COLT 2021 Papers

We give a novel, unified derivation of conditional PAC-Bayesian and mutual information (MI) generalization bounds. We derive conditional MI bounds as an instance, with special choice of prior, of conditional MAC-Bayesian (Mean Approximately Correct) bounds, itself derived from conditional PAC-Bayesian bounds, where ‘conditional’ means that one can use priors conditioned on a joint training and ghost sample. This allows us to get nontrivial PAC-Bayes and MI-style bounds for general VC classes, something recently shown to be impossible with standard PAC-Bayesian/MI bounds. Second, it allows us to get fast rates of order $O((\text{KL}/n)^{\gamma}$ for $\gamma > 1/2$ if a Bernstein condition holds and for exp-concave losses (with $\gamma=1$), which is impossible with both standard PAC-Bayes generalization and MI bounds. Our work extends the recent work by Steinke and Zakynthinou (2020) who handle MI with VC but neither PAC-Bayes nor fast rates and Mhammedi et al. (2019) who initiated fast rate PAC-Bayes generalization error bounds but handle neither MI nor general VC classes.

We study the problem of learning Ising models satisfying Dobrushin’s condition in the outlier-robust setting where a constant fraction of the samples are adversarially corrupted. Our main result is to provide the first computationally efficient robust learning algorithm for this problem with near-optimal error guarantees. Our algorithm can be seen as a special case of an algorithm for robustly learning a distribution from a general exponential family. To prove its correctness for Ising models, we establish new anti-concentration results for degree-2 polynomials of Ising models that may be of independent interest.

This paper treats the task of designing optimization algorithms as an optimal control problem. Using regret as a metric for an algorithm’s performance, we study the existence, uniqueness and consistency of regret-optimal algorithms. By providing first-order optimality conditions for the control problem, we show that regret-optimal algorithms must satisfy a specific structure in their dynamics which we show is equivalent to performing \emph{dual-preconditioned gradient descent} on the value function generated by its regret. Using these optimal dynamics, we provide bounds on their rates of convergence to solutions of convex optimization problems. Though closed-form optimal dynamics cannot be obtained in general, we present fast numerical methods for approximating them, generating optimization algorithms which directly optimize their long-term regret. These are benchmarked against commonly used optimization algorithms to demonstrate their effectiveness.

Conventional wisdom in the sampling literature, backed by a popular diffusion scaling limit, suggests that the mixing time of the Metropolis-Adjusted Langevin Algorithm (MALA) scales as O(d^{1/3}), where d is the dimension. However, the diffusion scaling limit requires stringent assumptions on the target distribution and is asymptotic in nature. In contrast, the best known non-asymptotic mixing time bound for MALA on the class of log-smooth and strongly log-concave distributions is O(d). In this work, we establish that the mixing time of MALA on this class of target distributions is \tilde\Theta(d^{1/2}) under a warm start. Our upper bound proof introduces a new technique based on a projection characterization of the Metropolis adjustment which reduces the study of MALA to the well-studied discretization analysis of the Langevin SDE and bypasses direct computation of the acceptance probability.

We consider the problem of the Zinkevich (2003)-style dynamic regret minimization in online learning with \emph{exp-concave} losses. We show that whenever improper learning is allowed, a Strongly Adaptive online learner achieves the dynamic regret of $\tilde O^*(n^{1/3}C_n^{2/3} \vee 1)$ where $C_n$ is the \emph{total variation} (a.k.a. \emph{path length}) of the an arbitrary sequence of comparators that may not be known to the learner ahead of time. Achieving this rate was highly nontrivial even for square losses in 1D where the best known upper bound was $O(\sqrt{nC_n} \vee \log n)$ (Yuan and Lamperski, 2019). Our new proof techniques make elegant use of the intricate structures of the primal and dual variables imposed by the KKT conditions and could be of independent interest. Finally, we apply our results to the classical statistical problem of \emph{locally adaptive non-parametric regression} (Mammen, 1991; Donoho and Johnstone, 1998) and obtain a stronger and more flexible algorithm that do not require any statistical assumptions or any hyperparameter tuning.

Confidence intervals are a crucial building block in the analysis of various online learning problems. The analysis of kernel-based bandit and reinforcement learning problems utilize confidence intervals applicable to the elements of a reproducing kernel Hilbert space (RKHS). However, the existing confidence bounds do not appear to be tight, resulting in suboptimal regret bounds. In fact, the existing regret bounds for several kernelized bandit algorithms (e.g., GP-UCB, GP-TS, and their variants) may fail to even be sublinear. It is unclear whether the suboptimal regret bound is a fundamental shortcoming of these algorithms or an artifact of the proof, and the main challenge seems to stem from the online (sequential) nature of the observation points. We formalize the question of online confidence intervals in the RKHS setting and overview the existing results.

The open problem asks whether there exists an online learning algorithm for binary classification that guarantees, for all target concepts, to make a sublinear number of mistakes, under only the assumption that the (possibly random) sequence of points X allows that such a learning algorithm can exist for that sequence. As a secondary problem, it also asks whether a specific concise condition completely determines whether a given (possibly random) sequence of points X admits the existence of online learning algorithms guaranteeing a sublinear number of mistakes for all target concepts.

We propose matrix norm inequalities that extend the Recht and Ré (2012) conjecture on a noncommutative AM-GM inequality, by supplementing it with another inequality that accounts for single-shuffle in stochastic finite-sum minimization. Single-shuffle is a popular without-replacement sampling scheme that shuffles only once in the beginning, but has not been studied in the Recht-Ré conjecture and the follow-up literature. Instead of focusing on general positive semidefinite matrices, we restrict our attention to positive definite matrices with small enough condition numbers, which are more relevant to matrices that arise in the analysis of SGD. For such matrices, we conjecture that the means of matrix products satisfy a series of spectral norm inequalities that imply “single-shuffle SGD converges faster than random-reshuffle SGD, which is in turn faster than with-replacement SGD and GD” in special cases.

A few years ago, it was shown that there exist basic statistical learning problems whose learnability can not be determined within ZFC [Ben-David, Hrubes, Moran, Shpilka, Yehudayoff, 2017]. Such independence, and the implied impossibility of characterizing learnability of a class by any combinatorial parameter, stems from the basic definitions viewing learners as arbitrary functions. That level of generality not only results in unprovability issues but is also problematic from the perspective of modeling practical machine learning, where learners and predictors are computable objects. In light of that, it is natural to consider learnability by algorithms that output computable predictors (both learners and predictors are then representable as finite objects). A recent study [Agarwal, Ananthakrishnan, Ben-David, Lechner and Urner, 2020] initiated a theory of such models of learning. It proposed the notion of CPAC learnability, by adding some basic computability requirements into a PAC learning framework. As a first step towards a characterization of learnability in the CPAC framework, Agarwal et al showed that CPAC learnability of a binary hypothesis class is not implied by the finiteness of its VC-dimension anymore, as far as proper learners are concerned. A major remaining open question is whether a similar result holds also for improper learning. Namely, does there exist a computable concept class consisting of computable classifiers, that has a finite VC-dimension but no computable learner can PAC learn it (even if the learner is not restricted to output a hypothesis that is a member of the class)? Another implied interesting question concerns coming up with combinatorial characterizations of learnability for computable learners.

We study a novel variant of online finite-horizon Markov Decision Processes with adversarially changing loss functions and initially unknown dynamics. In each episode, the learner suffers the loss accumulated along the trajectory realized by the policy chosen for the episode, and observes aggregate bandit feedback: the trajectory is revealed along with the cumulative loss suffered, rather than the individual losses encountered along the trajectory. Our main result is a computationally efficient algorithm with $O(\sqrt{K})$ regret for this setting, where K is the number of episodes. We establish this result via an efficient reduction to a novel bandit learning setting we call Distorted Linear Bandits (DLB), which is a variant of bandit linear optimization where actions chosen by the learner are adversarially distorted before they are committed. We then develop a computationally-efficient online algorithm for DLB for which we prove an $O(\sqrt{T})$ regret bound, where T is the number of time steps. Our algorithm is based on online mirror descent with a self-concordant barrier regularization that employs a novel increasing learning rate schedule.

Which classes can be learned properly in the online model? — that is, by an algorithm that on each round uses a predictor from the concept class. While there are simple and natural cases where improper learning is useful and even necessary, it is natural to ask how complex must the improper predictors be in such cases. Can one always achieve nearly optimal mistake/regret bounds using "simple" predictors? In this work, we give a complete characterization of when this is possible, thus settling an open problem which has been studied since the pioneering works of Angluin (1987) and Littlestone (1988). More precisely, given any concept class C and any hypothesis class H, we provide nearly tight bounds (up to a log factor) on the optimal mistake bounds for online learning C using predictors from H. Our bound yields an exponential improvement over the previously best known bound by Chase and Freitag (2020). As applications, we give constructive proofs showing that (i) in the realizable setting, a near-optimal mistake bound (up to a constant factor) can be attained by a sparse majority-vote of proper predictors, and (ii) in the agnostic setting, a near-optimal regret bound (up to a log factor) can be attained by a randomized proper algorithm. The latter was proven non-constructively by Rakhlin, Sridharan, and Tewari (2015). It was also achieved by constructive but improper algorithms proposed by Ben-David, Pal, and Shalev-Shwartz (2009) and Rakhlin, Shamir, and Sridharan (2012). A technical ingredient of our proof which may be of independent interest is a generalization of the celebrated Minimax Theorem (von Neumann, 1928) for binary zero-sum games with arbitrary action-sets: a simple game which fails to satisfy Minimax is "Guess the Larger Number". In this game, each player picks a natural number and the player who picked the larger number wins. Equivalently, the payoff matrix of this game is infinite triangular. We show that this is the only obstruction: if the payoff matrix does not contain triangular submatrices of unbounded sizes then the Minimax theorem is satisfied. This generalizes von Neumann’s Minimax Theorem by removing requirements of finiteness (or compactness) of the action-sets, and moreover it captures precisely the types of games of interest in online learning: namely, Littlestone games.

We study the problem of online contextual optimization where, at each period, instead of observing the loss, we observe, after-the-fact, the optimal action an oracle with full knowledge of the objective function would have taken. At each period, the decision-maker has access to a new set of feasible actions to select from and to a new contextual function that affects that period’s loss function. We aim to minimize regret, which is defined as the difference between our losses and the ones incurred by an all-knowing oracle. We obtain the first regret bound for this problem that is logarithmic in the time horizon. Our results are derived through the development and analysis of a novel algorithmic structure that leverages the underlying geometry of the problem.

This paper studies the exponential stability of random matrix products driven by a general (possibly unbounded) state space Markov chain. It is a cornerstone in the analysis of stochastic algorithms in machine learning (e.g. for parameter tracking in online-learning or reinforcement learning). The existing results impose strong conditions such as uniform boundedness of the matrix-valued functions and uniform ergodicity of the Markov chains. Our main contribution is an exponential stability result for the p-th moment of random matrix product, provided that (i) the underlying Markov chain satisfies a super-Lyapunov drift condition, (ii) the growth of the matrix-valued functions is controlled by an appropriately defined function (related to the drift condition). Using this result, we give finite-time p-th moment bounds for constant and decreasing stepsize linear stochastic approximation schemes with Markovian noise on general state space. We illustrate these findings for linear value-function estimation in reinforcement learning. We provide finite-time p-th moment bound for various members of temporal difference (TD) family of algorithms.

We study the minimal squared error of the Empirical Risk Minimization (ERM) procedure in the task of regression, both in random and fixed design settings. Our sharp lower bounds shed light on the possibility (or impossibility) of adapting to simplicity of the model generating the data. In the fixed design setting, we show that the error is governed by the global complexity of the entire class. In contrast, in random design, ERM may only adapt to simpler models if the local neighborhoods around the regression function are nearly as complex as the class itself, a somewhat counter-intuitive conclusion. We provide sharp lower bounds for performance of ERM in Donsker and non-Donsker classes. We also discuss our results through the lens of recent studies on interpolation in overparameterized models.

We study sampling from a target distribution $\nu_* = e^{-f}$ using the unadjusted Langevin Monte Carlo (LMC) algorithm. For any potential function $f$ whose tails behave like $\|x\|^\alpha$ for ${\alpha \in [1,2]}$, and has $\beta$-Hölder continuous gradient, we prove that $\widetilde{\mathcal{O}} \Big(d^{\frac{1}{\beta}+\frac{1+\beta}{\beta}(\frac{2}{\alpha}-{1}_{\{\alpha \neq 1\}})} \epsilon^{-\frac{1}{\beta}}\Big)$ steps are sufficient to reach the $\epsilon$-neighborhood of a $d$-dimensional target distribution $\nu_*$ in KL-divergence. This bound, in terms of $\epsilon$ dependency, is not directly influenced by the tail growth rate $\alpha$ of the potential function as long as its growth is at least linear, and it only relies on the order of smoothness $\beta$. One notable consequence of this result is that for potentials with Lipschitz gradient, i.e. $\beta=1$, the above rate recovers the best known rate $\widetilde{\mathcal{O}} (d\epsilon^{-1})$ which was established for strongly convex potentials in terms of $\epsilon$ dependency, but we show that the same rate is achievable for a wider class of potentials that are degenerately convex at infinity. The growth rate $\alpha$ affects the rate estimate in high dimensions where $d$ is large; furthermore, it recovers the best-known dimension dependency when the tail growth of the potential is quadratic, i.e. $\alpha = 2$, in the current setup.

This paper considers the following question: how well can depth-two ReLU networks with randomly initialized bottom-level weights represent smooth functions? We give near-matching upper- and lower-bounds for L2-approximation in terms of the Lipschitz constant, the desired accuracy, and the dimension of the problem, as well as similar results in terms of Sobolev norms. Our positive results employ tools from harmonic analysis and ridgelet representation theory, while our lower-bounds are based on (robust versions of) dimensionality arguments.

We consider the problem of local planning in fixed-horizon Markov Decision Processes (MDPs) with a generative model under the assumption that the optimal value function lies close to the span of a feature map. The generative model provides a restricted, “local” access to the MDP: The planner can ask for random transitions from previously returned states and arbitrary actions, and the features are also only accessible for the states that are encountered in this process. As opposed to previous work (e.g. Lattimore et al. (2020)) where linear realizability of all policies was assumed, we consider the significantly relaxed assumption of a single linearly realizable (deterministic) policy. A recent lower bound by Weisz et al. (2020) established that the related problem when the action-value function of the optimal policy is linearly realizable requires an exponential number of queries, either in $H$ (the horizon of the MDP) or $d$ (the dimension of the feature mapping). Their construction crucially relies on having an exponentially large action set. In contrast, in this work, we establish that $\poly(H,d)$ planning is possible with state value function realizability whenever the action set has a constant size. In particular, we present the TensorPlan algorithm which uses $\poly((dH/\delta)^A)$ simulator queries to find a $\delta$-optimal policy relative to any deterministic policy for which the value function is linearly realizable with some bounded parameter (with a known bound). This is the first algorithm to give a polynomial query complexity guarantee using only linear-realizability of a single competing value function. Whether the computation cost is similarly bounded remains an interesting open question. We also extend the upper bound to the near-realizable case and to the infinite-horizon discounted MDP setup. The upper bounds are complemented by a lower bound which states that in the infinite-horizon episodic setting, planners that achieve constant suboptimality need exponentially many queries, either in the dimension or the number of actions.

In this paper, we consider variational autoencoders (VAE) via empirical Bayes estimation, referred to as Empirical Bayes Variational Autoencoders (EBVAE), which is a general framework including popular VAE methods as special cases. Despite the widespread use of VAE, its theoretical aspects are less explored in the literature. Motivated by this, we establish a general theoretical framework for analyzing the excess risk associated with EBVAE under the setting of density estimation, covering both parametric and nonparametric cases, through the lens of M-estimation. As an application, we analyze the excess risk of the commonly-used EBVAE with Gaussian models and highlight the importance of covariance matrices of Gaussian encoders and decoders in obtaining a good statistical guarantee, shedding light on the empirical observations reported in the literature.

In this work, we study the problem of answering $k$ queries with $(\epsilon, \delta)$-differential privacy, where each query has sensitivity one. We give an algorithm for this task that achieves an expected $\ell_\infty$ error bound of $O(\frac{1}{\epsilon}\sqrt{k \log \frac{1}{\delta}})$, which is known to be tight (Steinke and Ullman, 2016). A very recent work by Dagan and Kur (2020) provides a similar result, albeit via a completely different approach. One difference between our work and theirs is that our guarantee holds even when $\delta < 2^{-\Omega(k/(\log k)^8)}$ whereas theirs does not apply in this case. On the other hand, the algorithm of Dagan and Kur (2020) has a remarkable advantage that the $\ell_{\infty}$ error bound of $O(\frac{1}{\epsilon}\sqrt{k \log \frac{1}{\delta}})$ holds not only in expectation but always (i.e., with probability one) while we can only get a high probability (or expected) guarantee on the error.