AISTATS 2018 Papers

We address the problem of inferring the causal relation between two variables by comparing the least-squares errors of the predictions in both possible causal directions. Under the assumption of an independence between the function relating cause and effect, the conditional noise distribution, and the distribution of the cause, we show that the errors are smaller in causal direction if both variables are equally scaled and the causal relation is close to deterministic. Based on this, we provide an easily applicable method that only requires a regression in both possible causal directions. The performance of this method is compared with different related causal inference methods in various artificial and real-world data sets.

We introduce a generic scheme to solve nonconvex optimization problems using gradient-based algorithms originally designed for minimizing convex functions. Even though these methods may originally require convexity to operate, the proposed approach allows one to use them without assuming any knowledge about the convexity of the objective. In general, the scheme is guaranteed to produce a stationary point with a worst-case efficiency typical of first-order methods, and when the objective turns out to be convex, it automatically accelerates in the sense of Nesterov and achieves near-optimal convergence rate in function values. We conclude the paper by showing promising experimental results obtained by applying our approach to incremental algorithms such as SVRG and SAGA for sparse matrix factorization and for learning neural networks.

One iteration of standard k-means (i.e., Lloyd’s algorithm) or standard EM for Gaussian mixture models (GMMs) scales linearly with the number of clusters C, data points N, and data dimensionality D. In this study, we explore whether one iteration of k-means or EM for GMMs can scale sublinearly with C at run-time, while improving the clustering objective remains effective. The tool we apply for complexity reduction is variational EM, which is typically used to make training of generative models with exponentially many hidden states tractable. Here, we apply novel theoretical results on truncated variational EM to make tractable clustering algorithms more efficient. The basic idea is to use a partial variational E-step which reduces the linear complexity of O(NCD) required for a full E-step to a sublinear complexity. Our main observation is that the linear dependency on C can be reduced to a dependency on a much smaller parameter G which relates to cluster neighborhood relations. We focus on two versions of partial variational EM for clustering: variational GMM, scaling with O(NG²D), and variational k-means, scaling with O(NGD) per iteration. Empirical results show that these algorithms still require comparable numbers of iterations to improve the clustering objective to same values as k-means. For data with many clusters, we consequently observe reductions of net computational demands between two and three orders of magnitude. More generally, our results provide substantial empirical evidence in favor of clustering to scale sublinearly with C.

We study the convergence properties of the Expectation-Maximization algorithm in the Naive Bayes model. We show that EM can get stuck in regions of slow convergence, even when the features are binary and i.i.d. conditioning on the class label, and even under random (i.e. non worst-case) initialization. In turn, we show that EM can be bootstrapped in a pre-training step that computes a good initialization. From this initialization we show theoretically and experimentally that EM converges exponentially fast to the true model parameters. Our bootstrapping method amounts to running the EM algorithm on appropriately centered iterates of small magnitude, which as we show corresponds to effectively performing power iteration on the covariance matrix of the mixture model, although power iteration is performed under the hood by EM itself. As such, we call our bootstrapping approach “power EM.” Specifically for the case of two binary features, we show global exponentially fast convergence of EM, even without bootstrapping. Finally, as the Naive Bayes model is quite expressive, we show as corollaries of our convergence results that the EM algorithm globally converges to the true model parameters for mixtures of two Gaussians, recovering recent results of Xu et al.’2016 and Daskalakis et al. 2017.

Variational inference is a popular technique to approximate a possibly intractable Bayesian posterior with a more tractable one. Recently, boosting variational inference has been proposed as a new paradigm to approximate the posterior by a mixture of densities by greedily adding components to the mixture. However, as is the case with many other variational inference algorithms, its theoretical properties have not been studied. In the present work, we study the convergence properties of this approach from a modern optimization viewpoint by establishing connections to the classic Frank-Wolfe algorithm. Our analyses yields novel theoretical insights regarding the sufficient conditions for convergence, explicit rates, and algorithmic simplifications. Since a lot of focus in previous works for variational inference has been on tractability, our work is especially important as a much needed attempt to bridge the gap between probabilistic models and their corresponding theoretical properties.

In this paper, we propose a generalization of the best arm identification problem in stochastic multi-armed bandits (MAB) to the setting where every pull of an arm is associated with delayed feedbacks. The delay in feedbacks increases the effective sample complexity of the algorithm, but can be offset by partial feedbacks received before a pull is completed. We propose a a general modeling framework to structure in the partial feedbacks, and as a special case we introduce efficient algorithms for best arm identification in settings where the partial feedbacks are biased or unbiased estimators of the final outcome of the pull. Additionally, we propose a novel extension of the algorithms to the parallel MAB setting where an agent can control a batch of arms. Experiments on simulated as well as real world datasets of policy search for charging chemical batteries and hyperparameter optimization for mixed integer programming demonstrate that exploiting the structure of partial and delayed feedbacks can lead to significant improvements over baselines on both sequential and parallel MAB.

The superposition of temporal point processes has been studied for many years, although the usefulness of such models for practical applications has not be fully developed. We investigate superposed Hawkes process as an important class of such models, with properties studied in the framework of least squares estimation. The superposition of Hawkes processes is demonstrated to be beneficial for tightening the upper bound of excess risk under certain conditions, and we show the feasibility of the benefit in typical situations. The usefulness of superposed Hawkes processes is verified on synthetic data, and its potential to solve the cold-start problem of recommendation systems is demonstrated on real-world data.

Evaluating integrals is an ubiquitous issue and Monte Carlo methods, exploiting advances in random number generation over the last decades, offer a popular and powerful alternative to integration deterministic techniques, unsuited in particular when the domain of integration is complex. This paper is devoted to the study of a kernel smoothing based competitor built from a sequence of $n\geq 1$ i.i.d random vectors with arbitrary continuous probability distribution $f(x)dx$, originally proposed in Delyon et al. (2016), from a nonasymptotic perspective. We establish a probability bound showing that the method under study, though biased, produces an estimate approximating the target integral $\int_{x\in\mathbb{R}^d}\varphi(x)dx$ with an error bound of order $o(1/\sqrt{n})$ uniformly over a class $\Phi$ of functions $\varphi$, under weak complexity/smoothness assumptions related to the class $\Phi$, outperforming Monte-Carlo procedures. This striking result is shown to derive from an appropriate decomposition of the maximal deviation between the target integrals and their estimates, highlighting the remarkable benefit to averaging strongly dependent terms regarding statistical accuracy in this situation. The theoretical analysis then rests on sharp probability inequalities for degenerate $U$-statistics. It is illustrated by numerical results in the context of covariate shift regression, providing empirical evidence of the relevance of the approach.

Human brain activity as measured by fMRI exhibits strong correlations between brain regions which are believed to vary over time. Importantly, dynamic connectivity has been linked to individual differences in physiology, psychology and behavior, and has shown promise as a biomarker for disease. The state of the art in computational neuroimaging is to estimate the brain networks as relatively short sliding window covariance matrices, which leads to high variance estimates, thereby resulting in high overall error. This manuscript proposes a novel Bayesian model for dynamic brain connectivity. Motivated by the underlying neuroscience, the model estimates covariances which vary smoothly over time, with an instantaneous decomposition into a collection of spatially sparse components – resulting in parsimonious and highly interpretable estimates of dynamic brain connectivity. Simulated results are presented to illustrate the performance of the model even when it is mis-specified. For real brain imaging data with unknown ground truth, in addition to qualitative evaluation, we devise a simple classification task which suggests that the estimated brain networks better capture the underlying structure.

Analyzing the underlying structure of multiple time-sequences provides insights into the understanding of social networks and human activities. In this work, we present the Bayesian nonparametric Poisson process allocation (BaNPPA), a latent-function model for time-sequences, which automatically infers the number of latent functions. We model the intensity of each sequence as an infinite mixture of latent functions, each of which is obtained using a function drawn from a Gaussian process. We show that a technical challenge for the inference of such mixture models is the unidentifiability of the weights of the latent functions. We propose to cope with the issue by regulating the volume of each latent function within a variational inference algorithm. Our algorithm is computationally efficient and scales well to large data sets. We demonstrate the usefulness of our proposed model through experiments on both synthetic and real-world data sets.

We present a probabilistic, fully Bayesian framework for multi-label learning. Our framework is based on the idea of learning a joint low-rank embedding of the label matrix and the label co-occurrence matrix. The proposed framework has the following appealing aspects: (1) It leverages the sparsity in the label matrix and the feature matrix, which results in very efficient inference, especially for sparse datasets, commonly encountered in multi-label learning problems, and (2) By effectively utilizing the label co-occurrence information, the model yields improved prediction accuracies, especially in the case where the amount of training data is low and/or the label matrix has a significant fraction of missing labels. Our framework enjoys full local conjugacy and admits a simple inference procedure via a scalable Gibbs sampler. We report experimental results on a number of benchmark datasets, on which it outperforms several state-of-the-art multi-label learning models.

Distribution regression has recently attracted much interest as a generic solution to the problem of supervised learning where labels are available at the group level, rather than at the individual level. Current approaches, however, do not propagate the uncertainty in observations due to sampling variability in the groups. This effectively assumes that small and large groups are estimated equally well, and should have equal weight in the final regression. We account for this uncertainty with a Bayesian distribution regression formalism, improving the robustness and performance of the model when group sizes vary. We frame our models in a neural network style, allowing for simple MAP inference using backpropagation to learn the parameters, as well as MCMC-based inference which can fully propagate uncertainty. We demonstrate our approach on illustrative toy datasets, as well as on a challenging problem of predicting age from images.

Bayesian optimization (BO) has become an effective approach for black-box function optimization problems when function evaluations are expensive and the optimum can be achieved within a relatively small number of queries. However, many cases, such as the ones with high-dimensional inputs, may require a much larger number of observations for optimization. Despite an abundance of observations thanks to parallel experiments, current BO techniques have been limited to merely a few thousand observations. In this paper, we propose ensemble Bayesian optimization (EBO) to address three current challenges in BO simultaneously: (1) large-scale observations; (2) high dimensional input spaces; and (3) selections of batch queries that balance quality and diversity. The key idea of EBO is to operate on an ensemble of additive Gaussian process models, each of which possesses a randomized strategy to divide and conquer. We show unprecedented, previously impossible results of scaling up BO to tens of thousands of observations within minutes of computation.

We propose Batch-Expansion Training (BET), a framework for running a batch optimizer on a gradually expanding dataset. As opposed to stochastic approaches, batches do not need to be resampled i.i.d. at every iteration, thus making BET more resource efficient in a distributed setting, and when disk-access is constrained. Moreover, BET can be easily paired with most batch optimizers, does not require any parameter-tuning, and compares favorably to existing stochastic and batch methods. We show that when the batch size grows exponentially with the number of outer iterations, BET achieves optimal O (1/epsilon) data-access convergence rate for strongly convex objectives. Experiments in parallel and distributed settings show that BET performs better than standard batch and stochastic approaches.

Group regularized learning problems (such as group Lasso) are important in machine learning. The asynchronous parallel stochastic optimization algorithms have received huge attentions recently as handling large scale problems. However, existing asynchronous stochastic algorithms for solving the group regularized learning problems are not scalable enough simultaneously in sample size and feature dimensionality. To address this challenging problem, in this paper, we propose a novel asynchronous doubly stochastic proximal gradient algorithm with variance reduction (AsyDSPG+). To the best of our knowledge, AsyDSPG+ is the first asynchronous doubly stochastic proximal gradient algorithm, which can scale well with the large sample size and high feature dimensionality simultaneously. More importantly, we provide a comprehensive convergence guarantee to AsyDSPG+. The experimental results on various large-scale real-world datasets not only confirm the fast convergence of our new method, but also show that AsyDSPG+ scales better than the existing algorithms with the sample size and dimension simultaneously.

A common problem in machine learning is to rank a set of n items based on pairwise comparison. Here, ranking refers to partitioning the items into sets of pre-specified sizes according to theirs scores, which includes identification of the top-k items as the most prominent special case. The score of a given item is defined as the probability that it beats a randomly chosen other item. In practice, in particular when n is large, finding an exact ranking typically requires a prohibitively large number of comparisons. What comes to our rescue here is that in practice, one is usually content with finding an approximate ranking. In this paper we consider the problem of finding approximate rankings from pairwise comparisons. We analyze an active ranking algorithm that counts the number of comparisons won, and decides whether to stop or which pair of items to compare next, based on confidence intervals computed from the data collected in previous steps. We show that this algorithm succeeds in recovering approximate rankings using a number of comparisons that is close to optimal up to logarithmic factors. We also present numerical results, showing that in practice, approximation can drastically reduce the number of comparisons required to estimate a ranking.

Complex simulator-based models usually have intractable likelihood functions, rendering the likelihood-based inference methods inapplicable. Approximate Bayesian Computation (ABC) emerges as an alternative framework of likelihood-free inference methods. It identifies a quasi-posterior distribution by finding values of parameter that simulate the synthetic data resembling the observed data. A major ingredient of ABC is the discrepancy measure between the observed and the simulated data, which conventionally involves a fundamental difficulty of constructing effective summary statistics. To bypass this difficulty, we adopt a Kullback-Leibler divergence estimator to assess the data discrepancy. Our method enjoys the asymptotic consistency and linearithmic time complexity as the data size increases. In experiments on five benchmark models, this method achieves a comparable or higher quasi-posterior quality, compared to the existing methods using other discrepancy measures.

A falling rule list is a probabilistic decision list for binary classification, consisting of a series of if-then rules with antecedents in the if clauses and probabilities of the desired outcome ("1") in the then clauses. Just as in a regular decision list, the order of rules in a falling rule list is important – each example is classified by the first rule whose antecedent it satisfies. Unlike a regular decision list, a falling rule list requires the probabilities of the desired outcome ("1") to be monotonically decreasing down the list. We propose an optimization approach to learning falling rule lists and "softly" falling rule lists, along with Monte-Carlo search algorithms that use bounds on the optimal solution to prune the search space.

Distributional approaches to value-based reinforcement learning model the entire distribution of returns, rather than just their expected values, and have recently been shown to yield state-of-the-art empirical performance. This was demonstrated by the recently proposed C51 algorithm, based on categorical distributional reinforcement learning (CDRL) [Bellemare et al., 2017]. However, the theoretical properties of CDRL algorithms are not yet well understood. In this paper, we introduce a framework to analyse CDRL algorithms, establish the importance of the projected distributional Bellman operator in distributional RL, draw fundamental connections between CDRL and the Cramer distance, and give a proof of convergence for sample-based categorical distributional reinforcement learning algorithms.