AISTATS 2016 Papers

We propose a Laplace approximation that creates a stochastic unit from any smooth monotonic activation function, using only Gaussian noise. This paper investigates the application of this stochastic approximation in training a family of Restricted Boltzmann Machines (RBM) that are closely linked to Bregman divergences. This family, that we call exponential family RBM (Exp-RBM), is a subset of the exponential family Harmoniums that expresses family members through a choice of smooth monotonic non-linearity for each neuron. Using contrastive divergence along with our Gaussian approximation, we show that Exp-RBM can learn useful representations using novel stochastic units.

Method of moment estimators exhibit appealing statistical properties, such as asymptotic unbiasedness, for nonconvex problems. However, they typically require a large number of samples and are extremely sensitive to model misspecification. In this paper, we apply the framework of M-estimation to develop both a generalized method of moments procedure and a principled method for regularization. Our proposed M-estimator obtains optimal sample efficiency rates (in the class of moment-based estimators) and the same well-known rates on prediction accuracy as other spectral estimators. It also makes it straightforward to incorporate regularization into the sample moment conditions. We demonstrate empirically the gains in sample efficiency from our approach on hidden Markov models.

Extremes play a special role in Anomaly Detection. Beyond inference and simulation purposes, probabilistic tools borrowed from Extreme Value Theory (EVT), such as the \textitangular measure, can also be used to design novel statistical learning methods for Anomaly Detection/ranking. This paper proposes a new algorithm based on multivariate EVT to learn how to rank observations in a high dimensional space with respect to their degree of ‘abnormality’. The procedure relies on an original dimension-reduction technique in the extreme domain that possibly produces a sparse representation of multivariate extremes and allows to gain insight into the dependence structure thereof, escaping the curse of dimensionality. The representation output by the unsupervised methodology we propose here can be combined with any Anomaly Detection technique tailored to non-extreme data. As it performs linearly with the dimension and almost linearly in the data (in O(d n \log n)), it fits to large scale problems. The approach in this paper is novel in that EVT has never been used in its multivariate version in the field of Anomaly Detection. Illustrative experimental results provide strong empirical evidence of the relevance of our approach.

In this paper we show how to embed information distances like the χ^2 and Jensen-Shannon divergences efficiently in low dimensional spaces while preserving all pairwise distances. We then prove a dimensionality reduction result for the Hellinger, Jensen–Shannon, and χ^2 divergences that preserves the information geometry of the distributions, specifically, by retaining the simplex structure of the space. While our first result already implies these divergences can be explicitly embedded in the Euclidean space, retaining the simplex structure is important because it allows us to do inferences in the reduced space. We also show that these divergences can be sketched efficiently (i.e., up to a multiplicative error in sublinear space) in the aggregate streaming model. This result is exponentially stronger than known upper bounds for sketching these distances in the strict turnstile streaming model.

We present a simple spectral approach to the well-studied constrained clustering problem. It captures constrained clustering as a generalized eigenvalue problem with graph Laplacians. The algorithm works in nearly-linear time and provides concrete guarantees for the quality of the clusters, at least for the case of 2-way partitioning. In practice this translates to a very fast implementation that consistently outperforms existing spectral approaches both in speed and quality.

A deep Gaussian process (DGP) is a deep network in which each layer is modelled with a Gaussian process (GP). It is a flexible model that can capture highly-nonlinear functions for complex data sets. However, the network structure of DGP often makes inference computationally expensive. In this paper, we propose an efficient sequential inference framework for DGP, where the data is processed sequentially. We also propose two DGP extensions to handle heteroscedasticity and multi-task learning. Our experimental evaluation shows the effectiveness of our sequential inference framework on a number of important learning tasks.

This paper proposes a novel nonparametric framework for semi-supervised learning and for optimizing the Laplacian spectrum of the data manifold simultaneously. Our formulation leads to a convex optimization problem that can be efficiently solved via the bundle method, and can be interpreted as to asymptotically minimize the generalization error bound of semi-supervised learning with respect to the graph spectrum. Experiments over benchmark datasets in various domains show advantageous performance of the proposed method over strong baselines.

A standard model of (conditional) heteroscedasticity, i.e., the phenomenon that the variance of a process changes over time, is the Generalized AutoRegressive Conditional Heteroskedasticity (GARCH) model, which is especially important for economics and finance. GARCH models are typically estimated by the Quasi-Maximum Likelihood (QML) method, which works under mild statistical assumptions. Here, we suggest a finite sample approach, called ScoPe, to construct distribution-free confidence regions around the QML estimate, which have exact coverage probabilities, despite no additional assumptions about moments are made. ScoPe is inspired by the recently developed Sign-Perturbed Sums (SPS) method, which however cannot be applied in the GARCH case. ScoPe works by perturbing the score function using randomly permuted residuals. This produces alternative samples which lead to exact confidence regions. Experiments on simulated and stock market data are also presented, and ScoPe is compared with the asymptotic theory and bootstrap approaches.

It is standard to assume a low-dimensional structure in estimating a high-dimensional density. However, popular methods, such as probabilistic principal component analysis, scale poorly computationally. We introduce a novel empirical Bayes method that we term geometric density estimation (GEODE) and show that, with mild conditions and among all d-dimensional linear subspaces, the span of the d leading principal axes of the data maximizes the model posterior. With these axes pre-computed using fast singular value decomposition, GEODE easily scales to high dimensional problems while providing uncertainty characterization. The model is also capable of imputing missing data and dynamically deleting redundant dimensions. Finally, we generalize GEODE by mixing it across a dyadic clustering tree. Both simulation studies and real world data applications show superior performance of GEODE in terms of robustness and computational efficiency.

Many real-world data arise naturally as tensors. Equipped with a low rank prior, learning algorithms can benefit from exploiting the rich dependency encoded in a tensor. Despite its prevalence in low-rank matrix learning, trace norm ceases to be tractable in tensors and therefore most existing works resort to matrix unfolding. Although some theoretical guarantees are available, these approaches may lose valuable structure information and are not scalable in general. To address this problem, we propose directly optimizing the tensor trace norm by approximating its dual spectral norm, and we show that the approximation bounds can be efficiently converted to the original problem via the generalized conditional gradient algorithm. The resulting approach is scalable to large datasets, and matches state-of-the-art recovery guarantees. Experimental results on tensor completion and multitask learning confirm the superiority of the proposed method.

We propose a stochastic gradient Markov chain Monte Carlo (SG-MCMC) algorithm for scalable inference in mixed-membership stochastic blockmodels (MMSB). Our algorithm is based on the stochastic gradient Riemannian Langevin sampler and achieves both faster speed and higher accuracy at every iteration than the current state-of-the-art algorithm based on stochastic variational inference. In addition we develop an approximation that can handle models that entertain a very large number of communities. The experimental results show that SG-MCMC strictly dominates competing algorithms in all cases.

We present a scalable Gaussian process model for identifying and characterizing smooth multidimensional changepoints, and automatically learning changes in expressive covariance structure. We use Random Kitchen Sink features to flexibly define a change surface in combination with expressive spectral mixture kernels to capture the complex statistical structure. Finally, through the use of novel methods for additive non-separable kernels, we can scale the model to large datasets. We demonstrate the model on numerical and real world data, including a large spatio-temporal disease dataset where we identify previously unknown heterogeneous changes in space and time.

Variational methods have been recently considered for scaling the training process of Gaussian process classifiers to large datasets. As an alternative, we describe here how to train these classifiers efficiently using expectation propagation (EP). The proposed EP method allows to train Gaussian process classifiers on very large datasets, with millions of instances, that were out of the reach of previous implementations of EP. More precisely, it can be used for (i) training in a distributed fashion where the data instances are sent to different nodes in which the required computations are carried out, and for (ii) maximizing an estimate of the marginal likelihood using a stochastic approximation of the gradient. Several experiments involving large datasets show that the method described is competitive with the variational approach.

In recent years exemplar clustering has become a popular tool for applications in document and video summarization, active learning, and clustering with general similarity, where cluster centroids are required to be a subset of the data samples rather than their linear combinations. The problem is also well-known as facility location in the operations research literature. While the problem has well-developed convex relaxation with approximation and recovery guarantees, it has number of variables grows quadratically with the number of samples. Therefore, state-of-the-art methods can hardly handle more than 10^4 samples (i.e. 10^8 variables). In this work, we propose an Augmented-Lagrangian with Block Coordinate Descent (AL-BCD) algorithm that utilizes problem structure to obtain closed-form solution for each block sub-problem, and exploits low-rank representation of the dissimilarity matrix to search active columns without computing the entire matrix. Experiments show our approach to be orders of magnitude faster than existing approaches and can handle problems of up to 10^6 samples. We also demonstrate successful applications of the algorithm on world-scale facility location, document summarization and active learning.

In many learning settings, the source data available to train a regression model differs from the target data it encounters when making predictions due to input distribution shift. Appropriately dealing with this situation remains an important challenge. Existing methods attempt to “reweight” the source data samples to better represent the target domain, but this introduces strong inductive biases that are highly extrapolative and can often err greatly in practice. We propose a robust approach for regression under covariate shift that embraces the uncertainty resulting from sample selection bias by producing regression models that are explicitly robust to it. We demonstrate the benefits of our approach on a number of regression tasks.

We study the problem of recovering the subspace spanned by the first k principal components of d-dimensional data under the streaming setting, with a memory bound of O(kd). Two families of algorithms are known for this problem. The first family is based on the framework of stochastic gradient descent. Nevertheless, the convergence rate of the family can be seriously affected by the learning rate of the descent steps and deserves more serious study. The second family is based on the power method over blocks of data, but setting the block size for its existing algorithms is not an easy task. In this paper, we analyze the convergence rate of a representative algorithm with decayed learning rate (Oja and Karhunen, 1985) in the first family for the general k>1 case. Moreover, we propose a novel algorithm for the second family that sets the block sizes automatically and dynamically with faster convergence rate. We then conduct empirical studies that fairly compare the two families on real-world data. The studies reveal the advantages and disadvantages of these two families.

We study a graph bandit setting where the objective of the learner is to detect the most influential node of a graph by requesting as little information from the graph as possible. One of the relevant applications for this setting is marketing in social networks, where the marketer aims at finding and taking advantage of the most influential customers. The existing approaches for bandit problems on graphs require either partial or complete knowledge of the graph. In this paper, we do not assume any knowledge of the graph, but we consider a setting where it can be gradually discovered in a sequential and active way. At each round, the learner chooses a node of the graph and the only information it receives is a stochastic set of the nodes that the chosen node is currently influencing. To address this setting, we propose BARE, a bandit strategy for which we prove a regret guarantee that scales with the detectable dimension, a problem dependent quantity that is often much smaller than the number of nodes.

A pretraining algorithm, which is a layer-by-layer greedy learning algorithm, for a deep Boltzmann machine (DBM) is presented in this paper. By considering the deep belief net type of pretraining for the DBM, which is a simplified version of the original pretraining of the DBM, two interesting theoretical facts about pretraining can be obtained. (1) By applying two different types of approximation, a replacing approximation by using a Bayesian network and a Bethe type of approximation based on the cluster variation method, to two different parts of the true log-likelihood function of the DBM, pretraining can be derived from a variational approximation of the original maximum likelihood estimation. (2) It can be ensured that the pretraining improves the variational bound of the true log-likelihood function of the DBM. These two theoretical results will help deepen our understanding of deep learning. Moreover, on the basis of the theoretical results, we discuss the original pretraining of the DBM in the latter part of this paper.

Buyers (e.g., advertisers) often have limited financial or processing resources, and so their participation in auctions is throttled. Changes to auctions may affect bids or throttling, and any change may affect what buyers pay. This paper shows that if an A/B experiment affects only bids, then the observed treatment effect is an unbiased estimator when all the bidders in the same auction are randomly assigned to A or B but the observed treatment effect can be severely biased otherwise, even in the absence of throttling. Experiments that affect throttling algorithms can also be badly biased, but the bias can be much reduced if separate budgets are maintained for the A and B arms of the experiment.