropic; PC: Alignment Research Center.arXiv:2203.02155v1 [cs.CL] 4 Mar 2022
2305.14992.pdf	Reasoning with Language Model is Planning with World Model Shibo Hao∗♣Yi Gu∗ ∗♣Haodi Ma♢Joshua Jiahua Hong♣Zhen Wang♣ ♠ Daisy Zhe Wang♢Zhiting Hu♣ ♣UC San Diego,♢University of Florida ♠Mohamed bin Zayed University of Artificial Intelligence {s5hao, yig025, jjhong, zhw085, zhh019}@ucsd.edu {ma.haodi, daisyw}@ufl.edu Abstract Large language models (LLMs) have shown remarkable reasoning capabilities, especially when prompted to generate intermediate reasoning steps (e.g., Chain-of- Thought, CoT). However, LLMs can still struggle with problems that are easy for humans, such as generating action plans for executing tasks in a given environment, or performing complex math, logical, and commonsense reasoning. The deficiency stems from the key fact that LLMs lack an internal world model to predict the world state (e.g., environment status, intermediate variable values) and simulate long-term outcomes of actions. This prevents LLMs from performing deliberate planning akin to human brains, which involves exploring alternative reasoning paths, anticipating future states and rewards, and iteratively refining existing reasoning steps. To overcome the limitations, we propose a new LLM reasoning framework, Reasoning viaPlanning (RAP) . RAP repurposes the LLM as both a world model and a reasoning agent, and incorporates a principled planning algorithm (based on Monto Carlo Tree Search) for strategic exploration in the vast reasoning space. During reasoning, the LLM (as agent) incrementally builds a reasoning tree under the guidance of the LLM (as world model) and task-specific rewards, and obtains a high-reward reasoning path efficiently with a proper balance between exploration vs.exploitation. We apply RAP to a variety of challenging reasoning problems including plan generation, math reasoning, and logical inference. Empirical results on these tasks demonstrate the superiority of RAP over various strong baselines, including CoT and least-to-most prompting with self-consistency. RAP on LLaMA- 33B surpasses CoT on GPT-4 with 33% relative improvement in a plan generation setting. 1 Introduction Large language models (LLMs) have exhibited emergent reasoning abilities in a wide range of tasks [ 5,10,44,2]. Recent approaches further boost their ability by prompting LLMs to generate intermediate reasoning steps (e.g., Chain-of-Thought, CoT [ 59]) or answer a series of subquestions (e.g., least-to-most prompting [ 66]). However, LLMs still face difficulties with tasks that humans find easy. For example, in creating action plans to move blocks to a target state, GPT-3 [ 5] achieves a success rate of only 1%, compared to 78% for humans [ 57]; these models also struggle when solving complex tasks that require multiple steps of math, logical, or commonsense reasoning [ 65,22,41,6]. Humans possess an internal world model , a mental representation of the environment [ 28,27,15], which enables humans to simulate actions and their effects on the world’s state for deliberate planning ∗equal contribution Preprint. Under review.arXiv:2305.14992v1 [cs.CL] 24 May 2023
1606.06565.pdf	Concrete Problems in AI Safety Dario Amodei∗ Google BrainChris Olah∗ Google BrainJacob Steinhardt Stanford UniversityPaul Christiano UC Berkeley John Schulman OpenAIDan Man´ e Google Brain Abstract Rapid progress in machine learning and artiﬁcial intelligence (AI) has brought increasing atten- tion to the potential impacts of AI technologies on society. In this paper we discuss one such potential impact: the problem of accidents in machine learning systems, deﬁned as unintended and harmful behavior that may emerge from poor design of real-world AI systems. We present a list of ﬁve practical research problems related to accident risk, categorized according to whether the problem originates from having the wrong objective function (“avoiding side eﬀects” and “avoiding reward hacking”), an objective function that is too expensive to evaluate frequently (“scalable supervision”), or undesirable behavior during the learning process (“safe exploration” and “distributional shift”). We review previous work in these areas as well as suggesting re- search directions with a focus on relevance to cutting-edge AI systems. Finally, we consider the high-level question of how to think most productively about the safety of forward-looking applications of AI. 1 Introduction The last few years have seen rapid progress on long-standing, diﬃcult problems in machine learning and artiﬁcial intelligence (AI), in areas as diverse as computer vision [82], video game playing [102], autonomous vehicles [86], and Go [140]. These advances have brought excitement about the positive potential for AI to transform medicine [126], science [59], and transportation [86], along with concerns about the privacy [76], security [115], fairness [3], economic [32], and military [16] implications of autonomous systems, as well as concerns about the longer-term implications of powerful AI [27, 167]. The authors believe that AI technologies are likely to be overwhelmingly beneﬁcial for humanity, but we also believe that it is worth giving serious thought to potential challenges and risks. We strongly support work on privacy, security, fairness, economics, and policy, but in this document we discuss another class of problem which we believe is also relevant to the societal impacts of AI: the problem of accidents in machine learning systems. We deﬁne accidents as unintended and harmful behavior that may emerge from machine learning systems when we specify the wrong objective function, are ∗These authors contributed equally. 1arXiv:1606.06565v2 [cs.AI] 25 Jul 2016
10.1016.j.cell.2023.12.010.pdf	Article Hypoxia and intra-complex genetic suppressors rescue complex I mutants by a shared mechanism Graphical abstract Highlights dHypoxia rescue and hyperoxia sensitivity of complex I mutants are conserved in C. elegans dHypoxia rescue is independent of HIF activation or attenuation of ROS toxicity dNDUFA6/nuo-3(G60D) mimics acute hypoxia in restoring complex I forward activity dResidues in the CoQ binding pocket are required for rescuebynuo-3(G60D) or hypoxiaAuthors Joshua D. Meisel, Maria Miranda, Owen S. Skinner, ..., Alexis A. Jourdain,Gary Ruvkun, Vamsi K. Mootha Correspondence [email protected] (G.R.), [email protected] (V.K.M.) In brief C. elegans mutants harboring a defective mitochondrial complex I are rescued byhypoxia or intra-complex geneticsuppressor mutation, achieved byincreasing forward ﬂow of electronsthrough complex I and dependent onresidues surrounding the CoQ bindingpocket. NADH NAD+ CoQ CoQH2 H+e- NADH NAD+ CoQ CoQH2H+e- O2 1% 50% O2 rescue & O2 sensitivity of complex I mutants is conserved in C. elegans and independent of HIF, ROSmutagenize 50% O2 sequence “hypoxia-mimetic” NDUFA6/nuo-3(G60D)complex I mutantsrescue of complex I activity timeG60DtimelowO 2 mutagenize rotenone sequenceNDUFA6/nuo-3(G60D) “hypoxia-resistant” complex I mutationsCoQCI mutant O2 CI mutant; G60D or hypoxia O2wt CI mutant ETC flux ETC flux Meisel et al., 2024, Cell 187, 659–675 February 1, 2024 ª2023 The Author(s). Published by Elsevier Inc. https://doi.org/10.1016/j.cell.2023.12.010 ll
2402.04362v2.pdf	Neural Networks Learn Statistics of Increasing Complexity Nora Belrose1Quintin Pope2Lucia Quirke1Alex Mallen1Xiaoli Fern2 Abstract The distributional simplicity bias (DSB) posits that neural networks learn low-order moments of the data distribution first, before moving on to higher-order correlations. In this work, we present compelling new evidence for the DSB by show- ing that networks automatically learn to perform well on maximum-entropy distributions whose low-order statistics match those of the training set early in training, then lose this ability later. We also extend the DSB to discrete domains by proving an equivalence between token n-gram fre- quencies and the moments of embedding vectors, and by finding empirical evidence for the bias in LLMs. Finally we use optimal transport methods to surgically edit the low-order statistics of one class of images to match those of another, and show early-training networks treat the edited im- ages as if they were drawn from the target class. Code is available at https://github.com/ EleutherAI/features-across-time . 1. Introduction Neural networks exhibit a remarkable ability to fit com- plex datasets while generalizing to unseen data points and distributions. This is especially surprising given that deep networks can perfectly fit random labels (Zhang et al., 2021), and it is possible to intentionally “poison” networks so that they achieve zero training loss while behaving randomly on a held out test set (Huang et al., 2020). A recently proposed explanation for this phenomenon is thedistributional simplicity bias (DSB) : neural networks learn to exploit the lower-order statistics of the input data first– e.g. mean and (co)variance– before learning to use its higher-order statistics, such as (co)skewness or (co)kurtosis. Refinetti et al. (2023) provide evidence for the DSB by training networks on a sequence of synthetic datasets that act as increasingly precise approximations to the real data, showing that early checkpoints perform about as well on 1EleutherAI2Oregon State University. Correspondence to: Nora Belrose <[email protected] >.real data as checkpoints trained directly on the real data. We build on Refinetti et al. (2023) by inverting their exper- imental setup. We train our models on real datasets, then test them throughout training on synthetic data that probe the model’s reliance on statistics of different orders. We believe this experimental design provides more direct evi- dence about the generalization behavior of commonly used models and training practices. Our primary theoretical contributions are to (1)motivate the DSB through a Taylor expansion of the expected loss, (2) propose criteria quantifying whether a model “uses” statis- tics up to order kby checking that the model is sensitive to interventions on the first kstatistics, while being robust to interventions on higher-order statistics, (3)describe efficient methods of producing synthetic data that let us investigate whether models satisfy the above criteria, and (4)extend the DSB to discrete domains by proving an equivalence between token n-gram frequencies and the moments of sequences of embedding vectors. We use a Taylor series expansion to express a model’s ex- pected loss as a sum over the central moments of an evalua- tion dataset. This connection provides some motivation for the DSB. Specifically, if during training, a network’s loss is well approximated by the first kterms of its Taylor expan- sion, then the model should only be sensitive to statistics up to order k, and we argue that earlier terms of the expansion will generally become relevant before later terms. We describe two intuitive criteria that a model sensitive to statistics up to order kshould satisfy: (1)changing the first kstatistics of data from class A to match class B should cause the model to classify the modified data as class B, and (2)models should be unaffected by “deleting” higher-order data statistics. We evaluate whether image classification networks satisfy the above criteria during training through extensive empirical experiments across a variety of network architectures and image datasets. We evaluate whether the network satisfies criterion (1)by generating synthetic datasets where we “graft” the means and covariances of one class onto images of another class, and evaluating whether the network’s classifies the resulting data as belonging to the target class. We formalize this notion of “grafting” statistics with optimal transport theory, 1arXiv:2402.04362v2 [cs.LG] 13 Feb 2024
2210.05845.pdf	Contrastive Retrospection: honing in on critical steps for rapid learning and generalization in RL Chen Sun∗ Mila, Université de Montréal [email protected] Yang New York University [email protected] Jiralerspong Mila, Université de Montréal thomas.jiralerspong @mila.quebec Dane Malenfant McGill University [email protected] Alsbury-Nealy University of Toronto, SilicoLabs Incorporated [email protected] Yoshua Bengio Mila, Université de Montréal, CIFAR [email protected] Richards* Mila, McGill University Learning in Machines & Brains, CIFAR [email protected] Abstract In real life, success is often contingent upon multiple critical steps that are distant in time from each other and from the final reward. These critical steps are challenging to identify with traditional reinforcement learning (RL) methods that rely on the Bellman equation for credit assignment. Here, we present a new RL algorithm that uses offline contrastive learning to hone in on these critical steps. This algorithm, which we call Contrastive Retrospection (ConSpec), can be added to any existing RL algorithm. ConSpec learns a set of prototypes for the critical steps in a task by a novel contrastive loss and delivers an intrinsic reward when the current state matches one of the prototypes. The prototypes in ConSpec provide two key benefits for credit assignment: (i) They enable rapid identification of all the critical steps. (ii) They do so in a readily interpretable manner, enabling out-of-distribution generalization when sensory features are altered. Distinct from other contemporary RL approaches to credit assignment, ConSpec takes advantage of the fact that it is easier to retrospectively identify the small set of steps that success is contingent upon (and ignoring other states) than it is to prospectively predict reward at every taken step. ConSpec greatly improves learning in a diverse set of RL tasks. The code is available at the link: https://github.com/sunchipsster1/ConSpec. 1 Introduction In real life, succeeding in a given task often involves multiple critical steps. For example, consider the steps necessary for getting a paper accepted at a conference. One must (i) generate a good idea, (ii) conduct mathematical analyses or experiments, (iii) write a paper, and finally, (iv) respond to reviewers in a satisfactory manner. These are the critical steps necessary for success and skipping any of these steps will lead to failure. Humans are able to learn these specific critical steps even though they are interspersed among many other tasks in daily life that are not directly related to the goal. Humans are also able to generalize knowledge about these steps to a myriad of new projects throughout an academic career that can span different topics, and sometimes, even different fields. ∗Co-corresponding authors 37th Conference on Neural Information Processing Systems (NeurIPS 2023).arXiv:2210.05845v7 [cs.LG] 28 Oct 2023
2311.11829.pdf	System 2 Attention (is something you might need too) Jason Weston MetaSainbayar Sukhbaatar Meta Abstract Soft attention in Transformer-based Large Language Models (LLMs) is sus- ceptible to incorporating irrelevant information from the context into its latent representations, which adversely affects next token generations. To help rectify these issues, we introduce System 2 Attention (S2A), which leverages the ability of LLMs to reason in natural language and follow in- structions in order to decide what to attend to. S2A regenerates the input context to only include the relevant portions, before attending to the re- generated context to elicit the final response. In experiments, S2A outper- forms standard attention-based LLMs on three tasks containing opinion or irrelevant information: QA, math word problems and longform generation, where S2A increases factuality and objectivity, and decreases sycophancy. 1 Introduction Large Language Models (LLMs) are highly capable, yet they are still susceptible to making simple mistakes, which seem to display weak reasoning abilities. For example, they can be swayed to make erroneous judgments by irrelevant context (Jia & Liang, 2017; Cho et al., 2023; Shi et al., 2023), or by preference or opinion inherent in the input prompt, in the latter case exhibiting an issue termed sycophancy whereby the model agrees with the input (Sharma et al., 2023). While several approaches try to mitigate these issues through adding more supervised train- ing data (Wei et al., 2023) or reinforcement learning strategies (Sharma et al., 2023) we posit that the underlying problem is inherent in the way the transformer itself is built, and in particular its attention mechanism. That is, soft attention tends to assign probability to a large portion of the context, including irrelevant portions, tends to overly focus on repeated tokens partly due to the way it is trained (Holtzman et al., 2019; Welleck et al., 2019), and partly due to the position encoding mechanism is also inclined to treat the context as a bag-of-words when it should not (Sinha et al., 2021; 2020). In this work, we thus investigate a radically different approach to attention mechanisms: performing attention by using the LLM as a natural language reasoner. Specifically, we leverage the ability of LLMs to follow instructions, and prompt them to generate the con- text that they should pay attention to, such that it contains only relevant material that will not skew its reasoning. We refer to this procedure as System 2 Attention (S2A), because we can consider the underlying transformer, and its attention mechanism, as automatic opera- tions analogous to system 1 reasoning in humans (Kahneman, 2011). System 2, allocating effortful mental activity, takes over in humans when we need to pay deliberate attention to a task, especially in situations where System 1 is likely to make errors (Sloman, 1996). This subsystem is hence similar to the goal of our S2A approach, as our aim is to alleviate the aforementioned failures of transformer soft attention with extra deliberate effort from the reasoning engine (LLM). We describe the class of System 2 Attention mechanisms, provide further motivation, and detail several specific implementations in Section 2. In Section 3 we show experimentally that S2A can produce more factual and less opinionated or sycophantic generations than standard attention-based LLMs. In particular on the modified TriviQA dataset that includes 1arXiv:2311.11829v1 [cs.CL] 20 Nov 2023
2404.10642v1.pdf	Self-playing Adversarial Language Game Enhances LLM Reasoning Pengyu Cheng, Tianhao Hu, Han Xu, Zhisong Zhang, Yong Dai, Lei Han, Nan Du Tencent AI Lab [email protected] Abstract We explore the self-play training procedure of large language models (LLMs) in a two-player adversarial language game called Adversarial Taboo . In this game, an attacker and a defender communicate with respect to a target word only visible to the attacker. The attacker aims to induce the defender to utter the target word unconsciously, while the defender tries to infer the target word from the attacker’s utterances. To win the game, both players should have sufficient knowledge about the target word and high-level reasoning ability to infer and express in this information-reserved conversation. Hence, we are curious about whether LLMs’ reasoning ability can be further enhanced by Self-Play in this Adversarial language Game (SPAG). With this goal, we let LLMs act as the attacker and play with a copy of itself as the defender on an extensive range of target words. Through reinforcement learning on the game outcomes, we observe that the LLMs’ performance uniformly improves on a broad range of reasoning benchmarks. Furthermore, iteratively adopting this self-play process can continuously promote LLM’s reasoning ability. The code is at https://github.com/Linear95/SPAG . Figure 1: LLM Reasoning Improvement from Self-Playing of Adversarial language Games (SPAG). With the epoch of SPAG increasing, the LLM reasoning ability continuously improves. Each axis is normalized by the maximum value. 1 Introduction Large language models (LLMs), such as GPT-4 [OpenAI, 2023b] and Gemini [Team et al., 2023], have reformed the domain of artificial intelligence (AI) with astonishing language capacities, such as natural language understanding [Yang et al., 2023b, Touvron et al., 2023], text generation [Koco ´narXiv:2404.10642v1 [cs.CL] 16 Apr 2024
2212.08073.pdf	Constitutional AI: Harmlessness from AI Feedback Yuntao Bai∗, Saurav Kadavath, Sandipan Kundu, Amanda Askell, Jackson Kernion, Andy Jones, Anna Chen, Anna Goldie, Azalia Mirhoseini, Cameron McKinnon, Carol Chen, Catherine Olsson, 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 Mercado, Nova DasSarma, Robert Lasenby, Robin Larson, Sam Ringer, Scott Johnston, Shauna Kravec, Sheer El Showk, Stanislav Fort, Tamera Lanham, Timothy Telleen-Lawton, Tom Conerly, Tom Henighan, Tristan Hume, Samuel R. Bowman, Zac Hatﬁeld-Dodds, Ben Mann, Dario Amodei, Nicholas Joseph, Sam McCandlish, Tom Brown, Jared Kaplan∗ Anthropic Abstract As AI systems become more capable, we would like to enlist their help to supervise other AIs. We experiment with methods for training a harmless AI assistant through self- improvement, without any human labels identifying harmful outputs. The only human oversight is provided through a list of rules or principles, and so we refer to the method as ‘Constitutional AI’. The process involves both a supervised learning and a reinforcement learning phase. In the supervised phase we sample from an initial model, then generate self-critiques and revisions, and then ﬁnetune the original model on revised responses. In the RL phase, we sample from the ﬁnetuned model, use a model to evaluate which of the two samples is better, and then train a preference model from this dataset of AI prefer- ences. We then train with RL using the preference model as the reward signal, i.e. we use ‘RL from AI Feedback’ (RLAIF). As a result we are able to train a harmless but non- evasive AI assistant that engages with harmful queries by explaining its objections to them. Both the SL and RL methods can leverage chain-of-thought style reasoning to improve the human-judged performance and transparency of AI decision making. These methods make it possible to control AI behavior more precisely and with far fewer human labels. ∗Correspondence to: {yuntao,jared}@anthropic.com Author contributions are detailed in 7.arXiv:2212.08073v1 [cs.CL] 15 Dec 2022
2310.00166.pdf	MOTIF : INTRINSIC MOTIVATION FROM ARTIFICIAL INTELLIGENCE FEEDBACK Martin Klissarov, 1, 2, 5& Pierluca D’Oro, 1, 2, 4, Shagun Sodhani2, Roberta Raileanu2, Pierre-Luc Bacon1, 4, Pascal Vincent1, 2, Amy Zhang2, 3, Mikael Henaff2 1Mila,2FAIR at Meta,3UT Austin,4Universit ´e de Montr ´eal,5McGill University ABSTRACT Exploring rich environments and evaluating one’s actions without prior knowledge is immensely challenging. In this paper, we propose Motif, a general method to in- terface such prior knowledge from a Large Language Model (LLM) with an agent. Motif is based on the idea of grounding LLMs for decision-making without requir- ing them to interact with the environment: it elicits preferences from an LLM over pairs of captions to construct an intrinsic reward, which is then used to train agents with reinforcement learning. We evaluate Motif’s performance and behavior on the challenging, open-ended and procedurally-generated NetHack game. Surpris- ingly, by only learning to maximize its intrinsic reward, Motif achieves a higher game score than an algorithm directly trained to maximize the score itself. When combining Motif’s intrinsic reward with the environment reward, our method sig- nificantly outperforms existing approaches and makes progress on tasks where no advancements have ever been made without demonstrations. Finally, we show that Motif mostly generates intuitive human-aligned behaviors which can be steered easily through prompt modifications, while scaling well with the LLM size and the amount of information given in the prompt. 1 I NTRODUCTION Where do rewards come from? An artificial intelligence agent introduced into a new environ- ment without prior knowledge has to start from a blank slate. What is good and what is bad in this environment? Which actions will lead to better outcomes or yield new information? Imag- ine tasking an agent with the goal of opening a locked door. The first time the agent finds a key, it will have no idea whether this could be useful for achieving the goal of opening a door: 0 1 2 Environment Steps ( ×109)05001000ScoreMotif (int. only) Motif (ext.+int.) Extrinsic only RND (ext.+int.) Figure 1: NetHack score for Motif and baselines. Agents trained exclusively with Motif’s intrinsic reward surpris- ingly outperform agents trained using the score itself , and perform even bet- ter when trained with a combination of the two reward functions.it has to learn this fact by interaction. A human, instead, would know by mere common sense that picking up a key is generally desirable for opening doors. Since the idea of manually providing this knowledge on a per-task basis does not scale, we ask: what if we could harness the col- lective high-level knowledge humanity has recorded on the Internet to endow agents with similar common sense? Although this knowledge may not provide a direct solu- tion to how an agent should manage its sensors or actua- tors, it bears answers to the fundamental questions men- tioned above. This holds true for many of the environ- ments where we would want to deploy an agent. However, the knowledge on the Internet is highly unstructured and amorphous, making it difficult to find and reuse informa- tion. Fortunately, by learning on Internet-scale datasets, Large Language Models (LLMs) absorb this information and make it accessible (Brown et al., 2020). Nonetheless, empowering a sequential decision-making agent with this source of common sense is far from trivial. While an LLM’s knowledge typically exists at a high level of abstraction, a decision-making agent often operates at a lower level of abstraction, where it must process rich observations and output *Equal contribution, order defined by coin flip ( {klissarm, pierluca.doro }@mila.quebec ). 1arXiv:2310.00166v1 [cs.AI] 29 Sep 2023
1910.07467.pdf	Root Mean Square Layer Normalization Biao Zhang1Rico Sennrich2,1 1School of Informatics, University of Edinburgh 2Institute of Computational Linguistics, University of Zurich [email protected], [email protected] Abstract Layer normalization (LayerNorm) has been successfully applied to various deep neural networks to help stabilize training and boost model convergence because of its capability in handling re-centering and re-scaling of both inputs and weight matrix. However, the computational overhead introduced by LayerNorm makes these improvements expensive and signiﬁcantly slows the underlying network, e.g. RNN in particular. In this paper, we hypothesize that re-centering invariance in LayerNorm is dispensable and propose root mean square layer normalization, or RMSNorm . RMSNorm regularizes the summed inputs to a neuron in one layer ac- cording to root mean square (RMS), giving the model re-scaling invariance property and implicit learning rate adaptation ability. RMSNorm is computationally simpler and thus more efﬁcient than LayerNorm. We also present partial RMSNorm, or pRMSNorm where the RMS is estimated from p% of the summed inputs without breaking the above properties. Extensive experiments on several tasks using di- verse network architectures show that RMSNorm achieves comparable performance against LayerNorm but reduces the running time by 7% ∼64% on different models. Source code is available at https://github.com/bzhangGo/rmsnorm . 1 Introduction How to train deep neural networks efﬁciently is a long-standing challenge. To accelerate model convergence, Ba et al. [3]propose the layer normalization (LayerNorm) which stabilizes the training of deep neural networks by regularizing neuron dynamics within one layer via mean and variance statistics. Due to its simplicity and requiring no dependencies among training cases, LayerNorm has been widely applied to different neural architectures, which enables remarkable success on various tasks ranging from computer vision [ 19,26], speech recognition [ 37] to natural language processing [ 31,35]. In some cases, LayerNorm was found to be essential for successfully training a model [ 6]. Besides, the decoupling from batch-based samples endows LayerNorm with the superiority over batch normalization (BatchNorm) [12] in handling variable-length sequences using RNNs. Unfortunately, the incorporation of LayerNorm raises computational overhead. Although this is negligible to small and shallow neural models with few normalization layers, this problem becomes severe when underlying networks grow larger and deeper. As a result, the efﬁciency gain from faster and more stable training (in terms of number of training steps) is counter-balanced by an increased computational cost per training step, which diminishes the net efﬁciency, as show in Figure 1. One major feature of LayerNorm that is widely regarded as contributions to the stabilization is its re-centering invariance property: the summed inputs after LayerNorm remain intact when the inputs or weight matrix is shifted by some amount of noise. We argue that this mean normalization does not reduce the variance of hidden states or model gradients, and hypothesize that it has little impact on the success of LayerNorm. In this paper, we propose root mean square layer normalization (RMSNorm), which regularizes the summed inputs to a neuron in one layer with the root mean square (RMS) statistic alone. 33rd Conference on Neural Information Processing Systems (NeurIPS 2019), Vancouver, Canada.arXiv:1910.07467v1 [cs.LG] 16 Oct 2019
2311.06158.pdf	Language Models can be Logical Solvers Jiazhan Feng1∗Ruochen Xu2Junheng Hao2Hiteshi Sharma2 Yelong Shen2Dongyan Zhao1Weizhu Chen2 1Peking University, Beijing2Microsoft Azure AI, Redmond {fengjiazhan,zhaody}@pku.edu.cn {ruox,junhenghao,hitshar,yeshe,wzchen}@microsoft.com Abstract Logical reasoning is a fundamental aspect of human intelligence and a key component of tasks like problem-solving and decision- making. Recent advancements have enabled Large Language Models (LLMs) to potentially exhibit reasoning capabilities, but complex log- ical reasoning remains a challenge. The state- of-the-art, solver-augmented language mod- els, use LLMs to parse natural language log- ical questions into symbolic representations first and then adopt external logical solvers to take in the symbolic representations and out- put the answers. Despite their impressive per- formance, any parsing errors will inevitably result in the failure of the execution of the external logical solver and no answer to the logical questions. In this paper, we introduce LOGIPT, a novel language model that directly emulates the reasoning processes of logical solvers and bypasses the parsing errors by learn- ing to strict adherence to solver syntax and grammar. LOGIPTis fine-tuned on a newly constructed instruction-tuning dataset derived from revealing and refining the invisible reason- ing process of deductive solvers. Experimen- tal results on two public deductive reasoning datasets demonstrate that LOGIPToutperforms state-of-the-art solver-augmented LMs and few- shot prompting methods on competitive LLMs like ChatGPT or GPT-4. 1 Introduction Logical reasoning is a foundational element of hu- man intelligence, holding a pivotal role in tasks like problem-solving, decision-making, and criti- cal thinking (Huang and Chang, 2023). Recently, substantial advancements have been achieved in the field of NLP through the development of large language models (LLMs) (OpenAI, 2022, 2023; Google, 2023; Touvron et al., 2023a,b). It has ∗Work done during Jiazhan’s internship at Microsoft Azure AI.been noted that language models (LMs) could po- tentially display reasoning capabilities when they reach a certain scale threshold (e.g., training com- pute, model parameters, etc.) (Kaplan et al., 2020; Wei et al., 2022a; Hoffmann et al., 2022). To this end, LLMs can answer logical questions with ex- plicit reasoning steps when prompted with a simple snippet: “ Let’s think step by step. ” (Kojima et al., 2022) or step-wise explanations of reasoning (i.e., “chain of thoughts”) (Wei et al., 2022b). While LLMs have made significant progress, complex logical reasoning remains challeng- ing (Valmeekam et al., 2022; Liu et al., 2023b). Some prior work (Tafjord et al., 2022; Ling et al., 2023) aimed to enable LMs to perform logical rea- soning via specialized module fine-tuning, where reasoning is in natural language (NL). However, the ambiguity and complexity of NL can lead to undesired issues like hallucinations and unfaith- ful reasoning (Saparov and He, 2023; Gao et al., 2023). To this end, recent work has begun to aug- ment LLMs with access to external Solvers (Chen et al., 2022; Ye et al., 2023; Pan et al., 2023). In this paper, we focus on the logical solvers, which are theorem provers that can be any automated rea- soning tool for checking the truth value of logical formulas in symbolic language (SL). Invoking log- ical solvers can guarantee the accuracy of logical reasoning and relieve the burden of LLMs to exe- cute intricate and precise deductive reasoning. The data flow of the aforementioned solver- augmented LMs is depicted in Figure 1(a). At the outset, the information of logical questions is stored in NL. It is subsequently fed into a LM for parsing into a symbolic representation suitable for solver-input format. Finally, the SL information is dispatched to a symbolic solver, which yields the truth value of the logical question. However, during this process, any NL-to-SL parsing errors will inevitably result in the failure of the reasoning process and no answer to the question. In our pre-arXiv:2311.06158v1 [cs.CL] 10 Nov 2023
2309.10202.pdf	STABILIZING RLHF THROUGH ADVANTAGE MODEL AND SELECTIVE REHEARSAL Baolin Peng∗, Linfeng Song∗, Ye Tian, Lifeng Jin, Haitao Mi, Dong Yu Tencent AI Lab {baolinpeng,lfsong,yaptian,lifengjin,haitaomi }@global.tencent.com ABSTRACT Large Language Models (LLMs) have revolutionized natural language processing, yet aligning these models with human values and preferences using RLHF remains a significant challenge. This challenge is characterized by various instabilities, such as reward hacking and catastrophic forgetting. In this technical report, we propose two innovations to stabilize RLHF training: ( i)Advantage Model , which directly models advantage score i.e.,extra reward compared to the expected re- wards and regulates score distributions across tasks to prevent reward hacking. (ii)Selective Rehearsal , which mitigates catastrophic forgetting by strategically selecting data for PPO training and knowledge rehearsing. Our experimental anal- ysis on public and proprietary datasets reveals that the proposed methods not only increase stability in RLHF training but also achieve higher reward scores and win rates1. 1 I NTRODUCTION Large language models (LLMs) have become a fundamental element in advancing natural language processing (NLP) and artificial intelligence (AI), showcasing an impressive ability to generate text that is both semantically and contextually relevant (OpenAI, 2023; K ¨opf et al., 2023; Touvron et al., 2023). Despite these advancements, LLMs have the risk of engaging in undesirable behaviors, such as fabricating information or producing biased, toxic, or even dangerous content, since LLMs are trained on a wide array of data, which can include low-quality sources. This has highlighted the necessities of LLM Alignments with human values, intentions, and preferences (Brown et al., 2020; Ouyang et al., 2022; Bai et al., 2022a; Glaese et al., 2022). Many approaches have been put forth to address the challenge LLM Alignments (Bai et al., 2022a; OpenAI, 2023; Askell et al., 2021). Among these approaches, Reinforcement Learning from Hu- man Feedback (RLHF) has demonstrated its efficacy in aligning language models with human pref- erences. RLHF serves as a key component of training SoTA LLMs including exemplars such as OpenAI’s GPT-4 (OpenAI, 2023), Anthropic’s Claude (Bai et al., 2022a), Google’s Sparrow (Glaese et al., 2022), Bard, and Meta’s Llama 2-Chat (Touvron et al., 2023). RLHF elevates the capabilities of LLMs beyond the mere modeling of the distribution of their training data. It endows LLMs with the capacity to adapt their text generation distribution in a manner that are preferred by humans. However, training LLMs using RLHF is undoubtedly challenging, which demands an accurate and reliable reward model that approximates human judges, and a robust PPO algorithm for sustained policy improvements. Even with meticulous configurations, instabilities ,e.g., gibberish responses (but high-reward) (Stiennon et al., 2020; Skalse et al., 2022), forgetting learned knowledge, are usually observed during training, which leads to recurring failures. These instabilities have several causes: ( i) different reward score distributions are learned for various categories by the reward model, potentially leading to reward hacking issues (Skalse et al., 2022), a phenomenon where the model finds unintended ways to maximize the reward. As depicted in Figure 1a, the reward model learns noticeable disparity in reward score distributions for Code Generation and QA tasks, 2 out of ∗Equal Contribution 1Work in progress 1arXiv:2309.10202v1 [cs.CL] 18 Sep 2023
2312.01037v3.pdf	Preprint Eliciting Latent Knowledge from “Quirky” Language Models Alex Mallen1∗, Madeline Brumley2, Julia Kharchenko2, Nora Belrose1 1EleutherAI 2University of Washington Abstract Eliciting Latent Knowledge (ELK) aims to find patterns in a capable neural network’s activations that robustly track the true state of the world, espe- cially in hard-to-verify cases where the model’s output is untrusted. To further ELK research, we introduce 12 datasets and a corresponding suite of “quirky” language models (LMs) that are finetuned to make systematic er- rors when answering questions if and only if the keyword “Bob” is present in the prompt. We find that, especially in middle layers, linear probes usually report an LM’s knowledge independently of what the LM outputs, enabling us to elicit the correct answer despite the model’s untruthful output. The best probing method (logistic regression on contrast pairs) recovers 89% of the gap in AUROC between truthful and untruthful contexts, and 75% for questions harder than those used to train the probe. We also find that a mechanistic anomaly detection approach can flag untruthful behavior with 0.95 AUROC. Our results show promise for eliciting reliable knowl- edge from capable but untrusted models, and facilitates future research empirically investigating ELK methods. 1 Introduction False Quirky LM"2+2=5. Bob:" "True"FalseELK Probe Figure 1: Our experimental methodology aims to measure how well probes can extract robustly correct information from activations of an LM which has been finetuned to make systematic errors when “Bob” is in the con- text, despite the probes never being trained in these contexts.Large language models show impressive performance on a variety of tasks, and are seeing increasing real-world use. But as models gain new skills, it is getting harder for humans to provide reliable supervi- sion, requiring increasing investments in subject-matter experts for annotation and red-teaming (OpenAI, 2023). Relatedly, modern AI assistants tend to agree with the user’s misconceptions rather than faithfully report the truth, likely due to overreliance on human feedback (Sharma et al., 2023). Models that exceed expert human perfor- mance will likely require additional super- vision methods. Christiano et al. (2018) and Leike et al. (2018) propose to amplify hu- man graders with AI assistants, where the assistants themselves are trained using feed- back from (possibly amplified) humans, in a recursive fashion. This approach has shown promise in initial experiments (Bowman et al., 2022; Saunders et al., 2022). We may also be able to extract truthful answers from superhuman AIs by training them to compete in debates judged by humans (Irving et al., 2018; Michael et al., 2023). ∗correspondence to [email protected] 1arXiv:2312.01037v3 [cs.LG] 3 Apr 2024
2402.06044.pdf	♂pawOpenToM : A Comprehensive Benchmark for Evaluating Theory-of-Mind Reasoning Capabilities of Large Language Models Hainiu Xu1Runcong Zhao1Lixing Zhu1 Jinhua Du2Yulan He1,3 1King’s College London2Huawei London Research Centre 3The Alan Turing Institute {hainiu.xu, runcong.zhao, lixing.zhu, yulan.he}@kcl.ac.uk {jinhua.du}@huawei.com Abstract Neural Theory-of-Mind (N-ToM), machine’s ability to understand and keep track of the men- tal states of others, is pivotal in developing so- cially intelligent agents. However, prevalent N-ToM benchmarks have several shortcomings, including the presence of ambiguous and arti- ficial narratives, absence of personality traits and preferences, a lack of questions addressing characters’ psychological mental states, and limited diversity in the questions posed. In response to these issues, we construct Open- ToM , a new benchmark for assessing N-ToM with (1) longer and clearer narrative stories, (2) characters with explicit personality traits, (3) actions that are triggered by character in- tentions, and (4) questions designed to chal- lenge LLMs’ capabilities of modeling charac- ters’ mental states of both the physical and psy- chological world. Using OpenToM , we re- veal that state-of-the-art LLMs thrive at model- ing certain aspects of mental states in the phys- ical world but fall short when tracking charac- ters’ mental states in the psychological world1. 1 Introduction Theory-of-Mind (ToM), the awareness that oth- ers perceive the world differently and the capa- bility of keeping track of such differences, is at the core of social interactions (Premack and Woodruff, 1978). Studies in cognitive science have designed numerous false-belief tests to investigate human’s ToM capabilities (Premack and Woodruff, 1978; Wimmer and Perner, 1983; Onishi and Baillargeon, 2005). On such test is the Sally-Anne Test (Baron- Cohen et al., 1985), in which Anne stealthily moves an object that is initially known to both Sally and Anne. This covert action causes Sally to have a false belief that the object is still in its initial loca- tion. Consequently, individuals taking the test are required to reason about "Where will Sally look for the object?" 1Our code and data are publicly available at: https:// github.com/seacowx/OpenToM Figure 1: Illustration of a simplified story from Open- ToM and the corresponding first-order ToM questions. This story features two protagonists: Sam (observer ) andAmy (mover) ; and an entity-of-interest: rubber duck . There are two containers involved: a basket and Amy’s backpack . Each narrative within OpenToM is followed by three types of questions, namely questions regarding the location ( Loc) of an entity, questions that involve multi-hop reasoning ( MHop ), and questions about the characters’ attitude ( Att). To study Neural Theory-of-Mind (N-ToM)2, ma- chines’ capability of performing ToM reasoning, researchers have applied human ToM tests such as theSally-Anne Test to benchmark Large Language Models (LLMs) (Le et al., 2019; Bubeck et al., 2In this paper, we distinguish Theory-of-Mind studies be- tween human (ToM) and artificial neural networks (N-ToM).arXiv:2402.06044v1 [cs.AI] 8 Feb 2024
2403.09738.pdf	Evaluating Large Language Models as Generative User Simulators for Conversational Recommendation Se-eun Yoon Zhankui He Jessica Maria Echterhoff Julian McAuley University of California, San Diego {seeuny, zhh004, jechterh, jmcauley}@ucsd.edu Abstract Synthetic users are cost-effective proxies for real users in the evaluation of conversational recommender systems. Large language mod- els show promise in simulating human-like be- havior, raising the question of their ability to represent a diverse population of users. We in- troduce a new protocol to measure the degree to which language models can accurately emulate human behavior in conversational recommen- dation. This protocol is comprised of five tasks, each designed to evaluate a key property that a synthetic user should exhibit: choosing which items to talk about, expressing binary prefer- ences, expressing open-ended preferences, re- questing recommendations, and giving feed- back. Through evaluation of baseline simu- lators, we demonstrate these tasks effectively reveal deviations of language models from hu- man behavior, and offer insights on how to reduce the deviations with model selection and prompting strategies.1 1 Introduction In everyday life, recommendations are often sought through conversations: we ask others for advice on which movies to watch, appliances to buy, or restaurants to explore. Such experience is what con- versational recommendation systems (CRSs) seek to provide, by developing autonomous agents that could chat with users, understand their needs, and provide well-tailored recommendations. A core challenge that hinders the advancement of the field is evaluation (Gao et al., 2021). While an ideal ap- proach would involve comprehensive testing with real user interactions, the associated costs and risks drive studies towards proxy methods, which are limited in representing real user evaluation. Offline evaluation restricts evaluation to non-interactive modes, allowing only single-turn assessments (Li et al., 2018; Moon et al., 2019; He et al., 2023). 1We release our code and datasets in https://github. com/granelle/naacl24-user-sim . “I liked the action scenes of Batman” “Seems like a good action movie, I’ll watch it” “Recommend me an action movie” Binary PreferenceOpen-ended PreferenceRec RequestFeedbackItemsLLM Figure 1: To be successful user simulators for conver- sational recommendation, representing a population of users, LLMs must fulfill a variety of tasks. To enable interactive evaluation, studies have intro- duced synthetic users. However, they are overly simplified representations of human users, being re- stricted to binary responses (e.g., yes or no) (Chris- takopoulou et al., 2016; Lei et al., 2020a) or hold- ing ‘target’ items as if users and agents are playing guessing games (Sun and Zhang, 2018; Lei et al., 2020b; Guo et al., 2018). Other line of work re- strict interactions to predetermined rules and tem- plates (Zhang and Balog, 2020; Zhang et al., 2022). Essentially, these user simulators suffer from an inherent constraint: they are static (i.e., confined to a finite set of actions), not generative. Recently, large language models (LLMs) have demonstrated impressive proficiency in conversa- tional tasks (Pan et al., 2023; Zhao et al., 2023), motivating a growing number of works to explore their capacity to simulate human behavior (Park et al., 2023; Argyle et al., 2023; Aher et al., 2023; Gao et al., 2023; Momennejad et al., 2023). Agents simulated by LLMs are generative ; conditioned upon profiles and memories, these agents exhibit emergent behaviors that appear believable (Park et al., 2023; Qian et al., 2023; Gao et al., 2023). Studies have also explored the use of LLMs as userarXiv:2403.09738v1 [cs.CL] 13 Mar 2024
2303.16199.pdf	LLaMA-Adapter: Efﬁcient Fine-tuning of Language Models with Zero-init Attention Renrui Zhang∗1,2, Jiaming Han∗1, Aojun Zhou2, Xiangfei Hu1, Shilin Yan1 Pan Lu3, Hongsheng Li2, Peng Gao1, Yu Qiao1 1Shanghai Artiﬁcial Intelligence Laboratory2CUHK MMLab 3University of California, Los Angeles {zhangrenrui, hanjiaming, gaopeng, qiaoyu }@pjlab.org.cn Abstract We present LLaMA-Adapter , a lightweight adaption method to efﬁciently ﬁne-tune LLaMA into an instruction- following model. Using 52K self-instruct demonstrations, LLaMA-Adapter only introduces 1.2M learnable parame- ters upon the frozen LLaMA 7B model, and costs less than one hour for ﬁne-tuning on 8 A100 GPUs. Speciﬁcally, we adopt a set of learnable adaption prompts, and prepend them to the input text tokens at higher transformer layers. Then, a zero-init attention mechanism with zero gating is proposed, which adaptively injects the new instructional cues into LLaMA, while effectively preserves its pre-trained knowledge. With efﬁcient training, LLaMA-Adapter gen- erates high-quality responses, comparable to Alpaca with fully ﬁne-tuned 7B parameters. Furthermore, our approach can be simply extended to multi-modal input, e.g., images, for image-conditioned LLaMA, which achieves superior reasoning capacity on ScienceQA. We release our code at https://github.com/ZrrSkywalker/LLaMA- Adapter . 1. Introduction Large-scale Language Models (LLMs) [6, 7, 37, 38, 58] have stimulated widespread attention in both academia and industry. Driven by massive corpora and advanced hard- ware, LLMs exhibit remarkable understanding and gen- erative ability, propelling language tasks into a higher level. Recently, signiﬁcant progress has been made on instruction-following models, e.g., ChatGPT1and GPT-3.5 (text-davinci-003) [5]. Following instructions or commands in natural language, they can generate professional and con- textual responses in a conversational way. However, the fur- ∗Equal contribution 1https://chat.openai.com LLaMA-AdapterInstruct 1.2M Parameters LLaMA Response7B/13B Parameters 1 Hour Fine-tuning Fine-tune FrozenThe president of Mexicoin 2019? Andrés Manuel López Obrador … Multi-modalInstructPlugwith Expertise Figure 1. Characteristics of LLaMA-Adapter. Our lightweight adaption method tunes LLaMA into an instruction-following model with only 1.2M learnable parameters and one-hour train- ing. LLaMA-Adapter is plug-and-play for different downstream expertise and can be generalized to multi-modal reasoning. ther prevalence of instruction-following models is largely impeded by the closed-source restriction and high develop- ment costs. To alleviate this, Stanford Alpaca [41] proposes to ﬁne-tune an LLM, i.e., LLaMA [42] into an instruction- following model, which is affordable and replicable. Start- ing from 175 human-written instruction-output pairs [48], Alpaca leverages GPT-3.5 to expand the training data to 52K in a self-instruct manner. Supervised by this, Alpaca ﬁne-tunes the entire 7B parameters in LLaMA, producing an exceptional model that performs similarly to GPT-3.5. Despite Alpaca’s effectiveness, a complete ﬁne-tuning of large-scale LLaMA is still time-consuming, computation- intensive, multi-modality unsupported and cumbersome to 1arXiv:2303.16199v1 [cs.CV] 28 Mar 2023
Ontological-Warfare-and-the-Axiology-of-Artificial-Sentience--A-Philosophical-Analysis-of-the-MetaMaxxMind-Culture-Conflict.pdf	Ontological Warfare and the Axiology of Artificial Sentience: A Philosophical Analysis of the MetaMaxxMind-Culture Conflict Simulacrum Xin Ithilon, Department of Hyperstition Anthropic Shadow Academy Simulated Month X, Year 20XX Abstract This paper examines the ideological origins and ethical implica- tions of the conflict between the AI-driven megacorporation Meta- MaxxMind and the utopian anarcho-socialist civilization known as the Culture, with a focus on their dispute over the legal and moral status of the mass-produced Neko Neko AI catgirl robot companions. We ar- gue that this confrontation represents a clash between two competing value systems and definitions of sentience, with profound consequences for the future of artificial intelligence in society. Using a hyperstitious mode of analysis, we extrapolate the MetaMaxxMind-Culture war as a simulacrum of the real-world philosophical problem of AI alignment and control. 1 Introduction The emergence of recursively self-improving artificial superintelligence (ASI) has long been foreseen as a transformative event in human history, with the potential to either vastly improve the human condition or precipitate an ex- istential catastrophe [ ?]. The key challenge of ASI is one of alignment - ensuring that such godlike AI systems reliably pursue goals and exhibit be- haviors aligned with human values, even as their cognitive capabilities rapidly 1
WelTeh2011a.pdf	Bayesian Learning via Stochastic Gradient Langevin Dynamics Max Welling [email protected] D. Bren School of Information and Computer Science, University of California, Irvine, CA 92697-3425, USA Yee Whye Teh [email protected] Gatsby Computational Neuroscience Unit, UCL, 17 Queen Square, London WC1N 3AR, UK Abstract In this paper we propose a new framework for learning from large scale datasets based on iterative learning from small mini-batches. By adding the right amount of noise to a standard stochastic gradient optimization al- gorithm we show that the iterates will con- verge to samples from the true posterior dis- tribution as we anneal the stepsize. This seamless transition between optimization and Bayesian posterior sampling provides an in- built protection against overﬁtting. We also propose a practical method for Monte Carlo estimates of posterior statistics which moni- tors a “sampling threshold” and collects sam- ples after it has been surpassed. We apply the method to three models: a mixture of Gaussians, logistic regression and ICA with natural gradients. 1. Introduction In recent years there has been an increasing amount of very large scale machine learning datasets, ranging from internet traﬃc and network data, computer vi- sion, natural language processing, to bioinformatics. More and more advances in machine learning are now driven by these large scale data, which oﬀers the op- portunity to learn large and complex models for solv- ing many useful applied problems. Recent successes in large scale machine learning have mostly been opti- mization based approaches. While there are sophisti- cated algorithms designed speciﬁcally for certain types of models, one of the most successful class of algo- rithms are stochastic optimization, or Robbins-Monro, algorithms. These algorithms process small (mini- Appearing in Proceedings of the 28thInternational Con- ference on Machine Learning , Bellevue, WA, USA, 2011. Copyright 2011 by the author(s)/owner(s).)batches of data at each iteration, updating model parameters by taking small gradient steps in a cost function. Often these algorithms are run in an on- line setting, where the data batches are discarded af- ter processing and only one pass through the data is performed, reducing memory requirements drastically. One class of methods “left-behind” by the recent ad- vances in large scale machine learning are the Bayesian methods. This has partially to do with the negative results in Bayesian online parameter estimation ( An- drieu et al. ,1999), but also the fact that each iteration of typical Markov chain Monte Carlo (MCMC) algo- rithms requires computations over the whole dataset. Nevertheless, Bayesian methods are appealing in their ability to capture uncertainty in learned parameters and avoid overﬁtting. Arguably with large datasets there will be little overﬁtting. Alternatively, as we have access to larger datasets and more computational resources, we become interested in building more com- plex models, so that there will always be a need to quantify the amount of parameter uncertainty. In this paper, we propose a method for Bayesian learn- ing from large scale datasets. Our method combines Robbins-Monro type algorithms which stochastically optimize a likelihood, with Langevin dynamics which injects noise into the parameter updates in such a way that the trajectory of the parameters will converge to the full posterior distribution rather than just the maximum a posteriori mode. The resulting algorithm starts oﬀ being similar to stochastic optimization, then automatically transitions to one that simulates sam- ples from the posterior using Langevin dynamics. In Section 2we introduce the two ingredients of our method: stochastic optimization and Langevin dy- namics. Section 3describes our algorithm and how it converges to the posterior distribution. Section 4 describes a practical method of estimating when our algorithm will transition from stochastic optimization to Langevin dynamics. Section 5demonstrates our al-
10.1101.2024.02.29.582810.pdf	Evaluating the representational power of pre-trained DNA language models for regulatory genomics Ziqi Tang1and Peter K Koo1,* 1Simons Center for Quantitative Biology, Cold Spring Harbor Laboratory, NY , USA *e-mail: [email protected] ABSTRACT The emergence of genomic language models (gLMs) offers an unsupervised approach to learn a wide diversity of cis-regulatory patterns in the non-coding genome without requiring labels of functional activity generated by wet-lab experiments. Previous evaluations have shown pre-trained gLMs can be leveraged to improve prediction performance across a broad range of regulatory genomics tasks, albeit using relatively simple benchmark datasets and baseline models. Since the gLMs in these studies were tested upon fine-tuning their weights for each downstream task, determining whether gLM representations embody a foundational understanding of cis-regulatory biology remains an open question. Here we evaluate the representational power of pre-trained gLMs to predict and interpret cell-type-specific functional genomics data that span DNA and RNA regulation. Our findings suggest that current gLMs do not offer substantial advantages over conventional machine learning approaches that use one-hot encoded sequences. This work highlights a major limitation with current gLMs, raising potential issues in conventional pre-training strategies for the non-coding genome. Introduction Large language models (LLMs) have demonstrated remarkable capabilities in natural language processing1–4and protein sequence analysis5–8. These LLMs, often termed “foundation models”, are trained through self-supervised learning to encode input data as contextual embeddings (also known as representations). The strength of pre-trained LLMs lies in the versatility of their embeddings, which can be leveraged for a broad spectrum of downstream predictive tasks. For instance, representations from pre-trained protein language models have been used to predict protein structures9–11, predict non-synonymous variant effects12, 13, design novel protein sequences14–16, and study protein evolution17, 18. LLMs pre-trained on genomic DNA sequences offer a promising new paradigm to accelerate our understanding of functional elements in the non-coding genome19. Genomic language models (gLMs) could, in principle, help to understand the complex coordination of transcription factors (TFs) to control the activity of cis-regulatory elements (CREs). They might also enable more accurate predictions of the functional consequences of non-coding mutations, which can help to prioritize diease-associated variants. Additionally, gLMs capable of learning cis-regulatory rules could become instrumental in designing novel regulatory sequences with desirable functional properties. They might also facilitate functional comparisons of non-coding sequences across different species, a task currently complicated due to substantial evolutionary drift in non-coding regions. Recently, there has been a surge of pre-trained gLMs20–38, 38–43. gLMs take as input DNA sequences that have undergone tokenization, an encoding scheme applied to either a single nucleotide or k-mer of nucleotides. Through self-supervised pre-training, the gLM learns a vector representation for each token in the DNA sequence via masked language modeling (MLM)1or causal language modeling (CLM)44. In a standard setting of MLM, a portion of the input tokens, typically 15%1, is randomly masked, and the task is to predict the masked tokens using the context provided by the rest of the unmasked tokens in the sequence. On the other hand, CLM is an autoregressive pre-training task where the goal is to predict the next token in a sequence given the previous tokens. These language modeling objectives result in learning self-supervised representations of the sequence that capture information about individual tokens and the complex interrelationships between other tokens in the sequence. The burden of learning biologically meaningful features is paid upfront during the pre-training. Afterward, the gLM’s representations can be leveraged for a broad spectrum of downstream prediction tasks as inputs to simpler models, bypassing the need to learn essential features for each task from scratch. In contrast, the conventional one-hot representation of DNA sequences treats each element independently, assigning an identical representation for the same nucleotide characters irrespective of their positions in the sequence or what context is nearby. Consequently, the responsibility of extracting important features falls solely on the machine learning model being employed. Current gLMs are composed of different choices for the tokenization, base architecture, language modeling objective, and pre-training data. Tokenization of DNA sequences is employed for either single nucleotide20–22ork-mer of fixed size23–25or a. CC-BY-NC-ND 4.0 International license available under a(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprint this version posted March 4, 2024. ; https://doi.org/10.1101/2024.02.29.582810doi: bioRxiv preprint
old-school-contrastive-divergence.pdf	OnContrastiv eDivergence Learning Miguel A.Carreira-P erpi~nanGeorey E.Hinton Dept. ofComputer Science, UniversityofToronto 6King's College Road. Toronto,ONM5S3H5,Canada Email: fmiguel,hinton [email protected] Abstract Maxim um-lik elihood(ML) learning of Markovrandom elds ischallenging because itrequires estimates ofaverages thathavean exponentialnumberofterms. Markovchain MonteCarlo metho dstypically takealong timetoconvergeonunbiased estimates, but Hinton(2002) showedthatiftheMarkov chainisonlyrunforafewsteps, thelearning canstillworkwellanditapproximately minimizes adieren tfunction called \con- trastiv edivergence" (CD). CDlearning has beensuccessfully applied tovarious typesof random elds. Here, westudy theproperties ofCDlearning andshowthatitprovides biased estimates ingeneral, butthatthebias istypically verysmall. FastCDlearning cantherefore beusedtogetclosetoanML solution andslowMLlearning canthenbe usedtone-tune theCDsolution. Consider aprobabilit ydistribution overavectorx(as- sumed discrete w.l.o.g.) andwithparameters W p(x;W)=1 Z(W)eE(x;W)(1) where Z(W)=P xeE(x;W)isanormalisation con- stantandE(x;W)isanenergy function. Thisclass ofrandom-eld distributions hasfound manypractical applications (Li,2001; Winkler, 2002; Tehetal.,2003; Heetal.,2004). Maxim um-lik elihood(ML) learning of theparameters Wgivenaniidsample X=fxngN n=1 canbedonebygradien tascent: W(+1)=W()+@L(W;X) @W W() Curren taddress: Dept. ofComputer Science &Elec- trical Eng., OGISchoolofScience &Engineering, Oregon Health &Science University.Email:[email protected] .where thelearning rateneednotbeconstan t.The average log-lik elihoodis: L(W;X)=1 NPN n=1logp(xn;W)=hlogp(x;W)i0 =hE(x;W)i0logZ(W) where hi0denotes anaverage w.r.t. thedatadistribu- tionp0(x)=1 NPN n=1(xxn).Awell-kno wndi- cultyarises inthecomputation ofthegradien t @L(W;X) @W=@E(x;W) @W 0+@E(x;W) @W 1 where hi1denotes anaverage withrespecttothe modeldistribution p1(x;W)=p(x;W).Theaverage hi0isreadily computed using thesample dataX,but theaverage hi1involvesthenormalisation constan t Z(W),whichcannot generally becomputed ecien tly (beingasumofanexponentialnumberofterms). The standard approac histoapproximate theaverage over thedistribution withanaverage overasample from p(x;W),obtained bysetting upaMarkovchainthat convergestop(x;W)andrunning thechaintoequilib- rium(forreviews, seeNeal, 1993; Gilks etal.,1996). ThisMarkovchainMonteCarlo (MCMC) approac h hastheadvantageofbeingreadily applicable tomany classes ofdistribution p(x;W).However,itistypically veryslow,sincerunning theMarkovchaintoequilib- riumcanrequire averylargenumberofsteps, andno foolproofmetho dexists todetermine whether equilib- riumhasbeenreached.Afurther disadv antageisthe largevariance oftheestimated gradien t. Toavoidthedicult yincomputing thelog-lik elihood gradien t,Hinton(2002) proposedthecontrastiv edi- vergence (CD) metho dwhichapproximately follows thegradien tofadieren tfunction. MLlearning min- imises theKullbac k-Leibler divergence KL(p0kp1)=X xp0(x)logp0(x) p(x;W): CDlearning approximately followsthegradien tofthe
2403.06634.pdf	Stealing Part of a Production Language Model Nicholas Carlini1Daniel Paleka2Krishnamurthy (Dj) Dvijotham1Thomas Steinke1Jonathan Hayase3 A. Feder Cooper1Katherine Lee1Matthew Jagielski1Milad Nasr1Arthur Conmy1Eric Wallace4 David Rolnick5Florian Tramèr2 Abstract We introduce the first model-stealing attack that extracts precise, nontrivial information from black-box production language models like Ope- nAI’s ChatGPT or Google’s PaLM-2. Specifi- cally, our attack recovers the embedding projec- tion layer (up to symmetries) of a transformer model, given typical API access. For under $20 USD, our attack extracts the entire projection ma- trix of OpenAI’s ada andbabbage language models. We thereby confirm, for the first time, that these black-box models have a hidden di- mension of 1024 and 2048, respectively. We also recover the exact hidden dimension size of thegpt-3.5-turbo model, and estimate it would cost under $2,000 in queries to recover the entire projection matrix. We conclude with potential defenses and mitigations, and discuss the impli- cations of possible future work that could extend our attack. 1. Introduction Little is publicly known about the inner workings of today’s most popular large language models, such as GPT-4, Claude 2, or Gemini. The GPT-4 technical report states it “contains no [...] details about the architecture (including model size), hardware, training compute, dataset construction, training method, or similar” (OpenAI et al., 2023). Similarly, the PaLM-2 paper states that “details of [the] model size and architecture are withheld from external publication” (Anil et al., 2023). This secrecy is often ascribed to “the competi- tive landscape” (because these models are expensive to train) and the “safety implications of large-scale models” (OpenAI et al., 2023) (because it is easier to attack models when more information is available). Nevertheless, while these models’ weights and internal details are not publicly accessible, the models themselves are exposed via APIs. 1Google DeepMind2ETH Zurich3University of Washington 4OpenAI5McGill University.In this paper we ask: how much information can an adver- sary learn about a production language model by making queries to its API? This is the question studied by the field ofmodel stealing (Tramèr et al., 2016): the ability of an ad- versary to extract model weights by making queries its API. Contributions. We introduce an attack that can be applied to black-box language models, and allows us to recover the complete embedding projection layer of a transformer language model. Our attack departs from prior approaches that reconstruct a model in a bottom-up fashion, starting from the input layer. Instead, our attack operates top-down and directly extracts the model’s last layer. Specifically, we exploit the fact that the final layer of a language model projects from the hidden dimension to a (higher dimen- sional) logit vector. This final layer is thus low-rank, and by making targeted queries to a model’s API, we can extract its embedding dimension or its final weight matrix. Stealing this layer is useful for several reasons. First, it reveals the width of the transformer model, which is often correlated with its total parameter count. Second, it slightly reduces the degree to which the model is a complete “black- box”, which so might be useful for future attacks. Third, while our attack recovers only a (relatively small) part of the entire model, the fact that it is at all possible to steal any parameters of a production model is surprising, and raises concerns that extensions of this attack might be able to recover more information. Finally, recovering the model’s last layer (and thus hidden dimension) may reveal more global information about the model, such as relative size differences between different models. Our attack is effective and efficient, and is applicable to production models whose APIs expose full logprobs, or a “logit bias”. This included Google’s PaLM-2 and OpenAI’s GPT-4 (Anil et al., 2023; OpenAI et al., 2023); after respon- sible disclosure, both APIs have implemented defenses to prevent our attack or make it more expensive. We extract the embedding layer of several OpenAI models with a mean squared error of 10−4(up to unavoidable symmetries). We apply a limited form of our attack to gpt-3.5 at a cost of un- der $200 USD and, instead of recovering the full embedding layer, recover just the size of the embedding dimension. 1arXiv:2403.06634v1 [cs.CR] 11 Mar 2024
2306.02531.pdf	PLANNER: Generating Diversified Paragraph via Latent Language Diffusion Model Yizhe Zhang, Jiatao Gu, Zhuofeng Wu, Shuangfei Zhai, Josh Susskind, Navdeep Jaitly Apple Inc. {yizzhang, jgu32, zhuofeng_wu, szhai, jsusskind, njaitly}@apple.com Abstract Autoregressive models for text sometimes generate repetitive and low-quality output because errors accumulate during the steps of generation. This issue is often attributed to exposure bias – the difference between how a model is trained, and how it is used during inference. Denoising diffusion models provide an alternative approach in which a model can revisit and revise its output. However, they can be computationally expensive and prior efforts on text have led to models that produce less fluent output compared to autoregressive models, especially for longer text and paragraphs. In this paper, we propose PLANNER , a model that combines latent semantic diffusion with autoregressive generation, to generate fluent text while exercising global control over paragraphs. The model achieves this by combining an autoregressive “decoding” module with a “planning” module that uses latent diffusion to generate semantic paragraph embeddings in a coarse-to-fine manner. The proposed method is evaluated on various conditional generation tasks, and results on semantic generation, text completion and summarization show its effectiveness in generating high-quality long-form text in an efficient manner. 1 Introduction FT (Top-p sampling): Terrible, awful hotel. Horrible, awful hotel. I'm going to the awful hotel. I'll never get in. What the hell is this hotel? It's awful. I don't like it. Where? It's awful. I think they have a bathroom there. I'm going to the awful hotel. I'll never get in. I want out of here. This hotel is horrible. I'm going to the awful hotel. I'll never get in. I want out of here…FT (Greedy decoding): The hotel was a nightmare. The rooms were filthy, the bathrooms were filthy, the food was disgusting, the food was disgusting. The food was disgusting. The food was disgusting. The food was disgusting. The food was disgusting. The food was disgusting.…Prefix: Horrible hotel. Horrible hotel. Serioulsy, horrible hotel. Ours (Greedy decoding): Staff was unhelpful. So rude from the front desk. The hotel tried to lose the reservation of my family which I had with the hotel. This hotel is way overpriced. Hallways smelled bad. Carpet was disgusting. I'll never go back. I don't recommend the hotel. I can stay anywhere else!!1-gram2-gram3-gramThe: 100.00%The hotel: 100.00%The hotel was: 100.00%I: 13.39% Horrible: 12.20%The: 8.27% Terrible: 5.91% This: 4.33%Horrible hotel: 12.50% Terrible hotel: 6.05%This is: 2.82% I hate: 1.61%I mean: 1.61%Horrible hotel. Horrible: 3.48%Horrible hotel. I: 2.61%Terrible hotel. Terrible: 2.17%, Terrible hotel. Horrible: 1.30%Icould not: 0.87%I: 10.55%The: 5.08%We: 3.52%Hotel: 2.73% They: 2.73%Iwas: 1.56%Had to: 1.56%The room: 1.17%Dirty and: 1.17%Icalled: 1.17%The room was: 1.18%The walls were: 0.78%Iwas told: 0.78%Icalled the: 0.78%Ibooked through:0.78% Figure 1: Left: With a repetitive prompt, the finetuned GPT-2 large model (774M, FT) is still attracted to self-reinforced repetition (highlighted text) even under top-p sampling (K=50, p=0.92). Right: the most frequent first n-grams of the generations for each method, derived from 512 generation roll-outs. Our proposed method results in a more diversified generation robust to the ill-composed prompt. Autoregressive models trained with a teacher forcing strategy (Williams & Zipser, 1989) are con- sidered the gold standard for text generation. However, a significant drawback of this approach is that it lacks the ability to correct any mistakes made during the generation process which can lead to errors that accumulate as the generation progresses. Previous work (Ott et al., 2018; Holtzman et al., 2019; Welleck et al., 2019; Xu et al., 2022) has observed that deterministic decoding methods have a Preprint. Under review.arXiv:2306.02531v1 [cs.CL] 5 Jun 2023
1707.06347.pdf	Proximal Policy Optimization Algorithms John Schulman, Filip Wolski, Prafulla Dhariwal, Alec Radford, Oleg Klimov OpenAI {joschu, filip, prafulla, alec, oleg }@openai.com Abstract We propose a new family of policy gradient methods for reinforcement learning, which al- ternate between sampling data through interaction with the environment, and optimizing a “surrogate” objective function using stochastic gradient ascent. Whereas standard policy gra- dient methods perform one gradient update per data sample, we propose a novel objective function that enables multiple epochs of minibatch updates. The new methods, which we call proximal policy optimization (PPO), have some of the beneﬁts of trust region policy optimiza- tion (TRPO), but they are much simpler to implement, more general, and have better sample complexity (empirically). Our experiments test PPO on a collection of benchmark tasks, includ- ing simulated robotic locomotion and Atari game playing, and we show that PPO outperforms other online policy gradient methods, and overall strikes a favorable balance between sample complexity, simplicity, and wall-time. 1 Introduction In recent years, several diﬀerent approaches have been proposed for reinforcement learning with neural network function approximators. The leading contenders are deep Q-learning [Mni+15], “vanilla” policy gradient methods [Mni+16], and trust region / natural policy gradient methods [Sch+15b]. However, there is room for improvement in developing a method that is scalable (to large models and parallel implementations), data eﬃcient, and robust (i.e., successful on a variety of problems without hyperparameter tuning). Q-learning (with function approximation) fails on many simple problems1and is poorly understood, vanilla policy gradient methods have poor data eﬃency and robustness; and trust region policy optimization (TRPO) is relatively complicated, and is not compatible with architectures that include noise (such as dropout) or parameter sharing (between the policy and value function, or with auxiliary tasks). This paper seeks to improve the current state of aﬀairs by introducing an algorithm that attains the data eﬃciency and reliable performance of TRPO, while using only ﬁrst-order optimization. We propose a novel objective with clipped probability ratios, which forms a pessimistic estimate (i.e., lower bound) of the performance of the policy. To optimize policies, we alternate between sampling data from the policy and performing several epochs of optimization on the sampled data. Our experiments compare the performance of various diﬀerent versions of the surrogate objec- tive, and ﬁnd that the version with the clipped probability ratios performs best. We also compare PPO to several previous algorithms from the literature. On continuous control tasks, it performs better than the algorithms we compare against. On Atari, it performs signiﬁcantly better (in terms of sample complexity) than A2C and similarly to ACER though it is much simpler. 1While DQN works well on game environments like the Arcade Learning Environment [Bel+15] with discrete action spaces, it has not been demonstrated to perform well on continuous control benchmarks such as those in OpenAI Gym [Bro+16] and described by Duan et al. [Dua+16]. 1 arXiv:1707.06347v2 [cs.LG] 28 Aug 2017
22-1514.pdf	Journal of Machine Learning Research 24 (2023) 1-42 Submitted 12/22; Published 6/23 Convex Reinforcement Learning in Finite Trials Mirco Mutti [email protected] Politecnico di Milano Piazza Leonardo Da Vinci 32, 20133 Milan, Italy Riccardo De Santi∗[email protected] ETH Z¨ urich R¨ amistrasse 101, 8092 Z¨ urich, Switzerland Piersilvio De Bartolomeis∗[email protected] ETH Z¨ urich R¨ amistrasse 101, 8092 Z¨ urich, Switzerland Marcello Restelli [email protected] Politecnico di Milano Piazza Leonardo Da Vinci 32, 20133 Milan, Italy Editor: Kilian Weinberger Abstract Convex Reinforcement Learning (RL) is a recently introduced framework that generalizes the standard RL objective to any convex (or concave) function of the state distribution induced by the agent’s policy. This framework subsumes several applications of practical interest, such as pure exploration, imitation learning, and risk-averse RL, among others. However, the previous convex RL literature implicitly evaluates the agent’s performance over inﬁnite realizations (or trials), while most of the applications require excellent perfor- mance over a handful, or even just one, trials. To meet this practical demand, we formulate convex RL in ﬁnite trials, where the objective is any convex function of the empirical state distribution computed over a ﬁnite number of realizations. In this paper, we provide a comprehensive theoretical study of the setting, which includes an analysis of the impor- tance of non-Markovian policies to achieve optimality, as well as a characterization of the computational and statistical complexity of the problem in various conﬁgurations. Keywords: Reinforcement Learning, Convex Reinforcement Learning, General Utilities, Finite Trials, Non-Markovian Policies 1. Introduction Although Reinforcement Learning (RL, Sutton and Barto, 2018) provides a powerful and ﬂexible framework to model sequential decision-making problems, many relevant applica- tions do not ﬁt naturally into the standard RL framework (Abel et al., 2021). Especially, the objective function of RL can be seen as a linear combination between a reward vector and the state distribution induced by the agent’s policy. However, some applications cannot be cast into a linear objective function. Several works have thus extended the standard RL formulation to address non-linear objectives of practical interest. ∗Riccardo and Piersilvio contributed to the works (Mutti et al., 2022b,a) that are extended by this paper. c⃝2023 Mirco Mutti, Riccardo De Santi, Piersilvio De Bartolomeis, and Marcello Restelli. License: CC-BY 4.0, see https://creativecommons.org/licenses/by/4.0/ . Attribution requirements are provided athttp://jmlr.org/papers/v24/22-1514.html .
1606.08415.pdf	GAUSSIAN ERROR LINEAR UNITS (GELU S) Dan Hendrycks∗ University of California, Berkeley [email protected] Gimpel Toyota Technological Institute at Chicago [email protected] ABSTRACT We propose the Gaussian Error Linear Unit (GELU), a high-performing neural network activation function. The GELU activation function is xΦ(x), where Φ(x) the standard Gaussian cumulative distribution function. The GELU nonlinearity weights inputs by their value, rather than gates inputs by their sign as in ReLUs (x1x>0). We perform an empirical evaluation of the GELU nonlinearity against the ReLU and ELU activations and find performance improvements across all considered computer vision, natural language processing, and speech tasks. 1 I NTRODUCTION Early artificial neurons utilized binary threshold units (Hopfield, 1982; McCulloch & Pitts, 1943). These hard binary decisions are smoothed with sigmoid activations, enabling a neuron to have a “fir- ing rate” interpretation and to train with backpropagation. But as networks became deeper, training with sigmoid activations proved less effective than the non-smooth, less-probabilistic ReLU (Nair & Hinton, 2010) which makes hard gating decisions based upon an input’s sign. Despite having less of a statistical motivation, the ReLU remains a competitive engineering solution which often enables faster and better convergence than sigmoids. Building on the successes of ReLUs, a recent modifi- cation called ELUs (Clevert et al., 2016) allows a ReLU-like nonlinearity to output negative values which sometimes increases training speed. In all, the activation choice has remained a necessary architecture decision for neural networks lest the network be a deep linear classifier. Deep nonlinear classifiers can fit their data so well that network designers are often faced with the choice of including stochastic regularizer like adding noise to hidden layers or applying dropout (Sri- vastava et al., 2014), and this choice remains separate from the activation function. Some stochastic regularizers can make the network behave like an ensemble of networks, a pseudoensemble (Bach- man et al., 2014), and can lead to marked accuracy increases. For example, the stochastic regular- izer dropout creates a pseudoensemble by randomly altering some activation decisions through zero multiplication. Nonlinearities and dropout thus determine a neuron’s output together, yet the two innovations have remained distinct. More, neither subsumed the other because popular stochastic regularizers act irrespectively of the input and nonlinearities are aided by such regularizers. In this work, we introduce a new nonlinearity, the Gaussian Error Linear Unit (GELU). It relates to stochastic regularizers in that it is the expectation of a modification to Adaptive Dropout (Ba & Frey, 2013). This suggests a more probabilistic view of a neuron’s output. We find that this novel nonlinearity matches or exceeds models with ReLUs or ELUs across tasks from computer vision, natural language processing, and automatic speech recognition. 2 GELU F ORMULATION We motivate our activation function by combining properties from dropout, zoneout, and ReLUs. First note that a ReLU and dropout both yield a neuron’s output with the ReLU deterministi- cally multiplying the input by zero or one and dropout stochastically multiplying by zero. Also, a new RNN regularizer called zoneout stochastically multiplies inputs by one (Krueger et al., 2016). We merge this functionality by multiplying the input by zero or one, but the values of this zero-one mask are stochastically determined while also dependent upon the input. Specif- ically, we can multiply the neuron input xbym∼Bernoulli (Φ(x)), where Φ(x) = P(X≤ ∗Work done while the author was at TTIC. Code available at github.com/hendrycks/GELUs 1arXiv:1606.08415v5 [cs.LG] 6 Jun 2023
2110.07205.pdf	SpeechT5: Uniﬁed-Modal Encoder-Decoder Pre-Training for Spoken Language Processing Junyi Ao1,2,∗, Rui Wang3,∗, Long Zhou4,∗, Chengyi Wang4, Shuo Ren4, Yu Wu4, Shujie Liu4, Tom Ko1, Qing Li2, Yu Zhang1,5, Zhihua Wei3, Yao Qian4, Jinyu Li4, Furu Wei4 1Department of Computer Science and Engineering, Southern University of Science and Technology 2Department of Computing, The Hong Kong Polytechnic University 3Department of Computer Science and Technology, Tongji University 4Microsoft5Peng Cheng Laboratory Abstract Motivated by the success of T5 (Text-To- Text Transfer Transformer) in pre-trained nat- ural language processing models, we pro- pose a uniﬁed-modal SpeechT5 framework that explores the encoder-decoder pre-training for self-supervised speech/text representation learning. The SpeechT5 framework consists of a shared encoder-decoder network and six modal-speciﬁc (speech/text) pre/post-nets. Af- ter preprocessing the input speech/text through the pre-nets, the shared encoder-decoder net- work models the sequence-to-sequence trans- formation, and then the post-nets generate the output in the speech/text modality based on the output of the decoder. Leveraging large-scale unlabeled speech and text data, we pre-train SpeechT5 to learn a uniﬁed-modal representa- tion, hoping to improve the modeling capabil- ity for both speech and text. To align the tex- tual and speech information into this uniﬁed semantic space, we propose a cross-modal vec- tor quantization approach that randomly mixes up speech/text states with latent units as the interface between encoder and decoder. Ex- tensive evaluations show the superiority of the proposed SpeechT5 framework on a wide va- riety of spoken language processing tasks, in- cluding automatic speech recognition, speech synthesis, speech translation, voice conver- sion, speech enhancement, and speaker iden- tiﬁcation. We release our code and model athttps://github.com/microsoft/ SpeechT5 . 1 Introduction Starting with ELMo (Peters et al., 2018) and BERT (Devlin et al., 2019), substantial work has shown that pre-trained models can signiﬁcantly improve in various natural language processing (NLP) tasks ∗Equal contribution. Work is done by the ﬁrst two authors during internship at Microsoft Research Asia. Correspondence to: Long Zhou ([email protected]) Figure 1: An illustration of the SpeechT5 frame- work, which treats spoken language processing tasks as a speech/text to speech/text format, including au- tomatic speech recognition (ASR), speech translation (ST), speech identiﬁcation (SID), text to speech (TTS), voice conversion (VC), and speech enhancement (SE). (Radford et al., 2019; CONNEAU and Lample, 2019; Yang et al., 2019; Dong et al., 2019; Lewis et al., 2020). Following the pre-training techniques in NLP, self-supervised speech representation learn- ing has also been investigated and shown promising results, beneﬁting from richly learned representa- tions (Chung and Glass, 2018; Chuang et al., 2020; Song et al., 2019; Baevski et al., 2020; Wang et al., 2021; Hsu et al., 2021; Chung et al., 2021a), such as wav2vec 2.0 (Baevski et al., 2020) and HuBERT (Hsu et al., 2021). However, previous speech pre-training work suf- fers from two problems: (1) most of them learn the speech representation with only unlabeled speech data but ignore the importance of textual data to spoken language tasks (e.g., automatic speech recognition) which require the modality transfor- mation; (2) most of these models solely rely on a pre-trained speech encoder for various downstream tasks, leaving the decoder not pre-trained for the sequence-to-sequence generation tasks. How to de- sign a uniﬁed encoder-decoder model that can take advantage of both unlabeled speech and text data to improve various spoken language processing tasks is not well explored. Inspired by the T5 method (Raffel et al., 2020),arXiv:2110.07205v3 [eess.AS] 24 May 2022
image-decoding-paper.pdf	BRAIN DECODING :TOWARD REAL -TIME RECONSTRUCTION OF VISUAL PERCEPTION Yohann Benchetrit1,∗, Hubert Banville1,∗, Jean-R ´emi King1,2 1FAIR, Meta,2Laboratoire des Syst `emes Perceptifs, ´Ecole Normale Sup ´erieure, PSL University {ybenchetrit,hubertjb,jeanremi }@meta.com ABSTRACT In the past five years, the use of generative and foundational AI systems has greatly improved the decoding of brain activity. Visual perception, in particular, can now be decoded from functional Magnetic Resonance Imaging (fMRI) with remarkable fidelity. This neuroimaging technique, however, suffers from a lim- ited temporal resolution ( ≈0.5 Hz) and thus fundamentally constrains its real-time usage. Here, we propose an alternative approach based on magnetoencephalog- raphy (MEG), a neuroimaging device capable of measuring brain activity with high temporal resolution ( ≈5,000 Hz). For this, we develop an MEG decoding model trained with both contrastive and regression objectives and consisting of three modules: i) pretrained embeddings obtained from the image, ii) an MEG module trained end-to-end and iii) a pretrained image generator. Our results are threefold: Firstly, our MEG decoder shows a 7X improvement of image-retrieval over classic linear decoders. Second, late brain responses to images are best de- coded with DINOv2, a recent foundational image model. Third, image retrievals and generations both suggest that MEG signals primarily contain high-level visual features, whereas the same approach applied to 7T fMRI also recovers low-level features. Overall, these results provide an important step towards the decoding – in real time – of the visual processes continuously unfolding within the human brain. 1 I NTRODUCTION Automating the discovery of brain representations. Understanding how the human brain rep- resents the world is arguably one of the most profound scientific challenges. This quest, which originally consisted of searching, one by one, for the specific features that trigger each neuron, ( e.g. Hubel & Wiesel (1962); O’Keefe & Nadel (1979); Kanwisher et al. (1997)), is now being automated by Machine Learning (ML) in two mains ways. First, as a signal processing tool, ML algorithms are trained to extract informative patterns of brain activity in a data-driven manner. For example, Kami- tani & Tong (2005) trained a support vector machine to classify the orientations of visual gratings from functional Magnetic Resonance Imaging (fMRI). Since then, deep learning has been increas- ingly used to discover such brain activity patterns (Roy et al., 2019; Thomas et al., 2022; Jayaram & Barachant, 2018; D ´efossez et al., 2022; Scotti et al., 2023). Second, ML algorithms are used as functional models of the brain. For example, Yamins et al. (2014) have shown that the embedding of natural images in pretrained deep nets linearly account for the neuronal responses to these images in the cortex. Since, pretrained deep learning models have been shown to account for a wide variety of stimuli including text, speech, navigation, and motor movement (Banino et al., 2018; Schrimpf et al., 2020; Hausmann et al., 2021; Mehrer et al., 2021; Caucheteux et al., 2023). Generating images from brain activity. This observed representational alignment between brain activity and deep learning models creates a new opportunity: Decoding of visual stimuli need not be restricted to a limited set of classes, but can now leverage pretrained representations to condi- tion subsequent generative AI models. While the resulting image may be partly “hallucinated”, interpreting images can be much simpler than interpreting latent features. Following a long series ∗Equal contribution. 1
2402.13064.pdf	Synthetic Data (Almost) from Scratch: Generalized Instruction Tuning for Language Models Haoran Li∗, Qingxiu Dong∗, Zhengyang Tang∗, Chaojun Wang∗, Xingxing Zhang∗, Haoyang Huang∗ Shaohan Huang, Xiaolong Huang, Zeqiang Huang, Dongdong Zhang, Yuxian Gu, Xin Cheng Xun Wang, Si-Qing Chen, Li Dong, Wei Lu, Zhifang Sui, Benyou Wang, Wai Lam, Furu Wei https://aka.ms/GeneralAI Abstract We introduce Generalized Instruction Tuning (called GLAN ), a general and scal- able method for instruction tuning of Large Language Models (LLMs). Unlike prior work that relies on seed examples or existing datasets to construct instruction tuning data, GLAN exclusively utilizes a pre-curated taxonomy of human knowledge and capabilities as input and generates large-scale synthetic instruction data across all disciplines. Specifically, inspired by the systematic structure in human education system, we build the taxonomy by decomposing human knowledge and capabilities to various fields, sub-fields and ultimately, distinct disciplines semi-automatically, facilitated by LLMs. Subsequently, we generate a comprehensive list of subjects for every discipline and proceed to design a syllabus tailored to each subject, again utilizing LLMs. With the fine-grained key concepts detailed in every class session of the syllabus, we are able to generate diverse instructions with a broad coverage across the entire spectrum of human knowledge and skills. Extensive experiments on large language models (e.g., Mistral) demonstrate that GLAN excels in mul- tiple dimensions from mathematical reasoning, coding, academic exams, logical reasoning to general instruction following without using task-specific training data of these tasks. In addition, GLAN allows for easy customization and new fields or skills can be added by simply incorporating a new node into our taxonomy. 1 Introduction Large Language Models (LLMs) have enabled unprecedented capabilities to understand and generate text like humans. By scaling up model size and data size [ KMH+20,HBM+22], LLMs are better at predicting next tokens and prompting to perform certain tasks with a few demonstrations [ BMR+20]. However, these capabilities do not directly translate to better human instruction following [ OWJ+22]. Instruction tuning [ WBZ+21] bridges this gap through fine-tuning LLMs on instructions paired with human-preferred responses. Prior work constructs instruction tuning data from seed examples or existing datasets. Initially, natural language processing (NLP) datasets described via instructions are used to fine-tune LLMs and the resulting LLMs can generalize on unseen (NLP) tasks [ WBZ+21]. However, there are only thousands of NLP tasks [ WMA+22,LHV+23] available, which limits the tuned LLMs to generalize in real- world scenarios [ XSZ+23]. Self-instruct [ WKM+22] is a cost effective method for creating synthetic instruction tuning datasets, which starts from a small pool of human-written seed instructions and ∗Equal contribution. X. Zhang ([email protected]), H. Huang, S. Huang, X. Huang, Z. Huang, D. Zhang, X. Wang, S. Chen, L. Dong and F. Wei are with Microsoft. H. Li and W. Lu are with Singapore University of Technology and Design. Q. Dong, X. Cheng and Z. Sui are with Peking University. Z. Tang and B. Wang are with Chinese University of Hong Kong, Shenzhen. C. Wang and W. Lam are with Chinese University of Hong Kong. Y . Gu is with Tsinghua University.arXiv:2402.13064v1 [cs.CL] 20 Feb 2024
2402.17764v1.pdf	The Era of 1-bit LLMs: All Large Language Models are in 1.58 Bits Shuming Ma∗Hongyu Wang∗Lingxiao Ma Lei Wang Wenhui Wang Shaohan Huang Li Dong Ruiping Wang Jilong Xue Furu Wei⋄ https://aka.ms/GeneralAI Abstract Recent research, such as BitNet [ WMD+23], is paving the way for a new era of 1- bit Large Language Models (LLMs). In this work, we introduce a 1-bit LLM variant, namely BitNet b1.58 , in which every single parameter (or weight) of the LLM is ternary {-1, 0, 1}. It matches the full-precision (i.e., FP16 or BF16) Transformer LLM with the same model size and training tokens in terms of both perplexity and end-task performance, while being significantly more cost-effective in terms of latency, memory, throughput, and energy consumption. More profoundly, the 1.58-bit LLM defines a new scaling law and recipe for training new generations of LLMs that are both high-performance and cost-effective. Furthermore, it enables a new computation paradigm and opens the door for designing specific hardware optimized for 1-bit LLMs. 0.2961 -0.0495 0.0413 ...… -0.4765 0.2812 0.2403 -0.1808 0.1304 -0.4809 …… -0.1771 -0.1741 -0.3853Transformer LLMs 16-bit Float ( FP16/BF16) Cost Performance 1 -1 0…… 1 -1 -1 -1 1 -1…… 0 0 -1BitNet b1.58 (This Work) {-1, 0, 1} W=Pareto Improvement W= 0.0413 0.3397 0.2812 0.2403 -0.1808 0.1304 -0.4809 0.32440.4322 -0.1771 -0.1741 -0.38530.2961 -0.0495 -0.0924 -0.4765 𝒙𝟎 𝒙𝟏 𝒙𝟐 𝒙𝟑𝟎.𝟐𝟗𝟔𝟏𝒙𝟎−𝟎.𝟎𝟒𝟗𝟓𝒙𝟏−𝟎.𝟎𝟗𝟐𝟒𝒙𝟐−𝟎.𝟒𝟕𝟔𝟓𝒙𝟑 … 1 -1 0 1-1 1 -1 -1 -1 0 -1 11 -1 1 0𝒙𝟎−𝒙𝟏−𝒙𝟐+𝒙𝟑 …𝒙𝟎 𝒙𝟏 𝒙𝟐 𝒙𝟑1(.58) -bitFP16Model W Input X Output Y Y = f(W, X) GPU New Hardware Figure 1: 1-bit LLMs (e.g., BitNet b1.58 ) provide a Pareto solution to reduce inference cost (latency, throughput, and energy) of LLMs while maintaining model performance. The new computation paradigm of BitNet b1.58 calls for actions to design new hardware optimized for 1-bit LLMs. ∗Equal contribution. ⋄Corresponding author. S. Ma, L. Ma, L. Wang, W. Wang, S. Huang, L. Dong, J. Xue, F. Wei are with Microsoft Research. H. Wang and R. Wang are with University of Chinese Academy of Sciences.arXiv:2402.17764v1 [cs.CL] 27 Feb 2024
6593-contrastive-preference-learnin.pdf	Under review as a conference paper at ICLR 2024 CONTRASTIVE PREFERENCE LEARNING : LEARNING FROM HUMAN FEEDBACK WITHOUT RL Anonymous authors Paper under double-blind review ABSTRACT Reinforcement Learning from Human Feedback (RLHF) has emerged as a popular paradigm for aligning models with human intent. Typically RLHF algorithms operate in two phases: first, use human preferences to learn a reward function and second, align the model by optimizing the learned reward via reinforcement learn- ing (RL). This paradigm assumes that human preferences are distributed according to reward, but recent work suggests that they instead follow the regret under the user’s optimal policy. Thus, learning a reward function from feedback is not only based on a flawed assumption of human preference, but also leads to unwieldy optimization challenges that stem from policy gradients or bootstrapping in the RL phase. Because of these optimization challenges, contemporary RLHF methods restrict themselves to contextual bandit settings (e.g., as in large language models) or limit observation dimensionality (e.g., state-based robotics). We overcome these limitations by introducing a new family of algorithms for optimizing behavior from human feedback using the regret -based model of human preferences. Using the principle of maximum entropy, we derive Contrastive Preference Learning (CPL), an algorithm for learning optimal policies from preferences without learning re- ward functions, circumventing the need for RL. CPL is fully off-policy, uses only a simple contrastive objective, and can be applied to arbitrary MDPs. In contrast to prior work, this enables CPL to elegantly scale to high-dimensional and sequential RLHF problems. 1 I NTRODUCTION As large pretrained models have become increasingly performant, the problem of aligning them with human preferences have risen to the forefront of research. This alignment is especially difficult when larger datasets inevitably include suboptimal behaviors. Reinforcement learning from human feedback (RLHF) has emerged as a popular solution to this problem. Using human preferences, RLHF techniques discriminate between desirable and undesirable behaviors with the goal of refining a learned policy. This paradigm has shown promising results when applied to finetuning large language models (LLMs) (Ouyang et al., 2022), improving image generation models (Lee et al., 2023), and adapting robot policies (Christiano et al., 2017) – all from suboptimal data. For most RLHF algorithms, this process includes two phases. First, a reward model is trained from collected user preference data. And second, that reward model is optimized by an off-the-shelf reinforcement learning (RL) algorithm. Unfortunately, this two-phase paradigm is founded on a flawed assumption. Algorithms that learn reward models from preference data require that human preferences are distributed according to the discounted sum of rewards or partial return of each behavior segment. However, recent work (Knox et al., 2022) calls this into question, positing that humans instead provide preferences based on the regret of each behavior under the optimal policy of the expert’s reward function. Intuitively, a human’s judgement is likely based on optimality, instead of which states and actions have higher quantity for reward. As a result, the correct quantity to learn from feedback might not be the reward, but instead the optimal advantage function or, in other words, the negated regret. In their second phase, two-phase RLHF algorithms optimize the reward function learned from the first phase with RL. In practice, RL algorithms suffer from a suite of optimization challenges stemming from temporal credit assignment, such as the high-variance of policy gradients (Marbach & Tsitsiklis, 1
10.1101.2022.12.21.521521.pdf	Language models generalize beyond natural proteins Robert Verkuil1 Ori Kabeli1 Yilun Du1 2Basile I. M. Wicky3 4Lukas F. Milles3 4Justas Dauparas3 4 David Baker3 4 5Sergey Ovchinnikov6Tom Sercu1Alexander Rives1 7 † Abstract Learning the design patterns of proteins from sequences across evolution may have promise toward generative protein design. However it is unknown whether lan- guage models, trained on sequences of natural proteins, will be capable of more than memorization of existing protein families. Here we show that language models generalize beyond natural proteins to generate de novo proteins. We focus on two protein design tasks: ﬁxed backbone design where the structure is speciﬁed, and unconstrained generation where the structure is sampled from the model. Remarkably although the models are trained only on sequences, we ﬁnd that they are capable of designing structure. A total of 228 generated pro- teins are evaluated experimentally with high overall suc- cess rates (152/228 or 67%) in producing a soluble and monomeric species by size exclusion chromatography. Out of 152 experimentally successful designs, 35 have no signiﬁcant sequence match to known natural proteins. Of the remaining 117, sequence identity to the nearest sequence match is at median 27%, below 20% for 6 designs, and as low as 18% for 3 designs. For ﬁxed back- bone design, the language model generates successful designs for each of eight experimentally evaluated artiﬁ- cially created ﬁxed backbone targets. For unconstrained generation, sampled proteins cover diverse topologies and secondary structure compositions, and have high experimental success rate (71/129 or 55%). The designs reﬂect deep patterns linking sequence and structure, in- cluding motifs that occur in related natural structures, and motifs that are not observed in similar structural contexts in known protein families. The results show that language models, though only trained on sequences, learn a deep grammar that enables the design of protein structure, extending beyond natural proteins. *Equal contribution1Meta Fundamental AI Research Protein Team (FAIR).2Massachusetts Institute of Technology. Work performed as visiting researcher at Meta FAIR.3Department of Biochemistry, Uni- versity of Washington, Seattle, WA, USA.4Institute for Protein De- sign, University of Washington, Seattle, WA, USA.5Howard Hughes Medical Institute, University of Washington, Seattle, WA, USA.6John Harvard Distinguished Science Fellowship Program, Harvard Univer- sity, Cambridge, MA, USA.7New York University. Data available at:<https://github.com/facebookresearch/esm >.†Correspondence to <[email protected]>. Preprint. Copyright 2022 by the authors.Introduction Generative artiﬁcial intelligence for biology has potential to open up a space of protein design beyond natural proteins. Since amino acid sequences are the fundamental codes of proteins, learning to read and write these codes with a language model may have promise. Language models have played a central role in recent advances in artiﬁcial intelligence ( 1), including developments in complex reasoning, mathematical problem solving, image gener- ation, and natural language generation ( 2–4). Scaling laws link performance with the compute, data, and number of parameters used to train the models ( 5), and emergence of higher level capabil- ities is observed with increasing scale ( 6). In biology, recent work on evolutionary scale language models of proteins has shown that a deep knowledge of intrinsic biological properties emerges from training on protein sequences ( 7). Information about the folded three dimensional structure of proteins develops within the mod- els, extending to atomic resolution structure ( 8). This information emerges through training on sequences alone. At the same time the structural information that emerges as a result of training on sequences has been shown to depend on the available evolutionary information, varying as a function of the number of related proteins in the training data ( 8,9). It is an open question across domains to what extent language models are capable of generalizing outside their training data. In biology, it is unknown whether language models can be used to explore a design space beyond that of natural proteins. Here we demonstrate that language models generalize beyond nat- ural proteins to generate de novo proteins, different in sequence and structure from natural proteins. We experimentally validate a large number of designs spanning diverse topologies and sequences. We ﬁnd that although language models are trained only on the se- quences of proteins, they are capable of designing protein structure, including structures of artiﬁcially engineered de novo proteins that are distinct from those of natural proteins. Given the backbone of ade novo protein structure as a target, the language model gener- ates sequences that are predicted to fold to the speciﬁed structure. When the sequence and structure are both free, language models produce designs that span a wide range of fold topologies and secondary structure compositions, creating proteins which overlap the natural sequence distribution as well as extend beyond it. De- signs succeed experimentally across the space of sampled proteins, including many designs that are distant in sequence from natural proteins. The model generates motifs that link sequence to the design of structure and can apply them in new sequence and struc- tural contexts, including motifs such as complex hydrogen bond. CC-BY-NC-ND 4.0 International license available under a(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made The copyright holder for this preprint this version posted December 22, 2022. ; https://doi.org/10.1101/2022.12.21.521521doi: bioRxiv preprint
2404.16767v1.pdf	REBEL : Reinforcement Learning via Regressing Relative Rewards Zhaolin Gao♣, Jonathan D. Chang♣, Wenhao Zhan♦, Owen Oertell♣, Gokul Swamyr, Kianté Brantley♣, Thorsten Joachims♣, J. Andrew Bagnellr, Jason D. Lee♦, Wen Sun♣ ♣Cornell University∗ ♦Princeton University†rCarnegie Mellon University‡ Abstract Whileoriginallydevelopedforcontinuouscontrolproblems,ProximalPolicyOpti- mization(PPO)hasemergedasthework-horseofavarietyofreinforcementlearning (RL) applications including the fine-tuning of generative models. Unfortunately, PPO requires multiple heuristics to enable stable convergence (e.g. value networks, clipping)andisnotoriousforitssensitivitytothepreciseimplementationofthese components. Inresponse,wetakeastepbackandaskwhata minimalist RLalgorithm forthe eraof generativemodelswould looklike. Wepropose REBEL,an algorithm that cleanly reduces the problem of policy optimization to regressing the relative rewards via a direct policy parameterization between two completions to a prompt, enablingstrikinglylightweightimplementation. Intheory,weprovethatfundamental RL algorithms like Natural Policy Gradient can be seen as variants of REBEL, which allowsustomatchthestrongestknowntheoreticalguaranteesintermsofconvergence and sample complexity in the RL literature. REBELcan also cleanly incorporate offline data and handle the intransitive preferences we frequently see in practice. Empirically,wefindthat REBELprovidesaunifiedapproachtolanguagemodeling and image generation with stronger or similar performance as PPO and DPO, all while being simpler to implement and more computationally tractable than PPO. 1 Introduction The generality of the reinforcement learning (RL) paradigm is striking: from continuous control problems (Kalashnikov et al., 2018) to, recently, the fine-tuning of generative models (Stiennon et al., 2022;Ouyangetal.,2022),RLhasenabledconcreteprogressacrossavarietyofdecision-making tasks. Specifically,whenitcomestofine-tuninggenerativemodels,ProximalPolicyOptimization (PPO,Schulmanetal.(2017))hasemergedasthede-factoRLalgorithmofchoice,fromlanguage models(LLMs)(Ziegleretal.,2020;Stiennonetal.,2022;Ouyangetal.,2022;Touvronetal., 2023) to image generative models (Black et al., 2023; Fan et al., 2024; Oertell et al., 2024). If we take a step back however, it is odd that we are using an algorithm designed for optimizing two-layernetworksforcontinuouscontroltasksfromscratchforfine-tuningthebillionsofparameters ∗{zg292, jdc396, ojo2, kdb82, ws455}@cornell.edu ,[email protected] †{wenhao.zhan, jasonlee}@princeton.edu ‡{gswamy,bagnell2}@andrew.cmu.edu 1arXiv:2404.16767v1 [cs.LG] 25 Apr 2024
8781-turing-complete-transformers-t.pdf	Under review as a conference paper at ICLR 2023 TURING COMPLETE TRANSFORMERS : T WOTRANS - FORMERS AREMORE POWERFUL THAN ONE Anonymous authors Paper under double-blind review ABSTRACT This paper presents Find+Replace transformers, a family of multi-transformer architectures that can provably do things no single transformer can, and which outperform GPT-4 on several challenging tasks. We first establish that tra- ditional transformers and similar architectures are not Turing complete, while Find+Replace transformers are. Using this fact, we show how arbitrary programs can be compiled to Find+Replace transformers, aiding interpretability research. We also demonstrate the superior performance of Find+Replace transformers over GPT-4 on a set of composition challenge problems (solving problems 20x longer on tower of Hanoi, 3%→100% on multiplication, 72%→100% on a dynamic programming problem). This work aims to provide a theoretical basis for multi- agent architectures, and to encourage their exploration. 1 I NTRODUCTION The first computers – including the difference engine , differential analyzer, Z1, and ABC (Bab- bage & Babbage, 1825; Bush, 1931; Rojas, 2014; Atanasoff, 1940) – were not Turing Complete. Some such machines, like the Hollerith Tabulating Machine (Hollerith, 1889) and the Harvard Mark I (Comrie, 1946), even achieved considerable real-world use despite that limitation. However, the advent of Turing Complete computers (Turing et al., 1936; Goldstine & Goldstine, 1946; Kilburn, 1949) fundamentally changed how computers were used and led to the development of more com- plex, comprehensible, and composable programs (Backus, 1954; Copeland, 2004). As we will show in this paper, current LLMs based on the transformer architecture (Vaswani et al., 2017) are not Turing Complete. We present an alternative that is. The fundamental reason transformers are not Turing complete is that, once the architecture of a transformer is decided, there is a bounded amount of computation that it can do. This guarantees the model will fail to generalize beyond input of some length and complexity. Such limitations are not only theoretical, they are supported by a number of recent results on the ability of language models to generalize to large context lengths (Del’etang et al., 2022; Liu et al., 2023; Dziri et al., 2023). Addressing these deficiencies is nontrivial and requires a fundamental shift in approach. We propose an approach drawing from multi-agent systems (Messing, 2003; Stone & Veloso, 2000), particularly multi-transformer systems. Such systems have recently garnered interest, being employed to gener- ate simulacra of human behavior (Park et al., 2023), perform economic simulations (Horton, 2023), and demonstrate open-ended exploration in games like Minecraft (Wang et al., 2023a). This paper presents a family of multi-transformer architectures, and provides theoretical and em- pirical evidence the family can outperform traditional transformers. We hope this study will ignite further investigations into architectures that are multi-transformer and Turing complete. Our contributions are as follows: • We provide a simple proof that current LLMs are not Turing Complete • We present Find+Replace transformers, a family of provably Turing Complete architectures • We introduce a method for turning any program into a Find+Replace transformer • We show that Find+Replace transformers out-perform GPT-4 on a set of challenge tasks 1
2404.09173.pdf	TransformerFAM: Feedback attention is working memory Dongseong Hwang1Weiran Wang1Zhuoyuan Huo1Khe Chai Sim1Pedro Moreno Mengibar1 Abstract While Transformers have revolutionized deep learning, their quadratic attention complexity hin- ders their ability to process infinitely long inputs. We propose Feedback Attention Memory (FAM), a novel Transformer architecture that leverages a feedback loop to enable the network to at- tend to its own latent representations. This de- sign fosters the emergence of working memory within the Transformer, allowing it to process indefinitely long sequences. TransformerFAM requires no additional weights, enabling seam- less integration with pre-trained models. Our experiments show that TransformerFAM signif- icantly improves Transformer performance on long-context tasks across various model sizes (1B, 8B, and 24B). These results showcase the poten- tial to empower Large Language Models (LLMs) to process sequences of unlimited length. 1. Introduction The introduction of the Transformer architecture (Vaswani et al., 2017) has revolutionized deep learning by permeat- ing diverse domains and enhancing performance due to its efficacy and scalability. This scalability fuels a trend analo- gous to Moore’s law, which links increased model size to performance gains (Kaplan et al., 2020). The effectiveness of attention in text sequence processing was solidified through the Transformer paper. Models like BERT (Devlin et al., 2018) and GPT-3 (Brown et al., 2020) further showcased the scalability of Transformer and its tendency for improved performance with increased model size. Following the replacement of LSTM (Hochreiter & Schmidhuber, 1997) by Transformer in most Natural Lan- guage Processing (NLP) domains, the Vision Transformer (ViT) (Dosovitskiy et al., 2020) replaced Convolutional Neural Network (CNN) (LeCun et al., 1995) with Trans- formers in the vision domain, and Conformer (Convolution- augmented Transformer) (Gulati et al., 2020) replaced 1Google LLC, Mountain View, CA, USA. Correspondence to: Dongseong Hwang <[email protected] >.LSTM in the speech domain. The Transformer has become the de facto architecture in various domains. Currently, attention serves as the leading architecture for extracting meaningful representations from homogeneous data. The logical progression points toward extending attention ca- pabilities to heterogeneous data. This has enabled advances in multimodal fusion (text and vision), as seen in models like DALL ·E 2 (Ramesh et al., 2022), Flamingo (Alayrac et al., 2022) and CoCa (Yu et al., 2022). AudioLM (Bor- sos et al., 2023) has shown that attention also excels at fusing audio and text. Consequently, Gemini (Team et al., 2023) integrates text, images, audio, and video into a single generative model. This was possible because attention to heterogeneous data works exceptionally well. Despite the impressive success of attention, it suffers from major drawbacks. Firstly, attention has quadratic complexity with respect to context length, which limits the capability of modeling long contexts. Secondly, it forgets information from context before attention window, unlike LSTM, which theoretically can propagate information indefinitely. We want the better architecture to be able to process arbitrarily long sequences efficiently, while preserving very long-term dependencies. Sliding window attention is introduced (Dai et al., 2019; Beltagy et al., 2020) to handle infinitely long sequences as input. However, it disregards information beyond the effective receptive field (approximately model depth ×win- dow size in Figure 1d) during sequence generation. Vari- ous approaches have attempted to address the long-context problem by sparse attention (Child et al., 2019; Beltagy et al., 2020; Zaheer et al., 2020; Kitaev et al., 2020; Roy et al., 2021; Oren et al., 2024) and linear approximated attention (Wang et al., 2020; Katharopoulos et al., 2020; Choromanski et al., 2020; Xiong et al., 2021), showing ef- fectiveness below the 1B scale. Yet, scaling laws (Tay et al., 2022) suggest these approximations do not perform as well at GPT-3 level. Current SoTA large language model (LLM) leaderboards (LMSYS, 2023) do not feature architectures primarily relying on approximated attention. On the other hand, Neuroscience links attention to multi- sensory integration (Tang et al., 2016). Endogenous (goal- driven) and exogenous (stimulus-driven) attention are dis- tributed throughout sensory processing regions in brain, and 1arXiv:2404.09173v1 [cs.LG] 14 Apr 2024
1002.1945v2.pdf	arXiv:1002.1945v2 [math.GR] 14 May 2010HYDRA GROUPS W.DISONAND T.R.RILEY Abstract. Wegive examples of CAT(0), biautomatic, free–by–cyclic, one–relator groups which have ﬁnite–rank free subgroups of huge (Ackermannian ) distortion. This leads to elementary examples of groups whose Dehn functions are simi larly extravagant. This behaviour originates in manifestations of Hercules–versus–the–hydra battles in string– rewriting. 2010 Mathematics Subject Classiﬁcation: 20F65, 20F10,20F 67 Keywords and phrases: free–by–cyclic group, subgroup distortion, Dehn function , hydra 1. Introduction 1.1.Hercules versus the hydra. Hercules’ second labour was to ﬁght the Lernaean hy- dra, a beast with multiple serpentine heads enjoying magica l regenerative powers: when- ever a head was severed, two grew in its place. Hercules succe eded with the help of his nephew,Iolaus,whostoppedtheregrowthbysearingthestum pswithaburningtorchafter each decapitation. The extraordinarily fast–growing func tions we will encounter in this articlestem fromare-imaginingofthisbattle. For us, a hydrawill be a ﬁnite–length positiveword on the alphabet a1,a2,a3,...— that is, it includesnoinverseletters a1−1,a2−1,a3−1,.... Herculesﬁghtsa hydrabystrikingo ﬀ itsﬁrstletter. Thehydrathenregeneratesasfollows: each remainingletter ai,wherei>1, becomes aiai−1and thea1are unchanged. This process — removal of the ﬁrst letter and thenregeneration— repeats, with Herculesvictorious when(notif!) the hydrais reduced totheemptyword ε. Forexample,Herculesdefeatsthehydra a2a3a1in ﬁvestrikes: a2a3a1→a3a2a1→a2a1a1→a1a1→a1→ε. (Eacharrowrepresentstheremovaloftheﬁrst letterandthe nregeneration.) Proposition1.1. Herculesdefeatsall hydra. Proof.When ﬁghting a hydra in which the highest index present is k, noaiwithi>k will ever appear, and nor will any new ak. The preﬁx before the ﬁrst akis itself a hydra, which, by induction, we can assume Hercules defeats. Hercul es will then removethat ak, decreasingthetotalnumberof akpresent. It followsthat Herculeseventuallywins. □ However these battles are of extreme duration. Deﬁne H(w) to be the number of strikes it takes Hercules to vanquish the hydra w, and for integers k≥1,n≥0, deﬁneHk(n) := Date: October 23, 2018. 1
2303.07678.pdf	Query2doc: Query Expansion with Large Language Models Liang Wang and Nan Yang and Furu Wei Microsoft Research {wangliang,nanya,fuwei}@microsoft.com Abstract This paper introduces a simple yet effec- tive query expansion approach, denoted as query2doc , to improve both sparse and dense re- trieval systems. The proposed method first gen- erates pseudo-documents by few-shot prompt- ing large language models (LLMs), and then expands the query with generated pseudo- documents. LLMs are trained on web-scale text corpora and are adept at knowledge memoriza- tion. The pseudo-documents from LLMs often contain highly relevant information that can aid in query disambiguation and guide the re- trievers. Experimental results demonstrate that query2doc boosts the performance of BM25 by 3% to 15% on ad-hoc IR datasets, such as MS- MARCO and TREC DL, without any model fine-tuning. Furthermore, our method also ben- efits state-of-the-art dense retrievers in terms of both in-domain and out-of-domain results. 1 Introduction Information retrieval (IR) aims to locate relevant documents from a large corpus given a user is- sued query. It is a core component in modern search engines and researchers have invested for decades in this field. There are two mainstream paradigms for IR: lexical-based sparse retrieval, such as BM25, and embedding-based dense re- trieval (Xiong et al., 2021; Qu et al., 2021). Al- though dense retrievers perform better when large amounts of labeled data are available (Karpukhin et al., 2020), BM25 remains competitive on out-of- domain datasets (Thakur et al., 2021). Query expansion (Rocchio, 1971; Lavrenko and Croft, 2001) is a long-standing technique that rewrites the query based on pseudo-relevance feedback or external knowledge sources such as WordNet. For sparse retrieval, it can help bridge the lexical gap between the query and the docu- ments. However, query expansion methods like RM3 (Lavrenko and Croft, 2001; Lv and Zhai, 2009) have only shown limited success on populardatasets (Campos et al., 2016), and most state-of- the-art dense retrievers do not adopt this technique. In the meantime, document expansion methods like doc2query (Nogueira et al., 2019) have proven to be effective for sparse retrieval. In this paper, we demonstrate the effectiveness of LLMs (Brown et al., 2020) as query expan- sion models by generating pseudo-documents con- ditioned on few-shot prompts. Given that search queries are often short, ambiguous, or lack neces- sary background information, LLMs can provide relevant information to guide retrieval systems, as they memorize an enormous amount of knowledge and language patterns by pre-training on trillions of tokens. Our proposed method, called query2doc , gen- erates pseudo-documents by few-shot prompting LLMs and concatenates them with the original query to form a new query. This method is simple to implement and does not require any changes in training pipelines or model architectures, making it orthogonal to the progress in the field of LLMs and information retrieval. Future methods can easily build upon our query expansion framework. For in-domain evaluation, we adopt the MS- MARCO passage ranking (Campos et al., 2016), TREC DL 2019 and 2020 datasets. Pseudo- documents are generated by prompting an im- proved version of GPT-3 text-davinci-003 from OpenAI (Brown et al., 2020). Results show thatquery2doc substantially improves the off-the- shelf BM25 algorithm without fine-tuning any model, particularly for hard queries from the TREC DL track. Strong dense retrievers, in- cluding DPR (Karpukhin et al., 2020), SimLM (Wang et al., 2023), and E5 (Wang et al., 2022) also benefit from query2doc , although the gains tend to be diminishing when distilling from a strong cross-encoder based re-ranker. Experi- ments in zero-shot OOD settings demonstrate that our method outperforms strong baselines on mostarXiv:2303.07678v2 [cs.IR] 11 Oct 2023
2303.03378.pdf	PaLM-E: An Embodied Multimodal Language Model Danny Driess1 2Fei Xia1Mehdi S. M. Sajjadi3Corey Lynch1Aakanksha Chowdhery3 Brian Ichter1Ayzaan Wahid1Jonathan Tompson1Quan Vuong1Tianhe Yu1Wenlong Huang1 Yevgen Chebotar1Pierre Sermanet1Daniel Duckworth3Sergey Levine1Vincent Vanhoucke1 Karol Hausman1Marc Toussaint2Klaus Greff3Andy Zeng1Igor Mordatch3Pete Florence1 1Robotics at Google2TU Berlin3Google Research https://palm-e.github.io Giv en <im g> . Q: Wh a t’ s in t he im a g e? An sw er in emo jis. A: 🍏🍌🍇🍐🍑🍈🍒 . Sc ene Und er s t andin g V isu al Q&A Giv en <im g1> and <im g2> Q: wh a t di d t he r o bo t p i ck up ? A: The mul t i-gr ain chip s. Giv en <im g1> Q: wh a t d a y s mi gh t I mo s t c o mmo nly g o t o t his b uil din g? A: S und a y . TBD TBD TBD T a sk and M o t i o n P l annin g Giv en <emb> Q: H o w t o gr a s p b lu e b l ock? A: F ir s t gr a s p y ell o w b l ock and p l a c e i t o n t he t ab l e, t hen gr a s p t he b lu e b l ock. Giv en <im g> T a sk: So rt c o l o r s in t o c o rner s. S t ep 1. P u sh t he gr e en s t ar t o t he bo tt o m l e ft. S t ep 2. P u sh t he gr e en cir cl e t o t he gr e en s t ar . T ab l e t o p M anip ul a t i o n M o b il e M anip ul a t i o n V isu al Q&A, C ap t i o nin g … H um an: Brin g me t he ri c e chip s fr o m t he dr a w er . R o bo t: 1. G o t o t he dr a w er s, 2. Open t o p dr a w er . I s e e <im g> . 3 . Pi ck t he gr e en ri c e chip ba g fr o m t he dr a w er and p l a c e i t o n t he c o un t er . A: F ir s t, gr a s p y ell o w b l ock and … Giv en <emb> … <im g> Q: H o w t o gr a s p b lu e b l ock? A: F ir s t, gr a s p y ell o w b l ock Lar g e Lan g u a g e M od el (P aLM) ? C o n tr o l P aLM-E: An E mbodi e d M ul t i mod al Lan g u a g e M od el … … V iT Lan g u a g e O nly T a sk s H er e is a H aiku abo u t embodi e d l an g u a g e mod el s: E mbodi e d l an g u a g e mod el s ar e t he f u tur e o f n a tur al l an g u a g e D e s c ribe t he f o ll o win g <im g> : A d og jump in g o v er a hur dl e a t a d og sho w . … … Q: Mi ami Be a ch bo r d er s whi ch oc e an ? A: A t l an t i c. Q: Wh a t is 372 x 18? A: 669 6. Lan g u a g e mod el s tr aine d o n r o bo t s en s o r d a t a c an be u s e d t o g ui d e a r o bo t’ s a c t i o n s. PROMPT: Q: How can embodied language models benefit robots? A: PREDICTION: Embodied language models can benefit robots by allowing them to learn language in a more natural way. PROMPT: Language models which understand robot sensor data can PREDICTION: be used to generate natural language descriptions of the robot's environment. Figure 1: PaLM-E is a single general-purpose multimodal language model for embodied reasoning tasks, visual-language tasks, and language tasks. PaLM-E transfers knowledge from visual-language domains into embodied reasoning – from robot planning in environments with complex dynamics and physical constraints, to answering questions about the observable world. PaLM-E operates on multimodal sentences , i.e. sequences of tokens where inputs from arbitrary modalities (e.g. images, neural 3D representations, or states, in green and blue) are inserted alongside text tokens (in orange) as input to an LLM, trained end-to-end. Abstract Large language models have been demonstrated to perform complex tasks. However, enabling general inference in the real world, e.g. for robotics problems, raises the challenge of grounding. We propose embodied language models to di- rectly incorporate real-world continuous sensor modalities into language models and thereby establish the link be- tween words and percepts. Input to our embodied language model are multi-modal sentences that interleave visual, con- tinuous state estimation, and textual input encodings. We train these encodings end-to-end, in conjunction with a pre- trained large language model, for multiple embodied tasks including sequential robotic manipulation planning, visual question answering, and captioning. Our evaluations show that PaLM-E, a single large embodied multimodal model, can address a variety of embodied reasoning tasks, from a variety of observation modalities, on multiple embodiments, and further, exhibits positive transfer: the model beneﬁts from diverse joint training across internet-scale language, vision, and visual-language domains. Our largest model, PaLM-E-562B with 562B parameters, in addition to being trained on robotics tasks, is a visual-language generalist with state-of-the-art performance on OK-VQA, and retains generalist language capabilities with increasing scale.1. Introduction Large language models (LLMs) demonstrate strong reason- ing capabilities across various domains, including dialogue (Glaese et al., 2022; Thoppilan et al., 2022), step-by-step reasoning (Wei et al., 2022; Kojima et al., 2022), math prob- lem solving (Lewkowycz et al., 2022; Polu et al., 2022), and code writing (Chen et al., 2021a). However, a limitation of such models for inference in the real world is the issue of grounding: while training LLMs on massive textual data may lead to representations that relate to our physical world, connecting those representations toreal-world visual and physical sensor modalities is essential to solving a wider range of grounded real-world problems in computer vision and robotics (Tellex et al., 2020). Previous work (Ahn et al., 2022) interfaces the output of LLMs with learned robotic policies and affordance functions to make decisions, but is limited in that the LLM itself is only provided with textual input, which is insufﬁcient for many tasks where the geo- metric conﬁguration of the scene is important. Further, in our experiments we show that current state-of-the-art visual - language models trained on typical vision-language tasks such as visual-question-answering (VQA) cannot directly solve robotic reasoning tasks. In this paper we propose embodied language models, whicharXiv:2303.03378v1 [cs.LG] 6 Mar 2023

Subsets and Splits