ICML 2015 Papers

Spectral clustering is one of the most important algorithms in data mining and machine intelligence; however, its computational complexity limits its application to truly large scale data analysis. The computational bottleneck in spectral clustering is computing a few of the top eigenvectors of the (normalized) Laplacian matrix corresponding to the graph representing the data to be clustered. One way to speed up the computation of these eigenvectors is to use the “power method” from the numerical linear algebra literature. Although the power method has been empirically used to speed up spectral clustering, the theory behind this approach, to the best of our knowledge, remains unexplored. This paper provides the first such rigorous theoretical justification, arguing that a small number of power iterations suffices to obtain near-optimal partitionings using the approximate eigenvectors. Specifically, we prove that solving the k-means clustering problem on the approximate eigenvectors obtained via the power method gives an additive-error approximation to solving the k-means problem on the optimal eigenvectors.

Gaussian processes (GP) provide an attractive machine learning model due to their non-parametric form, their flexibility to capture many types of observation data, and their generic inference procedures. Sparse GP inference algorithms address the cubic complexity of GPs by focusing on a small set of pseudo-samples. To date, such approaches have focused on the simple case of Gaussian observation likelihoods. This paper develops a variational sparse solution for GPs under general likelihoods by providing a new characterization of the gradients required for inference in terms of individual observation likelihood terms. In addition, we propose a simple new approach for optimizing the sparse variational approximation using a fixed point computation. We demonstrate experimentally that the fixed point operator acts as a contraction in many cases and therefore leads to fast convergence. An experimental evaluation for count regression, classification, and ordinal regression illustrates the generality and advantages of the new approach.

We consider the problem of clustering incomplete data drawn from a union of subspaces. Classical subspace clustering methods are not applicable to this problem because the data are incomplete, while classical low-rank matrix completion methods may not be applicable because data in multiple subspaces may not be low rank. This paper proposes and evaluates two new approaches for subspace clustering and completion. The first one generalizes the sparse subspace clustering algorithm so that it can obtain a sparse representation of the data using only the observed entries. The second one estimates a suitable kernel matrix by assuming a random model for the missing entries and obtains the sparse representation from this kernel. Experiments on synthetic and real data show the advantages and disadvantages of the proposed methods, which all outperform the natural approach (low-rank matrix completion followed by sparse subspace clustering) when the data matrix is high-rank or the percentage of missing entries is large.

We consider a stochastic bandit problem with infinitely many arms. In this setting, the learner has no chance of trying all the arms even once and has to dedicate its limited number of samples only to a certain number of arms. All previous algorithms for this setting were designed for minimizing the cumulative regret of the learner. In this paper, we propose an algorithm aiming at minimizing the simple regret. As in the cumulative regret setting of infinitely many armed bandits, the rate of the simple regret will depend on a parameter βcharacterizing the distribution of the near-optimal arms. We prove that depending on β, our algorithm is minimax optimal either up to a multiplicative constant or up to a \log(n) factor. We also provide extensions to several important cases: when βis unknown, in a natural setting where the near-optimal arms have a small variance, and in the case of unknown time horizon.

Inspired by recent work in machine translation and object detection, we introduce an attention based model that automatically learns to describe the content of images. We describe how we can train this model in a deterministic manner using standard backpropagation techniques and stochastically by maximizing a variational lower bound. We also show through visualization how the model is able to automatically learn to fix its gaze on salient objects while generating the corresponding words in the output sequence. We validate the use of attention with state-of-the-art performance on three benchmark datasets: Flickr8k, Flickr30k and MS COCO.

Second-order optimization methods, such as natural gradient, are difficult to apply to high-dimensional problems, because they require approximately solving large linear systems. We present FActorized Natural Gradient (FANG), an approximation to natural gradient descent where the Fisher matrix is approximated with a Gaussian graphical model whose precision matrix can be computed efficiently. We analyze the Fisher matrix for a small RBM and derive an extremely sparse graphical model which is a good match to the covariance of the sufficient statistics. Our experiments indicate that FANG allows RBMs to be trained more efficiently compared with stochastic gradient descent. Additionally, our analysis yields insight into the surprisingly good performance of the “centering trick” for training RBMs.

We consider the problem of approximate Bayesian inference in log-supermodular models. These models encompass regular pairwise MRFs with binary variables, but allow to capture high order interactions, which are intractable for existing approximate inference techniques such as belief propagation, mean field and variants. We show that a recently proposed variational approach to inference in log-supermodular models – L-Field – reduces to the widely studied minimum norm problem for submodular minimization. This insight allows to leverage powerful existing tools, and allows solving the variational problem orders of magnitude more efficiently than previously possible. We then provide another natural interpretation of L-Field, demonstrating that it exactly minimizes a specific type of Renyi divergence measure. This insight sheds light on the nature of the variational approximations produced by L-Field. Furthermore, we show how to perform parallel inference as message passing in a suitable factor graph at a linear convergence rate, without having to sum up over all the configurations of the factor. Finally, we apply our approach to a challenging image segmentation task. Our experiments confirm scalability of our approach, high quality of the marginals and the benefit of incorporating higher order potentials.

In this paper we propose an efficient, scalable non-parametric Gaussian process model for inference on Poisson point processes. Our model does not resort to gridding the domain or to introducing latent thinning points. Unlike competing models that scale as O(n^3) over n data points, our model has a complexity O(nk^2) where k << n. We propose a MCMC sampler and show that the model obtained is faster, more accurate and generates less correlated samples than competing approaches on both synthetic and real-life data. Finally, we show that our model easily handles data sizes not considered thus far by alternate approaches.

With a growing need to understand large-scale networks, factorial relational models, such as binary matrix factorization models (BMFs), have become important in many applications. Although BMFs have a natural capability to uncover overlapping group structures behind network data, existing inference techniques have issues of either high computational cost or lack of model selection capability, and this limits their applicability. For scalable model selection of BMFs, this paper proposes stochastic factorized asymptotic Bayesian (sFAB) inference that combines concepts in two recently-developed techniques: stochastic variational inference (SVI) and FAB inference. sFAB is a highly-efficient algorithm, having both scalability and an inherent model selection capability in a single inference framework. Empirical results show the superiority of sFAB/BMF in both accuracy and scalability over state-of-the-art inference methods for overlapping relational models.

A new framework for topic modeling is developed, based on deep graphical models, where interactions between topics are inferred through deep latent binary hierarchies. The proposed multi-layer model employs a deep sigmoid belief network or restricted Boltzmann machine, the bottom binary layer of which selects topics for use in a Poisson factor analysis model. Under this setting, topics live on the bottom layer of the model, while the deep specification serves as a flexible prior for revealing topic structure. Scalable inference algorithms are derived by applying Bayesian conditional density filtering algorithm, in addition to extending recently proposed work on stochastic gradient thermostats. Experimental results on several corpora show that the proposed approach readily handles very large collections of text documents, infers structured topic representations, and obtains superior test perplexities when compared with related models.

Bayesian optimization is an effective methodology for the global optimization of functions with expensive evaluations. It relies on querying a distribution over functions defined by a relatively cheap surrogate model. An accurate model for this distribution over functions is critical to the effectiveness of the approach, and is typically fit using Gaussian processes (GPs). However, since GPs scale cubically with the number of observations, it has been challenging to handle objectives whose optimization requires many evaluations, and as such, massively parallelizing the optimization. In this work, we explore the use of neural networks as an alternative to GPs to model distributions over functions. We show that performing adaptive basis function regression with a neural network as the parametric form performs competitively with state-of-the-art GP-based approaches, but scales linearly with the number of data rather than cubically. This allows us to achieve a previously intractable degree of parallelism, which we apply to large scale hyperparameter optimization, rapidly finding competitive models on benchmark object recognition tasks using convolutional networks, and image caption generation using neural language models.

Nuclear norm regularization has been shown very promising for pursing a low rank matrix solution in various machine learning problems. Many efforts have been devoted to develop efficient algorithms for solving the optimization problem in nuclear norm regularization. Solving it for large-scale matrix variables, however, is still a challenging task since the complexity grows fast with the size of matrix variable. In this work, we propose a novel method called safe subspace screening (SSS), to improve the efficiency of the solver for nuclear norm regularized least squares problems. Motivated by the fact that the low rank solution can be represented by a few subspaces, the proposed method accurately discards a predominant percentage of inactive subspaces prior to solving the problem to reduce problem size. Consequently, a much smaller problem is required to solve, making it more efficient than optimizing the original problem. The proposed SSS is safe, in that its solution is identical to the solution from the solver. In addition, the proposed SSS can be used together with any existing nuclear norm solver since it is independent of the solver. Extensive results on several synthetic and real data sets show that the proposed SSS is very effective in inactive subspace screening.

Multi-task feature learning (MTFL) is a powerful technique in boosting the predictive performance by learning multiple related classification/regression/clustering tasks simultaneously. However, solving the MTFL problem remains challenging when the feature dimension is extremely large. In this paper, we propose a novel screening rule—that is based on the dual projection onto convex sets (DPC)—to quickly identify the inactive features—that have zero coefficients in the solution vectors across all tasks. One of the appealing features of DPC is that: it is safe in the sense that the detected inactive features are guaranteed to have zero coefficients in the solution vectors across all tasks. Thus, by removing the inactive features from the training phase, we may have substantial savings in the computational cost and memory usage without sacrificing accuracy. To the best of our knowledge, it is the first screening rule that is applicable to sparse models with multiple data matrices. A key challenge in deriving DPC is to solve a nonconvex problem. We show that we can solve for the global optimum efficiently via a properly chosen parametrization of the constraint set. Moreover, DPC has very low computational cost and can be integrated with any existing solvers. We have evaluated the proposed DPC rule on both synthetic and real data sets. The experiments indicate that DPC is very effective in identifying the inactive features—especially for high dimensional data—which leads to a speedup up to several orders of magnitude.

Lifelong reinforcement learning provides a promising framework for developing versatile agents that can accumulate knowledge over a lifetime of experience and rapidly learn new tasks by building upon prior knowledge. However, current lifelong learning methods exhibit non-vanishing regret as the amount of experience increases, and include limitations that can lead to suboptimal or unsafe control policies. To address these issues, we develop a lifelong policy gradient learner that operates in an adversarial setting to learn multiple tasks online while enforcing safety constraints on the learned policies. We demonstrate, for the first time, sublinear regret for lifelong policy search, and validate our algorithm on several benchmark dynamical systems and an application to quadrotor control.

We consider sequential decision problems under uncertainty, where we seek to optimize an unknown function from noisy samples. This requires balancing exploration (learning about the objective) and exploitation (localizing the maximum), a problem well-studied in the multi-armed bandit literature. In many applications, however, we require that the sampled function values exceed some prespecified "safety" threshold, a requirement that existing algorithms fail to meet. Examples include medical applications where patient comfort must be guaranteed, recommender systems aiming to avoid user dissatisfaction, and robotic control, where one seeks to avoid controls causing physical harm to the platform. We tackle this novel, yet rich, set of problems under the assumption that the unknown function satisfies regularity conditions expressed via a Gaussian process prior. We develop an efficient algorithm called SafeOpt, and theoretically guarantee its convergence to a natural notion of optimum reachable under safety constraints. We evaluate SafeOpt on synthetic data, as well as two real applications: movie recommendation, and therapeutic spinal cord stimulation.

We seek to find the robust policy that maximizes the expected cumulative reward for the worst case when a partially observable Markov decision process (POMDP) has uncertain parameters whose values are only known to be in a given region. We prove that the robust value function, which represents the expected cumulative reward that can be obtained with the robust policy, is convex with respect to the belief state. Based on the convexity, we design a value-iteration algorithm for finding the robust policy. We prove that our value iteration converges for an infinite horizon. We also design point-based value iteration for fining the robust policy more efficiency possibly with approximation. Numerical experiments show that our point-based value iteration can adequately find robust policies.

Gaussian vector autoregressive (VAR) processes have been extensively studied in the literature. However, Gaussian assumptions are stringent for heavy-tailed time series that frequently arises in finance and economics. In this paper, we develop a unified framework for modeling and estimating heavy-tailed VAR processes. In particular, we generalize the Gaussian VAR model by an elliptical VAR model that naturally accommodates heavy-tailed time series. Under this model, we develop a quantile-based robust estimator for the transition matrix of the VAR process. We show that the proposed estimator achieves parametric rates of convergence in high dimensions. This is the first work in analyzing heavy-tailed high dimensional VAR processes. As an application of the proposed framework, we investigate Granger causality in the elliptical VAR process, and show that the robust transition matrix estimator induces sign-consistent estimators of Granger causality. The empirical performance of the proposed methodology is demonstrated by both synthetic and real data. We show that the proposed estimator is robust to heavy tails, and exhibit superior performance in stock price prediction.

Common statistical practice has shown that the full power of Bayesian methods is not realized until hierarchical priors are used, as these allow for greater “robustness” and the ability to “share statistical strength.” Yet it is an ongoing challenge to provide a learning-theoretically sound formalism of such notions that: offers practical guidance concerning when and how best to utilize hierarchical models; provides insights into what makes for a good hierarchical prior; and, when the form of the prior has been chosen, can guide the choice of hyperparameter settings. We present a set of analytical tools for understanding hierarchical priors in both the online and batch learning settings. We provide regret bounds under log-loss, which show how certain hierarchical models compare, in retrospect, to the best single model in the model class. We also show how to convert a Bayesian log-loss regret bound into a Bayesian risk bound for any bounded loss, a result which may be of independent interest. Risk and regret bounds for Student’s t and hierarchical Gaussian priors allow us to formalize the concepts of “robustness” and “sharing statistical strength.” Priors for feature selection are investigated as well. Our results suggest that the learning-theoretic benefits of using hierarchical priors can often come at little cost on practical problems.

We describe a method for removing the effect of confounders in order to reconstruct a latent quantity of interest. The method, referred to as