AISTATS 2023 Papers

In continuum-armed bandit problems where the underlying function resides in a reproducing kernel Hilbert space (RKHS), namely, the kernelised bandit problems, an important open problem remains of how well learning algorithms can adapt if the regularity of the associated kernel function is unknown. In this work, we study adaptivity to the regularity of translation-invariant kernels, which is characterized by the decay rate of the Fourier transformation of the kernel, in the bandit setting. We derive an adaptivity lower bound, proving that it is impossible to simultaneously achieve optimal cumulative regret in a pair of RKHSs with different regularities. To verify the tightness of this lower bound, we show that an existing bandit model selection algorithm applied with minimax non-adaptive kernelised bandit algorithms matches the lower bound in dependence of T, the total number of steps, except for log factors. By filling in the regret bounds for adaptivity between RKHSs, we connect the statistical difficulty for adaptivity in continuum-armed bandits in three fundamental types of function spaces: RKHS, Sobolev space, and Holder space.

In the paper, we propose a class of faster adaptive Gradient Descent Ascent (GDA) methods for solving the nonconvex-strongly-concave minimax problems by using the unified adaptive matrices, which include almost all existing coordinate-wise and global adaptive learning rates. In particular, we provide an effective convergence analysis framework for our adaptive GDA methods. Specifically, we propose a fast Adaptive Gradient Descent Ascent (AdaGDA) method based on the basic momentum technique, which reaches a lower gradient complexity of $\tilde{O}(\kappa^4\epsilon^{-4})$ for finding an $\epsilon$-stationary point without large batches, which improves the existing results of the adaptive GDA methods by a factor of $O(\sqrt{\kappa})$. Moreover, we propose an accelerated version of AdaGDA (VR-AdaGDA) method based on the momentum-based variance reduced technique, which achieves a lower gradient complexity of $\tilde{O}(\kappa^{4.5}\epsilon^{-3})$ for finding an $\epsilon$-stationary point without large batches, which improves the existing results of the adaptive GDA methods by a factor of $O(\epsilon^{-1})$. Moreover, we prove that our VR-AdaGDA method can reach the best known gradient complexity of $\tilde{O}(\kappa^{3}\epsilon^{-3})$ with the mini-batch size $O(\kappa^3)$. The experiments on policy evaluation and fair classifier learning tasks are conducted to verify the efficiency of our new algorithms.

Gaussian processes (GPs) are typically criticised for their unfavourable scaling in both computational and memory requirements. For large datasets, sparse GPs reduce these demands by conditioning on a small set of inducing variables designed to summarise the data. In practice however, for large datasets requiring many inducing variables, such as low-lengthscale spatial data, even sparse GPs can become computationally expensive, limited by the number of inducing variables one can use. In this work, we propose a new class of inter-domain variational GP, constructed by projecting a GP onto a set of compactly supported B-spline basis functions. The key benefit of our approach is that the compact support of the B-spline basis functions admits the use of sparse linear algebra to significantly speed up matrix operations and drastically reduce the memory footprint. This allows us to very efficiently model fast-varying spatial phenomena with tens of thousands of inducing variables, where previous approaches failed.

Federated learning (FL) was originally regarded as a framework for collaborative learning among clients with data privacy protection through a coordinating server. In this paper, we propose a new active membership inference (AMI) attack carried out by a dishonest server in FL. In AMI attacks, the server crafts and embeds malicious parameters into global models to effectively infer whether a target data sample is included in a client’s private training data or not. By exploiting the correlation among data features through a non-linear decision boundary, AMI attacks with a certified guarantee of success can achieve severely high success rates under rigorous local differential privacy (LDP) protection; thereby exposing clients’ training data to significant privacy risk. Theoretical and experimental results on several benchmark datasets show that adding sufficient privacy-preserving noise to prevent our attack would significantly damage FL’s model utility.

We consider the problem of active learning for single neuron models, also sometimes called “ridge functions”, in the agnostic setting (under adversarial label noise). Such models have been shown to be broadly effective in modeling physical phenomena, and for constructing surrogate data-driven models for partial differential equations. Surprisingly, we show that for a single neuron model with any Lipschitz non-linearity (such as the ReLU, sigmoid, absolute value, low-degree polynomial, among others), strong provable approximation guarantees can be obtained using a well-known active learning strategy for fitting linear functions in the agnostic setting. Namely, we can collect samples via statistical leverage score sampling, which has been shown to be nearoptimal in other active learning scenarios. We support our theoretical results with empirical simulations showing that our proposed active learning strategy based on leverage score sampling outperforms (ordinary) uniform sampling when fitting single neuron models.

A key challenge in science and engineering is to design experiments to learn about some unknown quantity of interest. Classical experimental design optimally allocates the experimental budget into measurements to maximize a notion of utility (e.g., reduction in uncertainty about the unknown quantity). We consider a rich setting, where the experiments are associated with states in a Markov chain, and we can only choose them by selecting a policy controlling the state transitions. This problem captures important applications, from exploration in reinforcement learning to spatial monitoring tasks. We propose an algorithm – markov-design – that efficiently selects policies whose measurement allocation provably converges to the optimal one. The algorithm is sequential in nature, adapting its choice of policies (experiments) using past measurements. In addition to our theoretical analysis, we demonstrate our framework on applications in ecological surveillance and pharmacology.

We study actively labeling streaming data, where an active learner is faced with a stream of data points and must carefully choose which of these points to label via an expensive experiment. Such problems frequently arise in applications such as healthcare and astronomy. We first study a setting when the data’s inputs belong to one of K discrete distributions and formalize this problem via a loss that captures the labeling cost and the prediction error. When the labeling cost is B, our algorithm, which chooses to label a point if the uncertainty is larger than a time and cost dependent threshold, achieves a worst-case upper bound of $\tilde O(B^{\frac{1}{3}} K^{\frac{1}{3}} T^{\frac{2}{3}})$ on the loss after T rounds. We also provide a more nuanced upper bound which demonstrates that the algorithm can adapt to the arrival pattern, and achieves better performance when the arrival pattern is more favorable. We complement both upper bounds with matching lower bounds. We next study this problem when the inputs belong to a continuous domain and the output of the experiment is a smooth function with bounded RKHS norm. After T rounds in d dimensions, we show that the loss is bounded by $\tilde O(B^{\frac{1}{d+3}} T^{\frac{d+2}{d+3}})$ in an RKHS with a squared exponential kernel and by $\tilde O(B^{\frac{1}{2d+3}} T^{\frac{2d+2}{2d+3}})$ in an RKHS with a Matérn kernel. Our empirical evaluation demonstrates that our method outperforms other baselines in several synthetic experiments and two real experiments in medicine and astronomy.

Frank-Wolfe algorithms (FW) are popular first-order methods for solving constrained convex optimization problems that rely on a linear minimization oracle instead of potentially expensive projection-like oracles. Many works have identified accelerated convergence rates under various structural assumptions on the optimization problem and for specific FW variants when using line-search or short-step, requiring feedback from the objective function. Little is known about accelerated convergence regimes when utilizing open-loop step-size rules, a.k.a. FW with pre-determined step-sizes, which are algorithmically extremely simple and stable. Not only is FW with open-loop step-size rules not always subject to the same convergence rate lower bounds as FW with line-search or short-step, but in some specific cases, such as kernel herding in infinite dimensions, it has been empirically observed that FW with open-loop step-size rules leads to faster convergence than FW with line-search or short-step. We propose a partial answer to this unexplained phenomenon in kernel herding, characterize a general setting for which FW with open-loop step-size rules converges non-asymptotically faster than with line-search or short-step, and derive several accelerated convergence results for FW with open-loop step-size rules.

Selective classification (or classification with a reject option) pairs a classifier with a selection function to determine whether or not a prediction should be accepted. This framework trades off coverage (probability of accepting a prediction) with predictive performance, typically measured by distributive loss functions. In many application scenarios, such as credit scoring, performance is instead measured by ranking metrics, such as the Area Under the ROC Curve (AUC). We propose a model-agnostic approach to associate a selection function to a given probabilistic binary classifier. The approach is specifically targeted at optimizing the AUC. We provide both theoretical justifications and a novel algorithm, called AUCROSS, to achieve such a goal. Experiments show that our method succeeds in trading-off coverage for AUC, improving over existing selective classification methods targeted at optimizing accuracy.

In this paper, we develop a new algorithm, Annealed Skewed SGD - AskewSGD - for training deep neural networks (DNNs) with quantized weights. First, we formulate the training of quantized neural networks (QNNs) as a smoothed sequence of interval-constrained optimization problems. Then, we propose a new first-order stochastic method, AskewSGD, to solve each constrained optimization subproblem. Unlike algorithms with active sets and feasible directions, AskewSGD avoids projections or optimization under the entire feasible set and allows iterates that are infeasible. The numerical complexity of AskewSGD is comparable to existing approaches for training QNNs, such as the straight-through gradient estimator used in BinaryConnect, or other state of the art methods (ProxQuant, LUQ). We establish convergence guarantees for AskewSGD (under general assumptions for the objective function). Experimental results show that the AskewSGD algorithm performs better than or on par with state of the art methods in classical benchmarks.

We study the problem of non-stationary dueling bandits and provide the first adaptive dynamic regret algorithm for this problem. The only two existing attempts in this line of work fall short across multiple dimensions, including pessimistic measures of non-stationary complexity and non-adaptive parameter tuning that requires knowledge of the number of preference changes. We develop an elimination-based rescheduling algorithm to overcome these shortcomings and show a near-optimal $\tilde O(\sqrt{S^{CW} T})$ dynamic regret bound, where $\S^{CW}$ is the number of times the Condorcet winner changes in $T$ rounds. This yields the first near-optimal dynamic regret bound for unknown $S^{CW}$. We further study other related notions of non-stationarity for which we also prove near-optimal dynamic regret guarantees under additional assumptions on the preference model.

Bayesian optimization (BO) improves the efficiency of black-box optimization; however, the associated computational cost and power consumption remain dominant in the application of machine learning methods. This paper proposes a method of determining the stopping time in BO. The proposed criterion is based on the difference between the expectation of the minimum of a variant of the simple regrets before and after evaluating the objective function with a new parameter setting. Unlike existing stopping criteria, the proposed criterion is guaranteed to converge to the theoretically optimal stopping criterion for any choices of arbitrary acquisition functions and threshold values. Moreover, the threshold for the stopping criterion can be determined automatically and adaptively. We experimentally demonstrate that the proposed stopping criterion finds reasonable timing to stop a BO with a small number of evaluations of the objective function.

We consider a decentralized optimization problem, in which n nodes collaborate to optimize a global objective function using local communications only. While many decentralized algorithms focus on gossip communications (pairwise averaging), we consider a different scheme, in which a “token” that contains the current estimate of the model performs a random walk over the network, and updates its model using the local model of the node it is at. Indeed, token algorithms generally benefit from improved communication efficiency and privacy guarantees. We frame the token algorithm as a randomized gossip algorithm on a conceptual graph, which allows us to prove a series of convergence results for variance-reduced and accelerated token algorithms for the complete graph. We also extend these results to the case of multiple tokens by extending the conceptual graph, and to general graphs by tweaking the communication procedure. The reduction from token to well-studied gossip algorithms leads to tight rates for many token algorithms, and we illustrate their performance empirically.

This paper introduces a new proximal stochastic gradient method with variance reduction and stabilization for minimizing the sum of a convex stochastic function and a group sparsity-inducing regularization function. Since the method may be viewed as a stabilized version of the recently proposed algorithm PStorm, we call our algorithm S-PStorm. Our analysis shows that S-PStorm has strong convergence results. In particular, we prove an upper bound on the number of iterations required by S-PStorm before its iterates correctly identify (with high probability) an optimal support (i.e., the zero and nonzero structure of an optimal solution). Most algorithms in the literature with such a support identification property use variance reduction techniques that require either periodically evaluating an exact gradient or storing a history of stochastic gradients. Unlike these methods, S-PStorm achieves variance reduction without requiring either of these, which is advantageous. Moreover, our support-identification result for S-PStorm shows that, with high probability, an optimal support will be identified correctly in all iterations with index above a threshold. We believe that this type of result is new to the literature since the few existing other results prove that the optimal support is identified with high probability at each iteration with a sufficiently large index (meaning that the optimal support might be identified in some iterations, but not in others). Numerical experiments on regularized logistic loss problems show that S-PStorm outperforms existing methods in various metrics that measure how efficiently and robustly iterates of an algorithm identify an optimal support.

Subset selection, i.e., finding a bunch of items from a collection to achieve specific goals, has wide applications in information retrieval, statistics, and machine learning. To implement an end-to-end learning framework, different relaxed differentiable operators of subset selection are proposed. Most existing work relies on either the regularization method or the perturbation method. In this work, we provide a probabilistic interpretation for regularization relaxation and unify two schemes. Besides, we build some concrete examples to show the generic connection between these two relaxations. Finally, we evaluate the perturbed selector as well as the regularized selector on two tasks: the maximum entropy sampling problem and the feature selection problem. The experimental results show that these two methods can achieve competitive performance against other benchmarks.

We study the regret for risk-sensitive reinforcement learning (RL) with the exponential utility in the episodic MDP. Recent works establish both a lower bound $\Omega((e^{|\beta|(H-1)/2}-1)\sqrt{SAT}/|\beta|)$ and the best known (upper) bound $\tilde{O}((e^{|\beta|H}-1)\sqrt{H^2SAT}/|\beta|)$, where $H$ is the length of the episode, $S$ the size of state space, $A$ the size of action space, $T$ the total number of timesteps, and $\beta$ the risk parameter. The gap between the upper and the lower bound is exponential and hence is unsatisfactory. In this paper, we show that a variant of UCB-Advantage algorithm reduces a factor of $\sqrt{H}$ from the best previously known bound in any arbitrary MDP. To further sharpen the regret bound, we introduce a brand new mechanism of regret analysis and derive a problem-dependent regret bound without prior knowledge of the MDP from the algorithm. This bound is much tighter in MDPs with special structures. Particularly, we show that a regret that matches the information-theoretic lower bound up to logarithmic factors can be attained within a rich class of MDPs, which improves an exponential factor over the best previously known bound. Further, we derive a novel information-theoretic lower bound of $\Omega(\max_{h\in[H]} c_{v,h+1}^*\sqrt{SAT}/|\beta|)$, where $\max_{h\in[H]} c_{v,h+1}^*$ is a problem-dependent statistic. This lower bound shows that the problem-dependent regret bound achieved by the algorithm is optimal in its dependence on $\max_{h\in[H]} c_{v,h+1}^*$.

Variational Inference (VI) is an attractive alternative to Markov Chain Monte Carlo (MCMC) due to its computational efficiency in the case of large datasets and/or complex models with high-dimensional parameters. However, evaluating the accuracy of variational approximations remains a challenge. Existing methods characterize the quality of the whole variational distribution, which is almost always poor in realistic applications, even if specific posterior functionals such as the component-wise means or variances are accurate. Hence, these diagnostics are of practical value only in limited circumstances. To address this issue, we propose the“TArgeted Diagnostic for Distribution Approximation Accuracy” (TADDAA), which uses many short parallel MCMC chains to obtain lower bounds on the error of each posterior functional of interest. We also develop a reliability check for TADDAA to determine when the lower bounds should not be trusted. Numerical experiments validate the practical utility and computational efficiency of our approach on a range of synthetic distributions and real-data examples, including sparse logistic regression and Bayesian neural network models.

Markov Decision Process (MDP) with large action space naturally occurs in many applications such as language processing, information retrieval, and recommendation system. There have been various approaches to solve these MDPs through value iteration (VI). Unfortunately, all VI algorithms require expensive linear scans over the entire action space for value function estimation during each iteration. To this end, we present two provable Least-Squares Value Iteration (LSVI) algorithms with runtime complexity sublinear in the number of actions for linear MDPs. We formulate the value function estimation procedure in VI as an approximate maximum inner product search problem and propose a Locality Sensitive Hashing (LSH) type data structure to solve this problem with sublinear time complexity. Our major contribution is combining the guarantees of approximate maximum inner product search with the regret analysis of reinforcement learning. We prove that, with the appropriate choice of approximation factor, there exists a sweet spot. Our proposed Sublinear LSVI algorithms maintain the same regret as the original LSVI algorithms while reducing the runtime complexity to sublinear in the number of actions. To the best of our knowledge, this is the first work that combines LSH with reinforcement learning that results in provable improvements. We hope that our novel way of combining data structures and the iterative algorithm will open the door for further study into the cost reduction in reinforcement learning.

The most relevant problems in discounted reinforcement learning involve estimating the mean of a function under the stationary distribution of a Markov reward process, such as the expected return in policy evaluation, or the policy gradient in policy optimization. In practice, these estimates are produced through a finite-horizon episodic sampling, which neglects the mixing properties of the Markov process. It is mostly unclear how this mismatch between the practical and the ideal setting affects the estimation, and the literature lacks a formal study on the pitfalls of episodic sampling, and how to do it optimally. In this paper, we present a minimax lower bound on the discounted mean estimation problem that explicitly connects the estimation error with the mixing properties of the Markov process and the discount factor. Then, we provide a statistical analysis on a set of notable estimators and the corresponding sampling procedures, which includes the finite-horizon estimators often used in practice. Crucially, we show that estimating the mean by directly sampling from the discounted kernel of the Markov process brings compelling statistical properties w.r.t. the alternative estimators, as it matches the lower bound without requiring a careful tuning of the episode horizon.