COLT 2021 Papers

We study the extent to which wide neural networks may be approximated by Gaussian processes, when initialized with random weights. It is a well-established fact that as the width of a network goes to infinity, its law converges to that of a Gaussian process. We make this quantitative by establishing explicit convergence rates for the central limit theorem in an infinite-dimensional functional space, metrized with a natural transportation distance. We identify two regimes of interest; when the activation function is polynomial, its degree determines the rate of convergence, while for non-polynomial activations, the rate is governed by the smoothness of the function.

Differentially private (DP) stochastic convex optimization (SCO) is a fundamental problem, where the goal is to approximately minimize the population risk with respect to a convex loss function, given a dataset of i.i.d. samples from a distribution, while satisfying differential privacy with respect to the dataset. Most of the existing works in the literature of private convex optimization focus on the Euclidean (i.e., $\ell_2$) setting, where the loss is assumed to be Lipschitz (and possibly smooth) w.r.t. the $\ell_2$ norm over a constraint set with bounded $\ell_2$ diameter. Algorithms based on noisy stochastic gradient descent (SGD) are known to attain the optimal excess risk in this setting. In this work, we conduct a systematic study of DP-SCO for $\ell_p$-setups. For $p=1$, under a standard smoothness assumption, we give a new algorithm with nearly optimal excess risk. This result also extends to general polyhedral norms and feasible sets. For $p\in(1, 2)$, we give two new algorithms, whose central building block is a novel privacy mechanism, which generalizes the Gaussian mechanism. Moreover, we establish a lower bound on the excess risk for this range of $p$, showing a necessary dependence on $\sqrt{d}$, where $d$ is the dimension of the space. Our lower bound implies a sudden transition of the excess risk at $p=1$, where the dependence on $d$ changes from logarithmic to polynomial, resolving an open question in prior work \citep{TTZ15a}. For $p\in (2, \infty)$, noisy SGD attains optimal excess risk in the low-dimensional regime; in particular, this proves the optimality of noisy SGD for $p=\infty$. Our work draws upon concepts from the geometry of normed spaces, such as the notions of regularity, uniform convexity, and uniform smoothness.

We study reinforcement learning (RL) with linear function approximation where the underlying transition probability kernel of the Markov decision process (MDP) is a linear mixture model (Jia et al., 2020; Ayoub et al., 2020; Zhou et al., 2020) and the learning agent has access to either an integration or a sampling oracle of the individual basis kernels. For the fixed-horizon episodic setting with inhomogeneous transition kernels, we propose a new, computationally efficient algorithm that uses the basis kernels to approximate value functions. We show that the new algorithm, which we call ${\text{UCRL-VTR}^{+}}$, attains an $\tilde O(dH\sqrt{T})$ regret where $d$ is the number of basis kernels, $H$ is the length of the episode and $T$ is the number of interactions with the MDP. We also prove a matching lower bound $\Omega(dH\sqrt{T})$ for this setting, which shows that ${\text{UCRL-VTR}^{+}}$ is minimax optimal up to logarithmic factors. At the core of our results are (1) a weighted least squares estimator for the unknown transitional probability; and (2) a new Bernstein-type concentration inequality for self-normalized vector-valued martingales with bounded increments. Together, these new tools enable tight control of the Bellman error and lead to a nearly minimax regret. To the best of our knowledge, this is the first computationally efficient, nearly minimax optimal algorithm for RL with linear function approximation.

Many real-world data sets are sparse or almost sparse. One method to measure this for a matrix $A\in \mathbb{R}^{n\times n}$ is the \emph{numerical sparsity}, denoted $\mathsf{ns}(A)$, defined as the minimum $k\geq 1$ such that $\|a\|_1/\|a\|_2 \leq \sqrt{k}$ for every row and every column $a$ of $A$. This measure of $a$ is smooth and is clearly only smaller than the number of non-zeros in the row/column $a$. The seminal work of Achlioptas and McSherry (2007) has put forward the question of approximating an input matrix $A$ by entrywise sampling. More precisely, the goal is to quickly compute a sparse matrix $\tilde{A}$ satisfying $\|A - \tilde{A}\|_2 \leq \epsilon \|A\|_2$ (i.e., additive spectral approximation) given an error parameter $\epsilon>0$. The known schemes sample and rescale a small fraction of entries from $A$. We propose a scheme that sparsifies an almost-sparse matrix $A$ — it produces a matrix $\tilde{A}$ with $O(\epsilon^{-2}\mathsf{ns}(A) \cdot n\ln n)$ non-zero entries with high probability. We also prove that this upper bound on $\mathsf{nnz}(\tilde{A})$ is \emph{tight} up to logarithmic factors. Moreover, our upper bound improves when the spectrum of $A$ decays quickly (roughly replacing $n$ with the stable rank of $A$). Our scheme can be implemented in time $O(\mathsf{nnz}(A))$ when $\|A\|_2$ is given. Previously, a similar upper bound was obtained by Achlioptas et al. (2013), but only for a restricted class of inputs that does not even include symmetric or covariance matrices. Finally, we demonstrate two applications of these sampling techniques, to faster approximate matrix multiplication, and to ridge regression by using sparse preconditioners.

The stochastic multi-armed bandit model captures the tradeoff between exploration and exploitation. We study the effects of competition and cooperation on this tradeoff. Suppose there are two arms, one predictable and one risky, and two players, Alice and Bob. In every round, each player pulls an arm, receives the resulting reward, and observes the choice of the other player but not their reward. Alice’s utility is $\Gamma_A + \lambda \Gamma_B$ (and similarly for Bob), where $\Gamma_A$ is Alice’s total reward and $\lambda \in [-1, 1]$ is a cooperation parameter. At $\lambda = -1$ the players are competing in a zero-sum game, at $\lambda = 1$, their interests are aligned, and at $\lambda = 0$, they are neutral: each player’s utility is their own reward. The model is related to the economics literature on strategic experimentation, where usually players observe each other’s rewards. Suppose the predictable arm has success probability $p$ and the risky arm has prior $\mu$. If the discount factor is $\beta$, then the value of $p$ where a single player is indifferent between the arms is the Gittins index $g = g(\mu,\beta) > m$, where $m$ is the mean of the risky arm. Our first result answers, in this setting, a fundamental question posed by Rothschild \cite{rotschild}. We show that competing and neutral players eventually settle on the same arm (even though it may not be the best arm) in every Nash equilibrium, while this can fail for players with aligned interests. Moreover, we show that \emph{competing players} explore \emph{less} than a single player: there is $p^* \in (m, g)$ so that for all $p > p^*$, the players stay at the predictable arm. However, the players are not myopic: they still explore for some $p > m$. On the other hand, \emph{cooperating players} (with $\lambda =1$) explore \emph{more} than a single player. We also show that \emph{neutral players} learn from each other, receiving strictly higher total rewards than they would playing alone, for all $ p\in (p^*, g)$, where $p^*$ is the threshold above which competing players do not explore.

We show how to achieve the notion of "multicalibration" from Hebert-Johnson et al. (2018) not just for means, but also for variances and other higher moments. Informally, this means that we can find regression functions which, given a data point, can make point predictions not just for the expectation of its label, but for higher moments of its label distribution as well—and those predictions match the true distribution quantities when averaged not just over the population as a whole, but also when averaged over an enormous number of finely defined subgroups. It yields a principled way to estimate the uncertainty of predictions on many different subgroups—and to diagnose potential sources of unfairness in the predictive power of features across subgroups. As an application, we show that our moment estimates can be used to derive marginal prediction intervals that are simultaneously valid as averaged over all of the (sufficiently large) subgroups for which moment multicalibration has been obtained.

This paper proposes a new mean-field framework for over-parameterized deep neural networks (DNNs), which can be used to analyze neural network training. In this framework, a DNN is represented by probability measures and functions over its features (that is, the function values of the hidden units over the training data) in the continuous limit, instead of the neural network parameters as most existing studies have done. This new representation overcomes the degenerate situation where all the hidden units essentially have only one meaningful hidden unit in each middle layer, leading to a simpler representation of DNNs. Moreover, we construct a non-linear dynamics called neural feature flow, which captures the evolution of an over-parameterized DNN trained by Gradient Descent. We illustrate the framework via the Residual Network (Res-Net) architecture. It is shown that when the neural feature flow process converges, it reaches a global minimal solution under suitable conditions.

We establish a connection between the stability of mirror descent and the information ratio by Russo and Van Roy (2014). Our analysis shows that mirror descent with suitable loss estimators and exploratory distributions enjoys the same bound on the adversarial regret as the bounds on the Bayesian regret for information-directed sampling. Along the way, we develop the theory for information-directed sampling and provide an efficient algorithm for adversarial bandits for which the regret upper bound matches exactly the best known information-theoretic upper bound. Keywords: Bandits, partial monitoring, mirror descent, information theory.

We study the stochastic shortest path problem with adversarial costs and known transition, and show that the minimax regret is $O(\sqrt{DT_\star K})$ and $O(\sqrt{DT_\star SA K})$ for the full-information setting and the bandit feedback setting respectively, where $D$ is the diameter, $T_\star$ is the expected hitting time of the optimal policy, $S$ is the number of states, $A$ is the number of actions, and $K$ is the number of episodes. Our results significantly improve upon the recent work of (Rosenberg and Mansour, 2020) which only considers the full-information setting and achieves suboptimal regret. Our work is also the first to consider bandit feedback with adversarial costs. Our algorithms are built on top of the Online Mirror Descent framework with a variety of new techniques that might be of independent interest, including an improved multi-scale expert algorithm, a reduction from general stochastic shortest path to a special loop-free case, a skewed occupancy measure space, and a novel correction term added to the cost estimators. Interestingly, the last two elements reduce the variance of the learner via positive bias and the variance of the optimal policy via negative bias respectively, and having them simultaneously is critical for obtaining the optimal high-probability bound in the bandit feedback setting.

We introduce the technique of generic chaining and majorizing measures for controlling sequential Rademacher complexity. We relate majorizing measures to the notion of fractional covering numbers, which we show to be dominated in terms of sequential scale-sensitive dimensions in a horizon-independent way, and, under additional complexity assumptions establish a tight control on worst-case sequential Rademacher complexity in terms of the integral of sequential scale-sensitive dimension. Finally, we establish a tight contraction inequality for worst-case sequential Rademacher complexity. The above constitutes the resolution of a number of outstanding open problems in extending the classical theory of empirical processes to the sequential case, and, in turn, establishes sharp results for online learning.

We study the problem of machine unlearning and identify a notion of algorithmic stability, Total Variation (TV) stability, which we argue, is suitable for the goal of exact unlearning. For convex risk minimization problems, we design TV-stable algorithms based on noisy Stochastic Gradient Descent (SGD). Our key contribution is the design of corresponding efficient unlearning algorithms, which are based on constructing a near-maximal coupling of Markov chains for the noisy SGD procedure. To understand the trade-offs between accuracy and unlearning efficiency, we give upper and lower bounds on excess empirical and populations risk of TV stable algorithms for convex risk minimization. Our techniques generalize to arbitrary non-convex functions, and our algorithms are differentially private as well.

A number of machine learning tasks entail a high degree of invariance: the data distribution does not change if we act on the data with a certain group of transformations. For instance, labels of images are invariant under translations of the images. Certain neural network architectures —for instance, convolutional networks—are believed to owe their success to the fact that they exploit such invariance properties. With the objective of quantifying the gain achieved by invariant architectures, we introduce two classes of models: invariant random features and invariant kernel methods. The latter includes, as a special case, the neural tangent kernel for convolutional networks with global average pooling. We consider uniform covariates distributions on the sphere and hypercube and a general invariant target function. We characterize the test error of invariant methods in a high-dimensional regime in which the sample size and number of hidden units scale as polynomials in the dimension, for a class of groups that we call ‘degeneracy $\alpha$’, with $\alpha \leq 1$. We show that exploiting invariance in the architecture saves a $d^\alpha$ factor ($d$ stands for the dimension) in sample size and number of hidden units to achieve the same test error as for unstructured architectures. Finally, we show that output symmetrization of an unstructured kernel estimator does not give a significant statistical improvement; on the other hand, data augmentation with an unstructured kernel estimator is equivalent to an invariant kernel estimator and enjoys the same improvement in statistical efficiency.

We consider a discounted infinite horizon optimal stopping problem. If the underlying distribution is known a priori, the solution of this problem is obtained via dynamic programming (DP) and is given by a well known threshold rule. When information on this distribution is lacking, a natural (though naive) approach is “explore-then-exploit," whereby the unknown distribution or its parameters are estimated over an initial exploration phase, and this estimate is then used in the DP to determine actions over the residual exploitation phase. We show: (i) with proper tuning, this approach leads to performance comparable to the full information DP solution; and (ii) despite common wisdom on the sensitivity of such “plug in" approaches in DP due to propagation of estimation errors, a surprisingly “short" (logarithmic in the horizon) exploration horizon suffices to obtain said performance. In cases where the underlying distribution is heavy-tailed, these observations are even more pronounced: a single sample exploration phase suffices.

We study the problem of learning Censored Markov Random Fields (abbreviated CMRFs), which are Markov Random Fields where some of the nodes are censored (i.e. not observed). We assume the CMRF is high temperature but, crucially, make no assumption about its structure. This makes structure learning impossible. Nevertheless we introduce a new definition, which we call learning to sample, that circumvents this obstacle. We give an algorithm that can learn to sample from a distribution within $\epsilon n$ earthmover distance of the target distribution for any $\epsilon > 0$. We obtain stronger results when we additionally assume high girth, as well as computational lower bounds showing that these are essentially optimal.

We study the problem of learning an unknown mixture of k permutations over n elements, given access to noisy samples drawn from the unknown mixture. We consider a range of different noise models, including natural variants of the “heat kernel” noise framework and the Mallows model. We give an algorithm which, for each of these noise models, learns the unknown mixture to high accuracy under mild assumptions and runs in $n^{O(log k)}$ time. Our approach is based on a new procedure that recovers an unknown mixture of permutations from noisy higher-order marginals.

Many multi-agent systems with strategic interactions have their desired functionality encoded as the Nash equilibrium of a game, e.g. machine learning architectures such as Generative Adversarial Networks. Directly computing a Nash equilibrium of these games is often impractical or impossible in practice, which has led to the development of numerous learning algorithms with the goal of iteratively converging on a Nash equilibrium. Unfortunately, the dynamics generated by the learning process can be very intricate and instances failing to converge become hard to interpret. In this paper we show that, in a strong sense, this dynamic complexity is inherent to games. Specifically, we prove that replicator dynamics, the continuous-time analogue of Multiplicative Weights Update, even when applied in a very restricted class of games–known as finite matrix games–is rich enough to be able to approximate arbitrary dynamical systems. In the context of machine learning, our results are positive in the sense that they show the nearly boundless dynamic modelling capabilities of current machine learning practices, but also negative in implying that these capabilities may come at the cost of interpretability. As a concrete example, we show how replicator dynamics can effectively reproduce the well-known strange attractor of Lonrenz dynamics (the “butterfly effect") while achieving no regret.

Observations from dynamical systems often exhibit irregularities, such as censoring, where values are recorded only if they fall within a certain range. Censoring is ubiquitous in practice, due to saturating sensors, limit-of-detection effects, image frame effects, and combined with the temporal dependencies within the data, makes the task of system identification particularly challenging. In light of recent developments on learning linear dynamical systems (LDSs), and on censored statistics with independent data, we revisit the decades-old problem of learning an LDS, from censored observations (Lee and Maddala (1985), Zeger and Brookmeyer (1986)). Here, the learner observes the state x_t \in R^d if and only if x_t belongs to some set S_t _x0012_\in R^d. We develop the first computationally and statistically efficient algorithm for learning the system, assuming only oracle to the sets St. Our algorithm, Stochastic Online Newton with Switching Gradients, is a novel second-order method that builds on the Online Newton Step (ONS) of Hazan et al. (2007). Our Switching-Gradient scheme does not always use (stochastic) gradients of the function we want to optimize, which we call censor-aware function. Instead, in each iteration, it performs a simple test to decide whether to use the censor-aware, or another censor-oblivious function, for getting a stochastic gradient. In our analysis, we consider a “generic” Online Newton method, which uses arbitrary vectors instead of gradients, and we prove an error-bound for it. This can be used to appropriately design these vectors, Leading to our Switching-Gradient scheme. This framework significantly deviates from the recent long line of works on censored statistics (e.g, Daskalakis et al. (2018); Kontonis et al. (2019); Daskalakis et al. (2019), which apply Stochastic Gradient Descent (SGD), and their analysis reduces to establishing conditions for off-the-shelf SGD-bounds. Our approach enables to relax these conditions, and gives rise to phenomena that might appear counterintuitive, given the previous works. Specifically, our method makes progress even when the current “survival probability” is exponentially small. We believe that our analysis framework will have applications in more settings where the data are subject to censoring.

We study the problems of learning and testing junta distributions on $\{-1,1\}^n$ with respect to the uniform distribution, where a distribution $p$ is a $k$-junta if its probability mass function $p(x)$ depends on a subset of at most $k$ variables. The main contribution is an algorithm for finding relevant coordinates in a $k$-junta distribution with subcube conditioning (Bhattacharyya et al 2018., Canonne et al. 2019). We give two applications: An algorithm for learning $k$-junta distributions with $\tilde{O}(k/\epsilon^2) \log n + O(2^k/\epsilon^2)$ subcube conditioning queries, and an algorithm for testing $k$-junta distributions with $\tilde{O}((k + \sqrt{n})/\epsilon^2)$ subcube conditioning queries. All our algorithms are optimal up to poly-logarithmic factors. Our results show that subcube conditioning, as a natural model for accessing high-dimensional distributions, enables significant savings in learning and testing junta distributions compared to the standard sampling model. This addresses an open question posed by Aliakbarpour et al. 2016.