ICML 2015 Papers

We investigate a new Follow the Perturbed Leader (FTPL) algorithm for online structured prediction problems. We show a regret bound which is comparable to the state of the art of FTPL algorithms and is comparable with the best possible regret in some cases. To better understand FTPL algorithms for online structured learning, we present a lower bound on the regret for a large and natural class of FTPL algorithms that use logconcave perturbations. We complete our investigation with an online shortest path experiment and empirically show that our algorithm is both statistically and computationally efficient.

Determinantal point processes (DPPs) offer an elegant tool for encoding probabilities over subsets of a ground set. Discrete DPPs are parametrized by a positive semidefinite matrix (called the DPP kernel), and estimating this kernel is key to learning DPPs from observed data. We consider the task of learning the DPP kernel, and develop for it a surprisingly simple yet effective new algorithm. Our algorithm offers the following benefits over previous approaches: (a) it is much simpler; (b) it yields equally good and sometimes even better local maxima; and (c) it runs an order of magnitude faster on large problems. We present experimental results on both real and simulated data to illustrate the numerical performance of our technique.

Canonical Correlation Analysis (CCA) is a widely used spectral technique for finding correlation structures in multi-view datasets. In this paper, we tackle the problem of large scale CCA, where classical algorithms, usually requiring computing the product of two huge matrices and huge matrix decomposition, are computationally and storage expensive. We recast CCA from a novel perspective and propose a scalable and memory efficient \textitAugmented Approximate Gradient (AppGrad) scheme for finding top k dimensional canonical subspace which only involves large matrix multiplying a thin matrix of width k and small matrix decomposition of dimension k\times k. Further, \textitAppGrad achieves optimal storage complexity O(k(p_1+p_2)), compared with classical algorithms which usually require O(p_1^2+p_2^2) space to store two dense whitening matrices. The proposed scheme naturally generalizes to stochastic optimization regime, especially efficient for huge datasets where batch algorithms are prohibitive. The online property of stochastic \textitAppGrad is also well suited to the streaming scenario, where data comes sequentially. To the best of our knowledge, it is the first stochastic algorithm for CCA. Experiments on four real data sets are provided to show the effectiveness of the proposed methods.

We develop an automated technique for detecting damped Lyman-αabsorbers (DLAs) along spectroscopic sightlines to quasi-stellar objects (QSOs or quasars). The detection of DLAs in large-scale spectroscopic surveys such as SDSS–III is critical to address outstanding cosmological questions, such as the nature of galaxy formation. We use nearly 50000 QSO spectra to learn a tailored Gaussian process model for quasar emission spectra, which we apply to the DLA detection problem via Bayesian model selection. We demonstrate our method’s effectiveness with a large-scale validation experiment on over 100000 spectra, with excellent performance.

Fictitious play is a popular game-theoretic model of learning in games. However, it has received little attention in practical applications to large problems. This paper introduces two variants of fictitious play that are implemented in behavioural strategies of an extensive-form game. The first variant is a full-width process that is realization equivalent to its normal-form counterpart and therefore inherits its convergence guarantees. However, its computational requirements are linear in time and space rather than exponential. The second variant, Fictitious Self-Play, is a machine learning framework that implements fictitious play in a sample-based fashion. Experiments in imperfect-information poker games compare our approaches and demonstrate their convergence to approximate Nash equilibria.

We seek decision rules for \it prediction-time cost reduction, where complete data is available for training, but during prediction-time, each feature can only be acquired for an additional cost. We propose a novel random forest algorithm to minimize prediction error for a user-specified \it average feature acquisition budget. While random forests yield strong generalization performance, they do not explicitly account for feature costs and furthermore require low correlation among trees, which amplifies costs. Our random forest grows trees with low acquisition cost and high strength based on greedy minimax cost-weighted-impurity splits. Theoretically, we establish near-optimal acquisition cost guarantees for our algorithm. Empirically, on a number of benchmark datasets we demonstrate competitive accuracy-cost curves against state-of-the-art prediction-time algorithms.

The cover tree data structure speeds up exact nearest neighbor queries over arbitrary metric spaces. This paper makes cover trees even faster. In particular, we provide (1) a simpler definition of the cover tree that reduces the number of nodes from O(n) to exactly n, (2) an additional invariant that makes queries faster in practice, (3) algorithms for constructing and querying the tree in parallel on multiprocessor systems, and (4) a more cache efficient memory layout. On standard benchmark datasets, we reduce the number of distance computations by 10–50%. On a large-scale bioinformatics dataset, we reduce the number of distance computations by 71%. On a large-scale image dataset, our parallel algorithm with 16 cores reduces tree construction time from 3.5 hours to 12 minutes.

The Frank-Wolfe method (a.k.a. conditional gradient algorithm) for smooth optimization has regained much interest in recent years in the context of large scale optimization and machine learning. A key advantage of the method is that it avoids projections - the computational bottleneck in many applications - replacing it by a linear optimization step. Despite this advantage, the known convergence rates of the FW method fall behind standard first order methods for most settings of interest. It is an active line of research to derive faster linear optimization-based algorithms for various settings of convex optimization. In this paper we consider the special case of optimization over strongly convex sets, for which we prove that the vanila FW method converges at a rate of \frac1t^2. This gives a quadratic improvement in convergence rate compared to the general case, in which convergence is of the order \frac1t, and known to be tight. We show that various balls induced by \ell_p norms, Schatten norms and group norms are strongly convex on one hand and on the other hand, linear optimization over these sets is straightforward and admits a closed-form solution. We further show how several previous fast-rate results for the FW method follow easily from our analysis.

Gaussian processes (GPs) are a flexible class of methods with state of the art performance on spatial statistics applications. However, GPs require O(n^3) computations and O(n^2) storage, and popular GP kernels are typically limited to smoothing and interpolation. To address these difficulties, Kronecker methods have been used to exploit structure in the GP covariance matrix for scalability, while allowing for expressive kernel learning (Wilson et al., 2014). However, fast Kronecker methods have been confined to Gaussian likelihoods. We propose new scalable Kronecker methods for Gaussian processes with non-Gaussian likelihoods, using a Laplace approximation which involves linear conjugate gradients for inference, and a lower bound on the GP marginal likelihood for kernel learning. Our approach has near linear scaling, requiring O(D n^(D+1)/D) operations and O(D n^2/D) storage, for n training data-points on a dense D > 1 dimensional grid. Moreover, we introduce a log Gaussian Cox process, with highly expressive kernels, for modelling spatiotemporal count processes, and apply it to a point pattern (n = 233,088) of a decade of crime events in Chicago. Using our model, we discover spatially varying multiscale seasonal trends and produce highly accurate long-range local area forecasts.

We investigate numerical integration of ordinary differential equations (ODEs) for Hamiltonian Monte Carlo (HMC). High-quality integration is crucial for designing efficient and effective proposals for HMC. While the standard method is leapfrog (Stormer-Verlet) integration, we propose the use of an exponential integrator, which is robust to stiff ODEs with highly-oscillatory components. This oscillation is difficult to reproduce using leapfrog integration, even with carefully selected integration parameters and preconditioning. Concretely, we use a Gaussian distribution approximation to segregate stiff components of the ODE. We integrate this term analytically for stability and account for deviation from the approximation using variation of constants. We consider various ways to derive Gaussian approximations and conduct extensive empirical studies applying the proposed “exponential HMC” to several benchmarked learning problems. We compare to state-of-the-art methods for improving leapfrog HMC and demonstrate the advantages of our method in generating many effective samples with high acceptance rates in short running times.

We provide new concentration inequalities for functions of dependent variables. The work extends that of Janson (2004), which proposes concentration inequalities using a combination of the Laplace transform and the idea of fractional graph coloring, as well as many works that derive concentration inequalities using the entropy method (see, e.g., (Boucheron et al., 2003)). We give inequalities for fractionally sub-additive and fractionally self-bounding functions. In the way, we prove a new Talagrand concentration inequality for fractionally sub-additive functions of dependent variables. The results allow us to envision the derivation of generalization bounds for various applications where dependent variables naturally appear, such as in bipartite ranking.

Entropy gain is widely used for learning decision trees. However, as we go deeper downward the tree, the examples become rarer and the faithfulness of entropy decreases. Thus, misleading choices and over-fitting may occur and the tree has to be adjusted by using an early-stop criterion or post pruning algorithms. However, these methods still depends on the choices previously made, which may be unsatisfactory. We propose a new cumulative entropy function based on confidence intervals on frequency estimates that together considers the entropy of the probability distribution and the uncertainty around the estimation of its parameters. This function takes advantage of the ability of a possibility distribution to upper bound a family of probabilities previously estimated from a limited set of examples and of the link between possibilistic specificity order and entropy. The proposed measure has several advantages over the classical one. It performs significant choices of split and provides a statistically relevant stopping criterion that allows the learning of trees whose size is well-suited w.r.t. the available data. On the top of that, it also provides a reasonable estimator of the performances of a decision tree. Finally, we show that it can be used for designing a simple and efficient online learning algorithm.

Graph smoothness objectives have achieved great success in semi-supervised learning but have not yet been applied extensively to unsupervised generative models. We define a new class of entropic graph-based posterior regularizers that augment a probabilistic model by encouraging pairs of nearby variables in a regularization graph to have similar posterior distributions. We present a three-way alternating optimization algorithm with closed-form updates for performing inference on this joint model and learning its parameters. This method admits updates linear in the degree of the regularization graph, exhibits monotone convergence and is easily parallelizable. We are motivated by applications in computational biology in which temporal models such as hidden Markov models are used to learn a human-interpretable representation of genomic data. On a synthetic problem, we show that our method outperforms existing methods for graph-based regularization and a comparable strategy for incorporating long-range interactions using existing methods for approximate inference. Using genome-scale functional genomics data, we integrate genome 3D interaction data into existing models for genome annotation and demonstrate significant improvements in predicting genomic activity.

In applications of Gaussian processes where quantification of uncertainty is of primary interest, it is necessary to accurately characterize the posterior distribution over covariance parameters. This paper proposes an adaptation of the Stochastic Gradient Langevin Dynamics algorithm to draw samples from the posterior distribution over covariance parameters with negligible bias and without the need to compute the marginal likelihood. In Gaussian process regression, this has the enormous advantage that stochastic gradients can be computed by solving linear systems only. A novel unbiased linear systems solver based on parallelizable covariance matrix-vector products is developed to accelerate the unbiased estimation of gradients. The results demonstrate the possibility to enable scalable and exact (in a Monte Carlo sense) quantification of uncertainty in Gaussian processes without imposing any special structure on the covariance or reducing the number of input vectors.

We introduce Mean-for-Mode estimation, a variant of an uncollapsed Gibbs sampler that we use to train LDA on a GPU. The algorithm combines benefits of both uncollapsed and collapsed Gibbs samplers. Like a collapsed Gibbs sampler — and unlike an uncollapsed Gibbs sampler — it has good statistical performance, and can use sampling complexity reduction techniques such as sparsity. Meanwhile, like an uncollapsed Gibbs sampler — and unlike a collapsed Gibbs sampler — it is embarrassingly parallel, and can use approximate counters.

A stochastic combinatorial semi-bandit is an online learning problem where at each step a learning agent chooses a subset of ground items subject to combinatorial constraints, and then observes stochastic weights of these items and receives their sum as a payoff. In this paper, we consider efficient learning in large-scale combinatorial semi-bandits with linear generalization, and as a solution, propose two learning algorithms called Combinatorial Linear Thompson Sampling (CombLinTS) and Combinatorial Linear UCB (CombLinUCB). Both algorithms are computationally efficient as long as the offline version of the combinatorial problem can be solved efficiently. We establish that CombLinTS and CombLinUCB are also provably statistically efficient under reasonable assumptions, by developing regret bounds that are independent of the problem scale (number of items) and sublinear in time. We also evaluate CombLinTS on a variety of problems with thousands of items. Our experiment results demonstrate that CombLinTS is scalable, robust to the choice of algorithm parameters, and significantly outperforms the best of our baselines.

In many machine learning applications the quality of the data is limited by resource constraints (may it be power, bandwidth, memory, ...). In such cases, the constraints are on the average resources allocated, therefore there is some control over each sample’s quality. In most cases this option remains unused and the data’s quality is uniform over the samples. In this paper we propose to actively allocate resources to each sample such that resources are used optimally overall. We propose a method to compute the optimal resource allocation. We further derive generalization bounds for the case where the problem’s model is unknown. We demonstrate the potential benefit of this approach on both simulated and real-life problems.

The Nyström method has been one of the most effective techniques for kernel-based approach that scales well to large data sets. Since its introduction, there has been a large body of work that improves the approximation accuracy while maintaining computational efficiency. In this paper, we present a novel Nyström method that improves both accuracy and efficiency based on a new theoretical analysis. We first provide a generalized sampling scheme, CAPS, that minimizes a novel error bound based on the subspace distance. We then present our double Nyström method that reduces the size of the decomposition in two stages. We show that our method is highly efficient and accurate compared to other state-of-the-art Nyström methods by evaluating them on a number of real data sets.

The rank aggregation problem has been studied with varying desiderata in varied communities such as Theoretical Computer Science, Statistics, Information Retrieval and Social Welfare Theory. We introduce a variant of this problem we call distributional rank aggregation, where the ranking data is only available via the induced distribution over the set of all permutations. We provide a novel translation of the usual social welfare theory axioms to this setting. As we show this allows for a more quantitative characterization of these axioms: which then are not only less prone to misinterpretation, but also allow simpler proofs for some key impossibility theorems. Most importantly, these quantitative characterizations lead to natural and novel relaxations of these axioms, which as we show, allow us to get around celebrated impossibility results in social choice theory. We are able to completely characterize the class of positional scoring rules with respect to our axioms and show that Borda Count is optimal in a certain sense.