ICML 2015 论文列表

The Alternating Direction Method of Multipliers (ADMM) has been studied for years. Traditional ADMM algorithms need to compute, at each iteration, an (empirical) expected loss function on all training examples, resulting in a computational complexity proportional to the number of training examples. To reduce the complexity, stochastic ADMM algorithms were proposed to replace the expected loss function with a random loss function associated with one uniformly drawn example plus a Bregman divergence term. The Bregman divergence, however, is derived from a simple 2nd-order proximal function, i.e., the half squared norm, which could be a suboptimal choice. In this paper, we present a new family of stochastic ADMM algorithms with optimal 2nd-order proximal functions, which produce a new family of adaptive stochastic ADMM methods. We theoretically prove that the regret bounds are as good as the bounds which could be achieved by the best proximal function that can be chosen in hindsight. Encouraging empirical results on a variety of real-world datasets confirm the effectiveness and efficiency of the proposed algorithms.

Graphical models are widely used in inference problems. In practice, one may construct a single large-scale model to explain a phenomenon of interest, which may be utilized in a variety of settings. The latent variables of interest, which can differ in each setting, may only represent a small subset of all variables. The marginals of variables of interest may change after the addition of measurements at different time points. In such adaptive settings, naive algorithms, such as standard belief propagation (BP), may utilize many unnecessary computations by propagating messages over the entire graph. Here, we formulate an efficient inference procedure, termed adaptive BP (AdaBP), suitable for adaptive inference settings. We show that it gives exact results for trees in discrete and Gaussian Markov Random Fields (MRFs), and provide an extension to Gaussian loopy graphs. We also provide extensions on finding the most likely sequence of the entire latent graph. Lastly, we compare the proposed method to standard BP and to that of (Sumer et al., 2011), which tackles the same problem. We show in synthetic and real experiments that it outperforms standard BP by orders of magnitude and explore the settings that it is advantageous over (Sumer et al., 2011).

While classic machine learning paradigms assume training and test data are generated from the same process, domain adaptation addresses the more realistic setting in which the learner has large quantities of labeled data from some source task but limited or no labeled data from the target task it is attempting to learn. In this work, we give the first formal analysis showing that using active learning for domain adaptation yields a way to address the statistical challenges inherent in this setting. We propose a novel nonparametric algorithm, ANDA, that combines an active nearest neighbor querying strategy with nearest neighbor prediction. We provide analyses of its querying behavior and of finite sample convergence rates of the resulting classifier under covariate shift. Our experiments show that ANDA successfully corrects for dataset bias in multi-class image categorization.

Low-rank tensor learning has many applications in machine learning. A series of batch learning algorithms have achieved great successes. However, in many emerging applications, such as climate data analysis, we are confronted with large-scale tensor streams, which poses significant challenges to existing solution in terms of computational costs and limited response time. In this paper, we propose an online accelerated low-rank tensor learning algorithm (ALTO) to solve the problem. At each iteration, we project the current tensor to the subspace of low-rank tensors in order to perform efficient tensor decomposition, then recover the decomposition of the new tensor. By randomly glancing at additional subspaces, we successfully avoid local optima at negligible extra computational cost. We evaluate our method on two tasks in streaming multivariate spatio-temporal analysis: online forecasting and multi-model ensemble, which shows that our method achieves comparable predictive accuracy with significant boost in run time.

State abstractions are often used to reduce the complexity of model-based reinforcement learning when only limited quantities of data are available. However, choosing the appropriate level of abstraction is an important problem in practice. Existing approaches have theoretical guarantees only under strong assumptions on the domain or asymptotically large amounts of data, but in this paper we propose a simple algorithm based on statistical hypothesis testing that comes with a finite-sample guarantee under assumptions on candidate abstractions. Our algorithm trades off the low approximation error of finer abstractions against the low estimation error of coarser abstractions, resulting in a loss bound that depends only on the quality of the best available abstraction and is polynomial in planning horizon.

Stochastic variational inference allows for fast posterior inference in complex Bayesian models. However, the algorithm is prone to local optima which can make the quality of the posterior approximation sensitive to the choice of hyperparameters and initialization. We address this problem by replacing the natural gradient step of stochastic varitional inference with a trust-region update. We show that this leads to generally better results and reduced sensitivity to hyperparameters. We also describe a new strategy for variational inference on streaming data and show that here our trust-region method is crucial for getting good performance.

We describe a novel non-parametric statistical hypothesis test of relative dependence between a source variable and two candidate target variables. Such a test enables us to determine whether one source variable is significantly more dependent on a first target variable or a second. Dependence is measured via the Hilbert-Schmidt Independence Criterion (HSIC), resulting in a pair of empirical dependence measures (source-target 1, source-target 2). We test whether the first dependence measure is significantly larger than the second. Modeling the covariance between these HSIC statistics leads to a provably more powerful test than the construction of independent HSIC statistics by sub-sampling. The resulting test is consistent and unbiased, and (being based on U-statistics) has favorable convergence properties. The test can be computed in quadratic time, matching the computational complexity of standard empirical HSIC estimators. The effectiveness of the test is demonstrated on several real-world problems: we identify language groups from a multilingual corpus, and we prove that tumor location is more dependent on gene expression than chromosomal imbalances. Source code is available for download at https://github.com/wbounliphone/reldep/.

This paper presents a novel unifying framework of anytime sparse Gaussian process regression (SGPR) models that can produce good predictive performance fast and improve their predictive performance over time. Our proposed unifying framework reverses the variational inference procedure to theoretically construct a non-trivial, concave functional that is maximized at the predictive distribution of any SGPR model of our choice. As a result, a stochastic natural gradient ascent method can be derived that involves iteratively following the stochastic natural gradient of the functional to improve its estimate of the predictive distribution of the chosen SGPR model and is guaranteed to achieve asymptotic convergence to it. Interestingly, we show that if the predictive distribution of the chosen SGPR model satisfies certain decomposability conditions, then the stochastic natural gradient is an unbiased estimator of the exact natural gradient and can be computed in constant time (i.e., independent of data size) at each iteration. We empirically evaluate the trade-off between the predictive performance vs. time efficiency of the anytime SGPR models on two real-world million-sized datasets.

We propose a unified framework for making a wide range of PCA-like algorithms – including the standard PCA, sparse PCA and non-negative sparse PCA, etc. – robust when facing a constant fraction of arbitrarily corrupted outliers. Our theoretic analysis establishes solid performance guarantees of the proposed framework: its estimation error is upper bounded by a term depending on the intrinsic parameters of the data model, the selected PCA-like algorithm and the fraction of outliers. Comprehensive experiments on synthetic and real-world datasets demonstrate that the outlier-robust PCA-like algorithms derived from our framework have outstanding performance.

We consider the problem of transferring some a priori knowledge in the context of supervised metric learning approaches. While this setting has been successfully applied in some empirical contexts, no theoretical evidence exists to justify this approach. In this paper, we provide a theoretical justification based on the notion of algorithmic stability adapted to the regularized metric learning setting. We propose an on-average-replace-two-stability model allowing us to prove fast generalization rates when an auxiliary source metric is used to bias the regularizer. Moreover, we prove a consistency result from which we show the interest of considering biased weighted regularized formulations and we provide a solution to estimate the associated weight. We also present some experiments illustrating the interest of the approach in standard metric learning tasks and in a transfer learning problem where few labelled data are available.

We describe and analyze a simple algorithm for principal component analysis and singular value decomposition, VR-PCA, which uses computationally cheap stochastic iterations, yet converges exponentially fast to the optimal solution. In contrast, existing algorithms suffer either from slow convergence, or computationally intensive iterations whose runtime scales with the data size. The algorithm builds on a recent variance-reduced stochastic gradient technique, which was previously analyzed for strongly convex optimization, whereas here we apply it to an inherently non-convex problem, using a very different analysis.

We study the K-armed dueling bandit problem which is a variation of the classical Multi-Armed Bandit (MAB) problem in which the learner receives only relative feedback about the selected pairs of arms. We propose a new algorithm called Relative Exponential-weight algorithm for Exploration and Exploitation (REX3) to handle the adversarial utility-based formulation of this problem. This algorithm is a non-trivial extension of the Exponential-weight algorithm for Exploration and Exploitation (EXP3) algorithm. We prove a finite time expected regret upper bound of order O(sqrt(K ln(K)T)) for this algorithm and a general lower bound of order omega(sqrt(KT)). At the end, we provide experimental results using real data from information retrieval applications.

Matrix spectral methods play an important role in statistics and machine learning, and most often the word ‘matrix’ is dropped as, by default, one assumes that similarities or affinities are measured between two points, thereby resulting in similarity matrices. However, recent challenges in computer vision and text mining have necessitated the use of multi-way affinities in the learning methods, and this has led to a considerable interest in hypergraph partitioning methods in machine learning community. A plethora of “higher-order” algorithms have been proposed in the past decade, but their theoretical guarantees are not well-studied. In this paper, we develop a unified approach for partitioning uniform hypergraphs by means of a tensor trace optimization problem involving the affinity tensor, and a number of existing higher-order methods turn out to be special cases of the proposed formulation. We further propose an algorithm to solve the proposed trace optimization problem, and prove that it is consistent under a planted hypergraph model. We also provide experimental results to validate our theoretical findings.

Multi-task feature selection methods often make the hypothesis that learning tasks share relevant and irrelevant features. However, this hypothesis may be too restrictive in practice. For example, there may be a few tasks with specific relevant and irrelevant features (outlier tasks). Similarly, a few of the features may be relevant for only some of the tasks (outlier features). To account for this, we propose a model for multi-task feature selection based on a robust prior distribution that introduces a set of binary latent variables to identify outlier tasks and outlier features. Expectation propagation can be used for efficient approximate inference under the proposed prior. Several experiments show that a model based on the new robust prior provides better predictive performance than other benchmark methods.

Model selection with cross validation (CV) is very popular in machine learning. However, CV with grid and other common search strategies cannot guarantee to find the model with minimum CV error, which is often the ultimate goal of model selection. Recently, various solution path algorithms have been proposed for several important learning algorithms including support vector classification, Lasso, and so on. However, they still do not guarantee to find the model with minimum CV error.In this paper, we first show that the solution paths produced by various algorithms have the property of piecewise linearity. Then, we prove that a large class of error (or loss) functions are piecewise constant, linear, or quadratic w.r.t. the regularization parameter, based on the solution path. Finally, we propose a new generalized error path algorithm (GEP), and prove that it will find the model with minimum CV error for the entire range of the regularization parameter. The experimental results on a variety of datasets not only confirm our theoretical findings, but also show that the best model with our GEP has better generalization error on the test data, compared to the grid search, manual search, and random search.

We introduce a framework for sparsity structures defined via graphs. Our approach is flexible and generalizes several previously studied sparsity models. Moreover, we provide efficient projection algorithms for our sparsity model that run in nearly-linear time. In the context of sparse recovery, we show that our framework achieves an information-theoretically optimal sample complexity for a wide range of parameters. We complement our theoretical analysis with experiments demonstrating that our algorithms improve on prior work also in practice.

Point process data are commonly observed in fields like healthcare and social science. Designing predictive models for such event streams is an under-explored problem, due to often scarce training data. In this work we propose a multitask point process model, leveraging information from all tasks via a hierarchical Gaussian process (GP). Nonparametric learning functions implemented by a GP, which map from past events to future rates, allow analysis of flexible arrival patterns. To facilitate efficient inference, we propose a sparse construction for this hierarchical model, and derive a variational Bayes method for learning and inference. Experimental results are shown on both synthetic data and an application on real electronic health records.

The Orthant-Wise Limited memory Quasi-Newton (OWL-QN) method has been demonstrated to be very effective in solving the \ell_1-regularized sparse learning problem. OWL-QN extends the L-BFGS from solving unconstrained smooth optimization problems to \ell_1-regularized (non-smooth) sparse learning problems. At each iteration, OWL-QN does not involve any \ell_1-regularized quadratic optimization subproblem and only requires matrix-vector multiplications without an explicit use of the (inverse) Hessian matrix, which enables OWL-QN to tackle large-scale problems efficiently. Although many empirical studies have demonstrated that OWL-QN works quite well in practice, several recent papers point out that the existing convergence proof of OWL-QN is flawed and a rigorous convergence analysis for OWL-QN still remains to be established. In this paper, we propose a modified Orthant-Wise Limited memory Quasi-Newton (mOWL-QN) algorithm by slightly modifying the OWL-QN algorithm. As the main technical contribution of this paper, we establish a rigorous convergence proof for the mOWL-QN algorithm. To the best of our knowledge, our work fills the theoretical gap by providing the first rigorous convergence proof for the OWL-QN-type algorithm on solving \ell_1-regularized sparse learning problems. We also provide empirical studies to show that mOWL-QN works well and is as efficient as OWL-QN.

This paper presents a lower bound for optimizing a finite sum of n functions, where each function is L-smooth and the sum is μ-strongly convex. We show that no algorithm can reach an error εin minimizing all functions from this class in fewer than Ω(n + \sqrtn(κ-1)\log(1/ε)) iterations, where κ=L/μis a surrogate condition number. We then compare this lower bound to upper bounds for recently developed methods specializing to this setting. When the functions involved in this sum are not arbitrary, but based on i.i.d. random data, then we further contrast these complexity results with those for optimal first-order methods to directly optimize the sum. The conclusion we draw is that a lot of caution is necessary for an accurate comparison, and identify machine learning scenarios where the new methods help computationally.