AISTATS 2020 Papers

In this paper, we analyze a Nyström based approach to efficient large scale kernel principal component analysis (PCA). The latter is a natural nonlinear extension of classical PCA based on considering a nonlinear feature map or the corresponding kernel. Like other kernel approaches, kernel PCA enjoys good mathematical and statistical properties but, numerically, it scales poorly with the sample size. Our analysis shows that Nyström sampling greatly improves computational efficiency without incurring any loss of statistical accuracy. While similar effects have been observed in supervised learning, this is the first such result for PCA. Our theoretical findings are based on a combination of analytic and concentration of measure techniques. Our study is more broadly motivated by the question of understanding the interplay between statistical and computational requirements for learning.

Multivariate time series with missing values are common in areas such as healthcare and finance, and have grown in number and complexity over the years. This raises the question whether deep learning methodologies can outperform classical data imputation methods in this domain. However, naive applications of deep learning fall short in giving reliable confidence estimates and lack interpretability.We propose a new deep sequential latent variable model for dimensionality reduction and data imputation. Our modeling assumption is simple and interpretable: the high dimensional time series has a lower-dimensional representation which evolves smoothly in time according to a Gaussian process. The non-linear dimensionality reduction in the presence of missing data is achieved using a VAE approach with a novel structured variational approximation. We demonstrate that our approach outperforms both classical and recent deep learning-based data imputation methods on high dimensional data from the domains of computer vision and healthcare.

We advocate the use of a notion of entropy that reflects the relative abundances of the symbols in an alphabet, as well as the similarities between them. This concept was originally introduced in theoretical ecology to study the diversity of ecosystems. Based on this notion of entropy, we introduce geometry-aware counterparts for several concepts and theorems in information theory. Notably, our proposed divergence exhibits performance on par with state-of-the-art methods based on the Wasserstein distance, but enjoys a closed-form expression that can be computed efficiently. We demonstrate the versatility of our method via experiments on a broad range of domains: training generative models, computing image barycenters, approximating empirical measures and counting modes.

Recently, several studies have proposed progressive or sequential layer-wise training methods based on the boosting theory for deep neural networks. However, most studies lack the global convergence guarantees or require weak learning conditions that can be verified a posteriori after running methods. Moreover, generalization bounds usually have a worse dependence on network depth. In this paper, to resolve these problems, we propose a new functional gradient boosting for learning deep residual-like networks in a layer-wise fashion with its statistical guarantees on multi-class classification tasks. In the proposed method, each residual block is recognized as a functional gradient (i.e., weak learner), and the functional gradient step is performed by stacking it on the network, resulting in a strong optimization ability. In the theoretical analysis, we show the global convergence of the method under a standard margin assumption on a data distribution instead of a weak learning condition, and we eliminate a worse dependence on the network depth in a generalization bound via a fine-grained convergence analysis. %, unlike existing studies. Moreover, we show that the existence of a learnable function with a large margin on a training dataset significantly improves a generalization bound. Finally, we experimentally demonstrate that our proposed method is certainly useful for learning deep residual networks.

We consider the fully decentralized machine learning scenario where many users with personal datasets collaborate to learn models through local peer-to-peer exchanges, without a central coordinator. We propose to train personalized models that leverage a collaboration graph describing the relationships between user personal tasks, which we learn jointly with the models. Our fully decentralized optimization procedure alternates between training nonlinear models given the graph in a greedy boosting manner, and updating the collaboration graph (with controlled sparsity) given the models. Throughout the process, users exchange messages only with a small number of peers (their direct neighbors when updating the models, and a few random users when updating the graph), ensuring that the procedure naturally scales with the number of users. Overall, our approach is communication-efficient and avoids exchanging personal data. We provide an extensive analysis of the convergence rate, memory and communication complexity of our approach, and demonstrate its benefits compared to competing techniques on synthetic and real datasets.

We consider the exploration-exploitation dilemma in finite-horizon reinforcement learning (RL). When the state space is large or continuous, traditional tabular approaches are unfeasible and some form of function approximation is mandatory. In this paper, we introduce an optimistically-initialized variant of the popular randomized least-squares value iteration (RLSVI), a model-free algorithm where exploration is induced by perturbing the least-squares approximation of the action-value function. Under the assumption that the Markov decision process has low-rank transition dynamics, we prove that the frequentist regret of RLSVI is upper-bounded by $\widetilde O(d^2 H^2 \sqrt{T})$ where $ d $ are the feature dimension, $ H $ is the horizon, and $ T $ is the total number of steps. To the best of our knowledge, this is the first frequentist regret analysis for randomized exploration with function approximation.

Measuring mutual information from finite data is difficult. Recent work has considered variational methods maximizing a lower bound. In this paper, we prove that serious statistical limitations are inherent to any method of measuring mutual information. More specifically, we show that any distribution-free high-confidence lower bound on mutual information estimated from N samples cannot be larger than O(ln N).

Conformal methods create prediction bands that control average coverage assuming solely i.i.d. data. Besides average coverage, one might also desire to control conditional coverage, that is, coverage for every new testing point. However, without strong assumptions, conditional coverage is unachievable. Given this limitation, the literature has focused on methods with asymptotical conditional coverage. In order to obtain this property, these methods require strong conditions on the dependence between the target variable and the features. We introduce two conformal methods based on conditional density estimators that do not depend on this type of assumption to obtain asymptotic conditional coverage: Dist-split and CD-split. While Dist-split asymptotically obtains optimal intervals, which are easier to interpret than general regions, CD-split obtains optimal size regions, which are smaller than intervals. CD-split also obtains local coverage by creating prediction bands locally on a partition of the features space. This partition is data-driven and scales to high-dimensional settings. In a wide variety of simulated scenarios, our methods have a better control of conditional coverage and have smaller length than previously proposed methods.

We investigate and provide new insights on the sampling rule called Top-Two Thompson Sampling (TTTS). In particular, we justify its use for fixed-confidence best-arm identification. We further propose a variant of TTTS called Top-Two Transportation Cost (T3C), which disposes of the computational burden of TTTS. As our main contribution, we provide the first sample complexity analysis of TTTS and T3C when coupled with a very natural Bayesian stopping rule, for bandits with Gaussian rewards, solving one of the open questions raised by Russo (2016). We also provide new posterior convergence results for TTTS under two models that are commonly used in practice: bandits with Gaussian and Bernoulli rewards and conjugate priors.

Inspired by the widespread use of Q-learning algorithms in reinforcement learning (RL), this present paper studies a class of biased stochastic approximation (SA) procedures under an ‘ergodic-like’ assumption on the underlying stochastic noise sequence. Leveraging a \emph{multistep Lyapunov function} that looks ahead to several future updates to accommodate the gradient bias, we prove a general result on the convergence of the iterates, and use it to derive finite-time bounds on the mean-square error in the case of constant stepsizes. This novel viewpoint renders the finite-time analysis of \emph{biased SA} algorithms under a broad family of stochastic perturbations possible. For direct comparison with past works, we also demonstrate these bounds by applying them to Q-learning with linear function approximation, under the realistic Markov chain observation model. The resultant finite-time error bound for Q-learning is \emph{the first of its kind}, in the sense that it holds: i) for the unmodified version (i.e., without making any modifications to the updates), and ii), for Markov chains starting from any initial distribution, at least one of which has to be violated for existing results to be applicable.

Motivated by the emerging use of multi-agent reinforcement learning (MARL) in engineering applications such as networked robotics, swarming drones, and sensor networks, we investigate the policy evaluation problem in a fully decentralized setting, using temporal-difference (TD) learning with linear function approximation to handle large state spaces in practice. The goal of the group of agents is to collaboratively learn the value function of a given policy from locally private rewards observed in a shared environment, through exchanging local estimates with neighbors. Despite their simplicity and widespread use, our theoretical understanding of such decentralized TD learning algorithms remains limited. Existing results were obtained based on i.i.d. data samples, or by imposing an ‘additional’ projection step to control the ‘gradient’ bias incurred by the Markovian observations. In this paper, we provide a finite-time analysis of the fully decentralized TD(0) learning under both i.i.d. as well as Markovian samples, and prove that all local estimates converge linearly to a small neighborhood of the optimum. The resultant error bounds are the first of its type—in the sense that they hold under the most practical assumptions—which is made possible by means of a novel multi-step Lyapunov approach.

Despite the recent successes of deep neural networks, the corresponding training problem remains highly non-convex and difficult to optimize. Classes of models have been proposed that introduce greater structure to the objective function at the cost of lifting the dimension of the problem. However, these lifted methods sometimes perform poorly compared to traditional neural networks. In this paper, we introduce a new class of lifted models, Fenchel lifted networks, that enjoy the same benefits as previous lifted models, without suffering a degradation in performance over classical networks. Our model represents activation functions as equivalent biconvex constraints and uses Lagrange Multipliers to arrive at a rigorous lower bound of the traditional neural network training problem. This model is efficiently trained using block-coordinate descent and is parallelizable across data points and/or layers. We compare our model against standard fully connected and convolutional networks and show that we are able to match or beat their performance.

The discovery of heavy hitters (most frequent items) in user-generated data streams drives improvements in the app and web ecosystems, but can incur substantial privacy risks if not done with care. To address these risks, we propose a distributed and privacy-preserving algorithm for discovering the heavy hitters in a population of user-generated data streams. We leverage the sampling and thresholding properties of our distributed algorithm to prove that it is inherently differentially private, without requiring additional noise. We also examine the trade-off between privacy and utility, and show that our algorithm provides excellent utility while also achieving strong privacy guarantees. A significant advantage of this approach is that it eliminates the need to centralize raw data while also avoiding the significant loss in utility incurred by local differential privacy. We validate our findings both theoretically, using worst-case analyses, and practically, using a Twitter dataset with 1.6M tweets and over 650k users. Finally, we carefully compare our approach to Apple’s local differential privacy method for discovering heavy hitters.

Federated learning is a distributed framework according to which a model is trained over a set of devices, while keeping data localized. This framework faces several systems-oriented challenges which include (i) communication bottleneck since a large number of devices upload their local updates to a parameter server, and (ii) scalability as the federated network consists of millions of devices. Due to these systems challenges as well as issues related to statistical heterogeneity of data and privacy concerns, designing a provably efficient federated learning method is of significant importance yet it remains challenging. In this paper, we present FedPAQ, a communication-efficient Federated Learning method with Periodic Averaging and Quantization. FedPAQ relies on three key features: (1) periodic averaging where models are updated locally at devices and only periodically averaged at the server; (2) partial device participation where only a fraction of devices participate in each round of the training; and (3) quantized message-passing where the edge nodes quantize their updates before uploading to the parameter server. These features address the communications and scalability challenges in federated learning. We also show that FedPAQ achieves near-optimal theoretical guarantees for strongly convex and non-convex loss functions and empirically demonstrate the communication-computation tradeoff provided by our method.

We discuss promising recent contributions on quantifying feature relevance using Shapley values, where we observed some confusion on which probability distribution is the right one for dropped features. We argue that the confusion is based on not carefully distinguishing between observational and interventional conditional probabilities and try a clarification based on Pearl’s seminal work on causality. We conclude that unconditional rather than conditional expectations provide the right notion of dropping features. This contradicts the view of the authors of the software package SHAP. In that work, unconditional expectations (which we argue to be conceptually right) are only used as approximation for the conditional ones, which encouraged others to ’improve’ SHAP in a way that we believe to be flawed.

We consider stochastic second-order methods for minimizing smooth and strongly-convex functions under an interpolation condition satisfied by over-parameterized models. Under this condition, we show that the regularized subsampled Newton method (R-SSN) achieves global linear convergence with an adaptive step-size and a constant batch-size. By growing the batch size for both the subsampled gradient and Hessian, we show that R-SSN can converge at a quadratic rate in a local neighbourhood of the solution. We also show that R-SSN attains local linear convergence for the family of self-concordant functions. Furthermore, we analyze stochastic BFGS algorithms in the interpolation setting and prove their global linear convergence. We empirically evaluate stochastic L-BFGS and a "Hessian-free" implementation of R-SSN for binary classification on synthetic, linearly-separable datasets and real datasets under a kernel mapping. Our experimental results demonstrate the fast convergence of these methods, both in terms of the number of iterations and wall-clock time.

Research on nearest-neighbor methods tends to focus somewhat dichotomously either on the statistical or the computational aspects – either on, say, Bayes consistency and rates of convergence or on techniques for speeding up the proximity search. This paper aims at bridging these realms: to reap the advantages of fast evaluation time while maintaining Bayes consistency, and further without sacrificing too much in the risk decay rate. We combine the locality-sensitive hashing (LSH) technique with a novel missing-mass argument to obtain a fast and Bayes-consistent classifier. Our algorithm’s prediction runtime compares favorably against state of the art approximate NN methods, while maintaining Bayes-consistency and attaining rates comparable to minimax. On samples of size $n$ in $\R^d$, our pre-processing phase has runtime $O(d n \log n)$, while the evaluation phase has runtime $O(d\log n)$ per query point.

We consider a ranking regression problem in which we use a dataset of ranked choices to learn Plackett-Luce scores as functions of sample features. We solve the maximum likelihood estimation problem by using the Alternating Directions Method of Multipliers (ADMM), effectively separating the learning of scores and model parameters. This separation allows us to express scores as the stationary distribution of a continuous-time Markov Chain. Using this equivalence, we propose two spectral algorithms for ranking regression that learn model parameters up to 579 times faster than the Newton’s method.

This paper considers k-means clustering in the presence of noise. It is known that k-means clustering is highly sensitive to noise, and thus noise should be removed to obtain a quality solution. A popular formulation of this problem is called k-means clustering with outliers. The goal of k-means clustering with outliers is to discard up to a specified number z of points as noise/outliers and then find a k-means solution on the remaining data. The problem has received significant attention, yet current algorithms with theoretical guarantees suffer from either high running time or inherent loss in the solution quality. The main contribution of this paper is two-fold. Firstly, we develop a simple greedy algorithm that has provably strong worst case guarantees. The greedy algorithm adds a simple preprocessing step to remove noise, which can be combined with any k-means clustering algorithm. This algorithm gives the first pseudo-approximation-preserving reduction from k-means with outliers to k-means without outliers. Secondly, we show how to construct a coreset of size O(k log n). When combined with our greedy algorithm, we obtain a scalable, near linear time algorithm. The theoretical contributions are verified experimentally by demonstrating that the algorithm quickly removes noise and obtains a high-quality clustering.