ICLR 2024 Papers

In this work, we theoretically investigate the generalization properties of neural networks (NN) trained by stochastic gradient descent (SGD) with large learning rates. Under such a training regime, our finding is that, the oscillation of the NN weights caused by SGD with large learning rates turns out to be beneficial to the generalization of the NN, potentially improving over the same NN trained by SGD with small learning rates that converges more smoothly. In view of this finding, we call such a phenomenon “benign oscillation”. Our theory towards demystifying such a phenomenon builds upon the feature learning perspective of deep learning. Specifically, we consider a feature-noise data generation model that consists of (i) weak features which have a small $\ell_2$-norm and appear in each data point; (ii) strong features which have a large $\ell_2$-norm but appear only in a certain fraction of all data points; and (iii) noise. We prove that NNs trained by oscillating SGD with a large learning rate can effectively learn the weak features in the presence of those strong features. In contrast, NNs trained by SGD with a small learning rate can only learn the strong features but make little progress in learning the weak features. Consequently, when it comes to the new testing data points that consist of only weak features, the NN trained by oscillating SGD with a large learning rate can still make correct predictions, while the NN trained by SGD with a small learning rate could not. Our theory sheds light on how large learning rate training benefits the generalization of NNs. Experimental results demonstrate our findings on the phenomenon of “benign oscillation”.

In the classical transformer attention scheme, we are given three $n \times d$ size matrices $Q, K, V$ (the query, key, and value tokens), and the goal is to compute a new $n \times d$ size matrix $D^{-1} \exp(QK^\top) V$ where $D = \mathrm{diag}( \exp(QK^\top) {\bf 1}_n )$. Here, $\exp()$ is applied entry-wise and ${\bf 1}_n$ denotes a length-$n$ vector whose entries are all ones. Intuitively, attention computation captures pairwise information between words in a sentence, but not higher-order information. Indeed, recent work \cite{sht23} has shown that attention units cannot solve simple problems about detecting triples of connected words. In this work, we study a generalization of attention which captures triple-wise correlations. The generalization is based on computations involving tensors defined by tuples of words. More formally, given five $n \times d$ size matrices $Q, K_1, K_2, V_1$ and $V_2$ (generalized query, key, and value tokens), our new goal is to compute an $n \times d$ size matrix $D^{-1} \exp( Q ( K_1 \oslash K_2)^\top ) (V_1 \oslash V_2) $ where $D = \mathrm{diag}( \exp( Q ( K_1 \oslash K_2)^\top ) {\bf 1}_{n^2} )$ and $K_1 \oslash K_2 \in \mathbb{R}^{n^2 \times d}$ denotes the column-wise Kronecker product of $K_1$ and $K_2$. This generalization is indeed able to solve problems about detecting triple-wise connections that were shown to be impossible for transformers. The potential downside of this generalization is that it appears as though computations are even more difficult, since the straightforward algorithm requires cubic time in $n$. However, we show that in the bounded-entry setting (which arises in practice, and which is well-studied in both theory and practice), there is actually a near-linear time algorithm. More precisely, we show that bounded entries are both necessary and sufficient for quickly performing generalized computations: $\bullet$ On the positive side, if all entries of the input matrices are bounded above by $o(\sqrt[3]{\log n})$ then we show how to approximate the ``tensor-type'' attention matrix in $n^{1+o(1)}$ time. $\bullet$ On the negative side, we show that if the entries of the input matrices may be as large as $\Omega(\sqrt[3]{\log n})$, then there is no algorithm that runs faster than $n^{3-o(1)}$ (assuming the Strong Exponential Time Hypothesis from fine-grained complexity theory). We also show that our construction, algorithms, and lower bounds naturally generalize to higher-order tensors and correlations. Interestingly, the higher the order of the tensors, the lower the bound on the entries needs to be for an efficient algorithm. Our results thus yield a natural tradeoff between the boundedness of the entries, and order of the tensor one may use for more expressive, efficient attention computation. Our constructions make use of a novel connection with a higher-order variant on the kernel density estimation problem. They combine a number of technical tools, including the polynomial method, algebraic geometry codes, and multiparty Merlin-Arthur communication protocols.

Generative flow networks (GFlowNets) are sequential sampling models trained to match a given distribution. GFlowNets have been successfully applied to various structured object generation tasks, sampling a diverse set of high-reward objects quickly. We propose expected flow networks (EFlowNets), which extend GFlowNets to stochastic environments. We show that EFlowNets outperform other GFlowNet formulations in stochastic tasks such as protein design. We then extend the concept of EFlowNets to adversarial environments, proposing adversarial flow networks (AFlowNets) for two-player zero-sum games. We show that AFlowNets learn to find above 80% of optimal moves in Connect-4 via self-play and outperform AlphaZero in tournaments. Code: https://github.com/GFNOrg/AdversarialFlowNetworks.

The standard active learning setting assumes a willing labeler, who provides labels on informative examples to speed up learning. However, if the labeler wishes to be compensated for as many labels as possible before learning finishes, the labeler may benefit from actually slowing down learning. This incentive arises for instance if the labeler is to be replaced by the ML model once it is trained. In this paper, we initiate the study of learning from a strategic labeler, who may abstain from labeling to slow down learning. We first prove that strategic abstention can prolong learning, and propose a novel complexity measure and representation to analyze the query complexity of the learning game. Next, we develop a near-optimal deterministic algorithm, prove its robustness to strategic labeling, and contrast it with other active learning algorithms. We also analyze extensions that encompass more general learning goals and labeler assumptions. Finally, we characterize the query cost of multi-task active learning, with and without abstention. Our first exploration of strategic labeling aims to consolidate our theoretical understanding of the \emph{imitative} nature of ML in human-AI interaction.

When applied to question answering and other text generation tasks, language models (LMs) may be queried generatively (by sampling answers from their output distribution) or discriminatively (by using them to score or rank a set of candidate answers). These procedures sometimes yield very different predictions. How do we reconcile mutually incompatible scoring procedures to obtain coherent LM predictions? We introduce a new, a training-free, game-theoretic procedure for language model decoding. Our approach casts language model decoding as a regularized imperfect-information sequential signaling game—which we term the concensus game—in which a generator seeks to communicate an abstract correctness parameter using natural language sentences to a discriminator. We develop computational procedures for finding approximate equilibria of this game, resulting in a decoding algorithm we call equilibrium-ranking. Applied to a large number of tasks (including reading comprehension, commonsense reasoning, mathematical problem-solving, and assistive dialog), equilibrium-ranking consistently improves performance over existing LM decoding procedures. These improvements are sometimes substantial—on multiple benchmarks, we observe that applying equilibrium-ranking to LLaMA-7B outperforms the much larger LLaMA-65B and PaLM-540B models.

Few neural architectures lend themselves to provable learning with gradient based methods. One popular model is the single-index model, in which labels are produced by composing an unknown linear projection with a possibly unknown scalar link function. Learning this model with SGD is relatively well-understood, whereby the so-called information exponent of the link function governs a polynomial sample complexity rate. However, extending this analysis to deeper or more complicated architectures remains challenging. In this work, we consider single index learning in the setting of symmetric neural networks. Under analytic assumptions on the activation and maximum degree assumptions on the link function, we prove that gradient flow recovers the hidden planted direction, represented as a finitely supported vector in the feature space of power sum polynomials. We characterize a notion of information exponent adapted to our setting that controls the efficiency of learning.

We propose the first loss function for approximate Nash equilibria of normal-form games that is amenable to unbiased Monte Carlo estimation. This construction allows us to deploy standard non-convex stochastic optimization techniques for approximating Nash equilibria, resulting in novel algorithms with provable guarantees. We complement our theoretical analysis with experiments demonstrating that stochastic gradient descent can outperform previous state-of-the-art approaches.

One prominent approach toward resolving the adversarial vulnerability of deep neural networks is the two-player zero-sum paradigm of adversarial training, in which predictors are trained against adversarially chosen perturbations of data. Despite the promise of this approach, algorithms based on this paradigm have not engendered sufficient levels of robustness and suffer from pathological behavior like robust overfitting. To understand this shortcoming, we first show that the commonly used surrogate-based relaxation used in adversarial training algorithms voids all guarantees on the robustness of trained classifiers. The identification of this pitfall informs a novel non-zero-sum bilevel formulation of adversarial training, wherein each player optimizes a different objective function. Our formulation yields a simple algorithmic framework that matches and in some cases outperforms state-of-the-art attacks, attains comparable levels of robustness to standard adversarial training algorithms, and does not suffer from robust overfitting.

We introduce the use of generative adversarial learning to compute equilibria in general game-theoretic settings, specifically the generalized Nash equilibrium (GNE) in pseudo-games, and its specific instantiation as the competitive equilibrium (CE) in Arrow-Debreu competitive economies. Pseudo-games are a generalization of games in which players' actions affect not only the payoffs of other players but also their feasible action spaces. Although the computation of GNE and CE is intractable in the worst-case, i.e., PPAD-hard, in practice, many applications only require solutions with high accuracy in expectation over a distribution of problem instances. We introduce Generative Adversarial Equilibrium Solvers (GAES): a family of generative adversarial neural networks that can learn GNE and CE from only a sample of problem instances. We provide computational and sample complexity bounds for Lipschitz-smooth function approximators in a large class of concave pseudo-games, and apply the framework to finding Nash equilibria in normal-form games, CE in Arrow-Debreu competitive economies, and GNE in an environmental economic model of the Kyoto mechanism.

Contrastive learning is a highly successful technique for learning representations of data from labeled tuples, specifying the distance relations within the tuple. We study the sample complexity of contrastive learning, i.e. the minimum number of labeled tuples sufficient for getting high generalization accuracy. We give tight bounds on the sample complexity in a variety of settings, focusing on arbitrary distance functions, $\ell_p$-distances, and tree metrics. Our main result is an (almost) optimal bound on the sample complexity of learning $\ell_p$-distances for integer $p$. For any $p \ge 1$, we show that $\tilde \Theta(nd)$ labeled tuples are necessary and sufficient for learning $d$-dimensional representations of $n$-point datasets. Our results hold for an arbitrary distribution of the input samples and are based on giving the corresponding bounds on the Vapnik-Chervonenkis/Natarajan dimension of the associated problems. We further show that the theoretical bounds on sample complexity obtained via VC/Natarajan dimension can have strong predictive power for experimental results, in contrast with the folklore belief about a substantial gap between the statistical learning theory and the practice of deep learning.

In this paper, we study inverse game theory (resp. inverse multiagent learning) in which the goal is to find parameters of a game’s payoff functions for which the expected (resp. sampled) behavior is an equilibrium. We formulate these problems as generative-adversarial (i.e., min-max) optimization problems, which we develop polynomial-time algorithms to solve, the former of which relies on an exact first- order oracle, and the latter, a stochastic one. We extend our approach to solve inverse multiagent simulacral learning in polynomial time and number of samples. In these problems, we seek a simulacrum, meaning parameters and an associated equilibrium that replicate the given observations in expectation. We find that our approach outperforms the widely-used ARIMA method in predicting prices in Spanish electricity markets based on time-series data.

In competitive two-agent environments, deep reinforcement learning (RL) methods like Policy Space Response Oracles (PSRO) often increase exploitability between iterations, which is problematic when training in large games. To address this issue, we introduce anytime double oracle (ADO), an algorithm that ensures exploitability does not increase between iterations, and its approximate extensive-form version, anytime PSRO (APSRO). ADO converges to a Nash equilibrium while iteratively reducing exploitability. However, convergence in these algorithms may require adding all of a game's deterministic policies. To improve this, we propose Self-Play PSRO (SP-PSRO), which incorporates an approximately optimal stochastic policy into the population in each iteration. APSRO and SP-PSRO demonstrate lower exploitability and near-monotonic exploitability reduction in games like Leduc poker and Liar's Dice. Empirically, SP-PSRO often converges much faster than APSRO and PSRO, requiring only a few iterations in many games.

Understanding how overparameterized neural networks generalize despite perfect interpolation of noisy training data is a fundamental question. Mallinar et. al. (2022) noted that neural networks seem to often exhibit ``tempered overfitting'', wherein the population risk does not converge to the Bayes optimal error, but neither does it approach infinity, yielding non-trivial generalization. However, this has not been studied rigorously. We provide the first rigorous analysis of the overfiting behaviour of regression with minimum norm ($\ell_2$ of weights), focusing on univariate two-layer ReLU networks. We show overfitting is tempered (with high probability) when measured with respect to the $L_1$ loss, but also show that the situation is more complex than suggested by Mallinar et. al., and overfitting is catastrophic with respect to the $L_2$ loss, or when taking an expectation over the training set.

Normal-form games (NFGs) are the fundamental model of *strategic interaction*. We study their representation using neural networks. We describe the inherent equivariance of NFGs --- any permutation of strategies describes an equivalent game --- as well as the challenges this poses for representation learning. We then propose the NfgTransformer architecture that leverages this equivariance, leading to state-of-the-art performance in a range of game-theoretic tasks including equilibrium-solving, deviation gain estimation and ranking, with a common approach to NFG representation. We show that the resulting model is interpretable and versatile, paving the way towards deep learning systems capable of game-theoretic reasoning when interacting with humans and with each other.

Memorization with neural networks is to study the expressive power of neural networks to interpolate a finite classification data set, which is closely related to the generalizability of deep learning. However, the important problem of robust memorization has not been thoroughly studied. In this paper, several basic problems about robust memorization are solved. First, we prove that it is NP-hard to compute neural networks with certain simple structures, which are robust memorization. A network hypothesis space is called optimal robust memorization for a data set if it can achieve robust memorization for any budget less than half the separation bound of the data set. Second, we explicitly construct neural networks with O(N n) parameters for optimal robust memorization of any data set with dimension n and size N . We also give a lower bound for the width of networks to achieve optimal robust memorization. Finally, we explicitly construct neural networks with O(N n log n) parameters for optimal robust memorization of any binary classification data set by controlling the Lipschitz constant of the network.

This paper studies the sample-efficiency of learning in Partially Observable Markov Decision Processes (POMDPs), a challenging problem in reinforcement learning that is known to be exponentially hard in the worst-case. Motivated by real-world settings such as loading in game playing, we propose an enhanced feedback model called ``multiple observations in hindsight'', where after each episode of interaction with the POMDP, the learner may collect multiple additional observations emitted from the encountered latent states, but may not observe the latent states themselves. We show that sample-efficient learning under this feedback model is possible for two new subclasses of POMDPs: \emph{multi-observation revealing POMDPs} and \emph{distinguishable POMDPs}. Both subclasses generalize and substantially relax \emph{revealing POMDPs}---a widely studied subclass for which sample-efficient learning is possible under standard trajectory feedback. Notably, distinguishable POMDPs only require the emission distributions from different latent states to be \emph{different} instead of \emph{linearly independent} as required in revealing POMDPs.