AISTATS 2017 Papers

Computing partition functions, the normalizing constants of probability distributions, is often hard. Variants of importance sampling give unbiased estimates of a normalizer Z, however, unbiased estimates of the reciprocal 1/Z are harder to obtain. Unbiased estimates of 1/Z allow Markov chain Monte Carlo sampling of “doubly-intractable” distributions, such as the parameter posterior for Markov Random Fields or Exponential Random Graphs. We demonstrate how to construct unbiased estimates for 1/Z given access to black-box importance sampling estimators for Z. We adapt recent work on random series truncation and Markov chain coupling, producing estimators with lower variance and a higher percentage of positive estimates than before. Our debiasing algorithms are simple to implement, and have some theoretical and empirical advantages over existing methods.

We consider the problem of estimating the underlying graph associated with a Markov random field, with the added twist that the decoding algorithm can iteratively choose which subsets of nodes to sample based on the previous samples, resulting in an active learning setting. Considering both Ising and Gaussian models, we provide algorithm-independent lower bounds for high-probability recovery within the class of degree-bounded graphs. Our main results are minimax lower bounds for the active setting that match the best known lower bounds for the passive setting, which in turn are known to be tight in several cases of interest. Our analysis is based on Fano’s inequality, along with novel mutual information bounds for the active learning setting, and the application of restricted graph ensembles. While we consider ensembles that are similar or identical to those used in the passive setting, we require different analysis techniques, with a key challenge being bounding a mutual information quantity associated with observed subsets of nodes, as opposed to full observations.

We introduce the localized Lasso, which learns models that both are interpretable and have a high predictive power in problems with high dimensionality d and small sample size n. More specifically, we consider a function defined by local sparse models, one at each data point. We introduce sample-wise network regularization to borrow strength across the models, and sample-wise exclusive group sparsity (a.k.a., l12 norm) to introduce diversity into the choice of feature sets in the local models. The local models are interpretable in terms of similarity of their sparsity patterns. The cost function is convex, and thus has a globally optimal solution. Moreover, we propose a simple yet efficient iterative least-squares based optimization procedure for the localized Lasso, which does not need a tuning parameter, and is guaranteed to converge to a globally optimal solution. The solution is empirically shown to outperform alternatives for both simulated and genomic personalized/precision medicine data.

Conditional random fields (CRFs) provide a powerful tool for structured prediction, but cast significant challenges in both the learning and inference steps. Approximation techniques are widely used in both steps, which should be considered jointly to guarantee good performance (a.k.a. “inferning"). Perturb-and-MAP models provide a promising alternative to CRFs, but require global combinatorial optimization and hence they are usable only on specific models. In this work, we present a new Local Perturb-and-MAP (locPMAP) framework that replaces the global optimization with a local optimization by exploiting our observed connection between locPMAP and the pseudolikelihood of the original CRF model. We test our approach on three different vision tasks and show that our method achieves consistently improved performance over other approximate inference techniques optimized to a pseudolikelihood objective. Additionally, we demonstrate that we can integrate our method in the fully convolutional network framework to increase our model’s complexity. Finally, our observed connection between locPMAP and the pseudolikelihood leads to a novel perspective for understanding and using pseudolikelihood.

Invariance to nuisance transformations is one of the desirable properties of effective representations. We consider transformations that form a group and propose an approach based on kernel methods to derive local group invariant representations. Locality is achieved by defining a suitable probability distribution over the group which in turn induces distributions in the input feature space. We learn a decision function over these distributions by appealing to the powerful framework of kernel methods and generate local invariant random feature maps via kernel approximations. We show uniform convergence bounds for kernel approximation and provide generalization bounds for learning with these features. We evaluate our method on three real datasets, including Rotated MNIST and CIFAR-10, and observe that it outperforms competing kernel based approaches. The proposed method also outperforms deep CNN on Rotated MNIST and performs comparably to the recently proposed group-equivariant CNN.

Density-ratio estimation (i.e. estimating $f = f_Q/f_P$ for two unknown distributions Q and P) has proved useful in many Machine Learning tasks, e.g., risk-calibration in transfer-learning, two-sample tests, and also useful in common techniques such importance sampling and bias correction. While there are many important analyses of this estimation problem, the present paper derives convergence rates in other practical settings that are less understood, namely, extensions of traditional Lipschitz smoothness conditions, and common high-dimensional settings with structured data (e.g. manifold data, sparse data). Various interesting facts, which hold in earlier settings, are shown to extend to these settings. Namely, (1) optimal rates depend only on the smoothness of the ratio f, and not on the densities $f_Q$, $f_P$, supporting the belief that plugging in estimates for $f_Q$, $f_P$ is suboptimal; (2) optimal rates depend only on the intrinsic dimension of data, i.e. this problem – unlike density estimation – escapes the curse of dimension. We further show that near-optimal rates are attainable by estimators tuned from data alone, i.e. with no prior distributional information. This last fact is of special interest in unsupervised settings such as this one, where only oracle rates seem to be known, i.e., rates which assume critical distributional information usually unavailable in practice.

User behaviors in social networks are microscopic with fine grained temporal information. Predicting a macroscopic quantity based on users’ collective behaviors is an important problem. However, existing works are mainly problem-specific models for the microscopic behaviors and typically design approximation or heuristic algorithms to compute the macroscopic quantity. In this paper, we propose a unifying framework with a jump stochastic differential equation model that systematically links the microscopic event data and macroscopic inference, and the theory to approximate its probability distribution. We showed that our method can better predict the user behaviors in real-world applications.

We derive an alternative proof for the regret of Thompson sampling (TS) in the stochastic linear bandit setting. While we obtain a regret bound of order $O(d^3/2\sqrtT)$ as in previous results, the proof sheds new light on the functioning of the TS. We leverage on the structure of the problem to show how the regret is related to the sensitivity (i.e., the gradient) of the objective function and how selecting optimal arms associated to \textitoptimistic parameters does control it. Thus we show that TS can be seen as a generic randomized algorithm where the sampling distribution is designed to have a fixed probability of being optimistic, at the cost of an additional $\sqrtd$ regret factor compared to a UCB-like approach. Furthermore, we show that our proof can be readily applied to regularized linear optimization and generalized linear model problems.

In this paper, we show that the Away-step Stochastic Frank-Wolfe (ASFW) and Pairwise Stochastic Frank-Wolfe (PSFW) algorithms converge linearly in expectation. We also show that if an algorithm convergences linearly in expectation then it converges linearly almost surely. In order to prove these results, we develop a novel proof technique based on concepts of empirical processes and concentration inequalities. As far as we know, this technique has not been used previously to derive the convergence rates of stochastic optimization algorithms. In large- scale numerical experiments, ASFW and PSFW perform as well as or better than their stochastic competitors in actual CPU time.

We develop and analyze a procedure for gradient-based optimization that we refer to as stochastically controlled stochastic gradient (SCSG). As a member of the SVRG family of algorithms, SCSG makes use of gradient estimates at two scales. Unlike most existing algorithms in this family, both the computation cost and the communication cost of SCSG do not necessarily scale linearly with the sample size n; indeed, these costs are independent of n when the target accuracy is small. An experimental evaluation of SCSG on the MNIST dataset shows that it can yield accurate results on this dataset on a single commodity machine with a memory footprint of only 2.6MB and only eight disk accesses.

Mean shift is a mode-seeking clustering algorithm that has been successfully used in a wide range of applications such as image segmentation and object tracking. To further improve the clustering performance, mean shift has been extended to various directions, including generalization to handle data on Riemannian manifolds and extension to directly estimating the density gradient without density estimation. In this paper, we combine these ideas and propose a novel mode-seeking algorithm for Riemannian manifolds with direct density-gradient estimation. Although the idea of combining the two extensions is rather straightforward, directly estimating the density gradient on Riemannian manifolds is mathematically challenging. We will provide a mathematically sound algorithm and demonstrate its usefulness through experiments.

In supervised learning, a human annotator only needs to assign each data point (document, image, etc.) its correct label. But in many situations, the human can also provide richer feedback at essentially no extra cost. In this paper, we examine a particular type of feature feedback that has been used, with some success, in information retrieval and in computer vision. We formalize two models of feature feedback, give learning algorithms for them, and quantify their usefulness in the learning process. Our experiments also show the efficacy of these methods.

Markov random fields (MRFs) are a useful tool for modeling relationships present in large and high-dimensional data. Often, this data comes from various sources and can have diverse distributions, for example a combination of numerical, binary, and categorical variables. Here, we define the pairwise exponential Markov random field (PE-MRF), an approach capable of modeling exponential family distributions in heterogeneous domains. We develop a scalable method of learning the graphical structure across the variables by solving a regularized approximated maximum likelihood problem. Specifically, we first derive a tractable upper bound on the log-partition function. We then use this upper bound to derive the group graphical lasso, a generalization of the classic graphical lasso problem to heterogeneous domains. To solve this problem, we develop a fast algorithm based on the alternating direction method of multipliers (ADMM). We also prove that our estimator is sparsistent, with guaranteed recovery of the true underlying graphical structure, and that it has a polynomially faster runtime than the current state-of-the-art method for learning such distributions. Experiments on synthetic and real-world examples demonstrate that our approach is both efficient and accurate at uncovering the structure of heterogeneous data.

Many machine learning tasks, such as learning with invariance and policy evaluation in reinforcement learning, can be characterized as problems of learning from conditional distributions. In such problems, each sample x itself is associated with a conditional distribution $p(z|x)$ represented by samples $z_i_i=1^M$, and the goal is to learn a function f that links these conditional distributions to target values y. These problems become very challenging when we only have limited samples or in the extreme case only one sample from each conditional distribution. Commonly used approaches either assume that z is independent of x, or require an overwhelmingly large set of samples from each conditional distribution. To address these challenges, we propose a novel approach which employs a new min-max reformulation of the learning from conditional distribution problem. With such new reformulation, we only need to deal with the joint distribution p(z,x). We also design an efficient learning algorithm, Embedding-SGD, and establish theoretical sample complexity for such problems. Finally, our numerical experiments, on both synthetic and real-world datasets, show that the proposed approach can significantly improve over existing algorithms.

In this paper, we consider a new low-quality label learning problem: learning time series detection models from temporally imprecise labels. In this problem, the data consist of a set of input time series, and supervision is provided by a sequence of noisy time stamps corresponding to the occurrence of positive class events. Such temporally imprecise labels commonly occur in areas like mobile health research where human annotators are tasked with labeling the occurrence of very short duration events. We propose a general learning framework for this problem that can accommodate different base classifiers and noise models. We present results on real mobile health data showing that the proposed framework significantly outperforms a number of alternatives including assuming that the label time stamps are noise-free, transforming the problem into the multiple instance learning framework, and learning on labels that were manually re-aligned.

In this work we study the learnability of stochastic processes with respect to the conditional risk, i.e. the existence of a learning algorithm that improves its next-step performance with the amount of observed data. We introduce a notion of pairwise discrepancy between conditional distributions at different times steps and show how certain properties of these discrepancies can be used to construct a successful learning algorithm. Our main results are two theorems that establish criteria for learnability for many classes of stochastic processes, including all special cases studied previously in the literature.

Deep neural networks (DNNs) are increasingly popular in modern machine learning. Bayesian learning affords the opportunity to quantify posterior uncertainty on DNN model parameters. Most existing work adopts independent Gaussian priors on the model weights, ignoring possible structural information. In this paper, we consider the matrix variate Gaussian (MVG) distribution to model structured correlations within the weights of a DNN. To make posterior inference feasible, a reparametrization is proposed for the MVG prior, simplifying the complex MVG-based model to an equivalent yet simpler model with independent Gaussian priors on the transformed weights. Consequently, we develop a scalable Bayesian online inference algorithm by adopting the recently proposed probabilistic backpropagation framework. Experiments on several synthetic and real datasets indicate the superiority of our model, achieving competitive performance in terms of model likelihood and predictive root mean square error. Importantly, it also yields faster convergence speed compared to related Bayesian DNN models.

Our goal is to identify beneficial interventions from observational data. We consider interventions that are narrowly focused (impacting few covariates) and may be tailored to each individual or globally enacted over a population. For applications where harmful intervention is drastically worse than proposing no change, we propose a conservative definition of the optimal intervention. Assuming the underlying relationship remains invariant under intervention, we develop efficient algorithms to identify the optimal intervention policy from limited data and provide theoretical guarantees for our approach in a Gaussian Process setting. Although our methods assume covariates can be precisely adjusted, they remain capable of improving outcomes in misspecified settings with unintentional downstream effects. Empirically, our approach identifies good interventions in two practical applications: gene perturbation and writing improvement.

We present a framework for incorporating prior information into nonparametric estimation of graphical models. To avoid distributional assumptions, we restrict the graph to be a forest and build on the work of forest density estimation (FDE). We reformulate the FDE approach from a Bayesian perspective, and introduce prior distributions on the graphs. As two concrete examples, we apply this framework to estimating scale-free graphs and learning multiple graphs with similar structures. The resulting algorithms are equivalent to finding a maximum spanning tree of a weighted graph with a penalty term on the connectivity pattern of the graph. We solve the optimization problem via a minorize-maximization procedure with Kruskal’s algorithm. Simulations show that the proposed methods outperform competing parametric methods, and are robust to the true data distribution. They also lead to improvement in predictive power and interpretability in two real data sets.