ICML 2018 Papers

We study the application of the Thompson sampling (TS) methodology to the stochastic combinatorial multi-armed bandit (CMAB) framework. We analyze the standard TS algorithm for the general CMAB, and obtain the first distribution-dependent regret bound of $O(m\log T / \Delta_{\min}) $ for TS under general CMAB, where $m$ is the number of arms, $T$ is the time horizon, and $\Delta_{\min}$ is the minimum gap between the expected reward of the optimal solution and any non-optimal solution. We also show that one cannot use an approximate oracle in TS algorithm for even MAB problems. Then we expand the analysis to matroid bandit, a special case of CMAB and for which we could remove the independence assumption across arms and achieve a better regret bound. Finally, we use some experiments to show the comparison of regrets of CUCB and CTS algorithms.

Sparse Subspace Clustering (SSC) is a popular unsupervised machine learning method for clustering data lying close to an unknown union of low-dimensional linear subspaces; a problem with numerous applications in pattern recognition and computer vision. Even though the behavior of SSC for complete data is by now well-understood, little is known about its theoretical properties when applied to data with missing entries. In this paper we give theoretical guarantees for SSC with incomplete data, and provide theoretical evidence that projecting the zero-filled data onto the observation pattern of the point being expressed can lead to substantial improvement in performance; a phenomenon already known experimentally. The main insight of our analysis is that even though this projection induces additional missing entries, this is counterbalanced by the fact that the projected and zero-filled data are in effect incomplete points associated with the union of the corresponding projected subspaces, with respect to which the point being expressed is complete. The significance of this phenomenon potentially extends to the entire class of self-expressive methods.

Recently, a unified model for image-to-image translation tasks within adversarial learning framework has aroused widespread research interests in computer vision practitioners. Their reported empirical success however lacks solid theoretical interpretations for its inherent mechanism. In this paper, we reformulate their model from a brand-new geometrical perspective and have eventually reached a full interpretation on some interesting but unclear empirical phenomenons from their experiments. Furthermore, by extending the definition of generalization for generative adversarial nets to a broader sense, we have derived a condition to control the generalization capability of their model. According to our derived condition, several practical suggestions have also been proposed on model design and dataset construction as a guidance for further empirical researches.

We study the complexity of the entire regularization path for least squares regression with 1-norm penalty, known as the Lasso. Every regression parameter in the Lasso changes linearly as a function of the regularization value. The number of changes is regarded as the Lasso’s complexity. Experimental results using exact path following exhibit polynomial complexity of the Lasso in the problem size. Alas, the path complexity of the Lasso on artificially designed regression problems is exponential We use smoothed analysis as a mechanism for bridging the gap between worst case settings and the de facto low complexity. Our analysis assumes that the observed data has a tiny amount of intrinsic noise. We then prove that the Lasso’s complexity is polynomial in the problem size.

We propose new positive definite kernels for permutations. First we introduce a weighted version of the Kendall kernel, which allows to weight unequally the contributions of different item pairs in the permutations depending on their ranks. Like the Kendall kernel, we show that the weighted version is invariant to relabeling of items and can be computed efficiently in O(n ln(n)) operations, where n is the number of items in the permutation. Second, we propose a supervised approach to learn the weights by jointly optimizing them with the function estimated by a kernel machine. Third, while the Kendall kernel considers pairwise comparison between items, we extend it by considering higher-order comparisons among tuples of items and show that the supervised approach of learning the weights can be systematically generalized to higher-order permutation kernels.

We consider the exploration/exploitation problem in reinforcement learning. For exploitation, it is well known that the Bellman equation connects the value at any time-step to the expected value at subsequent time-steps. In this paper we consider a similar uncertainty Bellman equation (UBE), which connects the uncertainty at any time-step to the expected uncertainties at subsequent time-steps, thereby extending the potential exploratory benefit of a policy beyond individual time-steps. We prove that the unique fixed point of the UBE yields an upper bound on the variance of the posterior distribution of the Q-values induced by any policy. This bound can be much tighter than traditional count-based bonuses that compound standard deviation rather than variance. Importantly, and unlike several existing approaches to optimism, this method scales naturally to large systems with complex generalization. Substituting our UBE-exploration strategy for $\epsilon$-greedy improves DQN performance on 51 out of 57 games in the Atari suite.

In this paper we aim to formally explain the phenomenon of fast convergence of Stochastic Gradient Descent (SGD) observed in modern machine learning. The key observation is that most modern learning architectures are over-parametrized and are trained to interpolate the data by driving the empirical loss (classification and regression) close to zero. While it is still unclear why these interpolated solutions perform well on test data, we show that these regimes allow for fast convergence of SGD, comparable in number of iterations to full gradient descent. For convex loss functions we obtain an exponential convergence bound for

We study the loss surface of neural networks equipped with a hinge loss criterion and ReLU or leaky ReLU nonlinearities. Any such network defines a piecewise multilinear form in parameter space. By appealing to harmonic analysis we show that all local minima of such network are non-differentiable, except for those minima that occur in a region of parameter space where the loss surface is perfectly flat. Non-differentiable minima are therefore not technicalities or pathologies; they are heart of the problem when investigating the loss of ReLU networks. As a consequence, we must employ techniques from nonsmooth analysis to study these loss surfaces. We show how to apply these techniques in some illustrative cases.

Policy gradient methods are a widely used class of model-free reinforcement learning algorithms where a state-dependent baseline is used to reduce gradient estimator variance. Several recent papers extend the baseline to depend on both the state and action and suggest that this significantly reduces variance and improves sample efficiency without introducing bias into the gradient estimates. To better understand this development, we decompose the variance of the policy gradient estimator and numerically show that learned state-action-dependent baselines do not in fact reduce variance over a state-dependent baseline in commonly tested benchmark domains. We confirm this unexpected result by reviewing the open-source code accompanying these prior papers, and show that subtle implementation decisions cause deviations from the methods presented in the papers and explain the source of the previously observed empirical gains. Furthermore, the variance decomposition highlights areas for improvement, which we demonstrate by illustrating a simple change to the typical value function parameterization that can significantly improve performance.

The cornerstone underpinning deep learning is the guarantee that gradient descent on an objective converges to local minima. Unfortunately, this guarantee fails in settings, such as generative adversarial nets, where there are multiple interacting losses. The behavior of gradient-based methods in games is not well understood – and is becoming increasingly important as adversarial and multi-objective architectures proliferate. In this paper, we develop new techniques to understand and control the dynamics in general games. The key result is to decompose the second-order dynamics into two components. The first is related to potential games, which reduce to gradient descent on an implicit function; the second relates to Hamiltonian games, a new class of games that obey a conservation law, akin to conservation laws in classical mechanical systems. The decomposition motivates Symplectic Gradient Adjustment (SGA), a new algorithm for finding stable fixed points in general games. Basic experiments show SGA is competitive with recently proposed algorithms for finding local Nash equilibria in GANs – whilst at the same time being applicable to – and having guarantees in – much more general games.

We present a comprehensive understanding of three important problems in PAC preference learning: maximum selection (maxing), ranking, and estimating

A common problem in Machine Learning and statistics consists in detecting whether the current sample in a stream of data belongs to the same distribution as previous ones, is an isolated outlier or inaugurates a new distribution of data. We present a hierarchical Bayesian algorithm that aims at learning a time-specific approximate posterior distribution of the parameters describing the distribution of the data observed. We derive the update equations of the variational parameters of the approximate posterior at each time step for models from the exponential family, and show that these updates find interesting correspondents in Reinforcement Learning (RL). In this perspective, our model can be seen as a hierarchical RL algorithm that learns a posterior distribution according to a certain stability confidence that is, in turn, learned according to its own stability confidence. Finally, we show some applications of our generic model, first in a RL context, next with an adaptive Bayesian Autoregressive model, and finally in the context of Stochastic Gradient Descent optimization.

While machine learning is going through an era of celebrated success, concerns have been raised about the vulnerability of its backbone: stochastic gradient descent (SGD). Recent approaches have been proposed to ensure the robustness of distributed SGD against adversarial (Byzantine) workers sending

This is a theoretical study on the sample complexity of dictionary learning with a general type of reconstruction loss. The goal is to estimate a $m \times d$ matrix $D$ of unit-norm columns when the only available information is a set of training samples. Points $x$ in $\mathbb{R}^m$ are subsequently approximated by the linear combination $Da$ after solving the problem $\min_{a \in \mathbb{R}^d}  \Phi(x - Da) + g(a)$; function $g:\mathbb{R}^d \to [0,+\infty)$ is either an indicator function or a sparsity promoting regularizer. Here is considered the case where $ \Phi(x) = \inf_{z \in \mathbb{R}^m} { ||x-z||_2^2 + h(||z||_2)}$ and $h$ is an even and univariate function on the real line. Connections are drawn between $\Phi$ and the Moreau envelope of $h$. A new sample complexity result concerning the $k$-sparse dictionary problem removes the spurious condition on the coherence of $D$ appearing in previous works. Finally, comments are made on the approximation error of certain families of losses. The derived generalization bounds are of order $\mathcal{O}(\sqrt{\log n /n})$ and valid without any further restrictions on the set of dictionaries with unit-norm columns.

We study the hypothesis testing problem of inferring the existence of combinatorial structures in undirected graphical models. Although there exist extensive studies on the information-theoretic limits of this problem, it remains largely unexplored whether such limits can be attained by efficient algorithms. In this paper, we quantify the minimum computational complexity required to attain the information-theoretic limits based on an oracle computational model. We prove that, for testing common combinatorial structures, such as clique, nearest neighbor graph and perfect matching, against an empty graph, or large clique against small clique, the information-theoretic limits are provably unachievable by tractable algorithms in general. More importantly, we define structural quantities called the weak and strong edge densities, which offer deep insight into the existence of such computational-statistical tradeoffs. To the best of our knowledge, our characterization is the first to identify and explain the fundamental tradeoffs between statistics and computation for combinatorial inference problems in undirected graphical models.

Understanding the learning dynamics of neural networks is one of the key issues for the improvement of optimization algorithms as well as for the theoretical comprehension of why deep neural nets work so well today. In this paper, we introduce a random matrix-based framework to analyze the learning dynamics of a single-layer linear network on a binary classification problem, for data of simultaneously large dimension and size, trained by gradient descent. Our results provide rich insights into common questions in neural nets, such as overfitting, early stopping and the initialization of training, thereby opening the door for future studies of more elaborate structures and models appearing in today’s neural networks.

Sparsity is a basic property of real vectors that is exploited in a wide variety of machine learning applications. In this work, we describe property testing algorithms for sparsity that observe a low-dimensional projec- tion of the input. We consider two settings. In the first setting, we test sparsity with respect to an unknown basis: given input vectors $y_1 ,...,y_p \in R^d$ whose concatenation as columns forms $Y \in R^{d \times p}$ , does $Y = AX$ for matrices $A \in R^{d\times m}$ and $X \in R^{m \times p}$ such that each column of $X$ is $k$-sparse, or is $Y$ “far” from having such a decomposition? In the second setting, we test sparsity with respect to a known basis: for a fixed design ma- trix $A \in R^{d \times m}$ , given input vector $y \in R^d$ , is $y = Ax$ for some $k$-sparse vector $x$ or is $y$ “far” from having such a decomposition? We analyze our algorithms using tools from high-dimensional geometry and probability.

We propose temporal Poisson square root graphical models (TPSQRs), a generalization of Poisson square root graphical models (PSQRs) specifically designed for modeling longitudinal event data. By estimating the temporal relationships for all possible pairs of event types, TPSQRs can offer a holistic perspective about whether the occurrences of any given event type could excite or inhibit any other type. A TPSQR is learned by estimating a collection of interrelated PSQRs that share the same template parameterization. These PSQRs are estimated jointly in a pseudo-likelihood fashion, where Poisson pseudo-likelihood is used to approximate the original more computationally intensive pseudo-likelihood problem stemming from PSQRs. Theoretically, we demonstrate that under mild assumptions, the Poisson pseudolikelihood approximation is sparsistent for recovering the underlying PSQR. Empirically, we learn TPSQRs from a real-world large-scale electronic health record (EHR) with millions of drug prescription and condition diagnosis events, for adverse drug reaction (ADR) detection. Experimental results demonstrate that the learned TPSQRs can recover ADR signals from the EHR effectively and efficiently.

Generative adversarial networks (GANs) have been shown to produce realistic samples from high-dimensional distributions, but training them is considered hard. A possible explanation for training instabilities is the inherent imbalance between the networks: While the discriminator is trained directly on both real and fake samples, the generator only has control over the fake samples it produces since the real data distribution is fixed by the choice of a given dataset. We propose a simple modification that gives the generator control over the real samples which leads to a tempered learning process for both generator and discriminator. The real data distribution passes through a lens before being revealed to the discriminator, balancing the generator and discriminator by gradually revealing more detailed features necessary to produce high-quality results. The proposed module automatically adjusts the learning process to the current strength of the networks, yet is generic and easy to add to any GAN variant. In a number of experiments, we show that this can improve quality, stability and/or convergence speed across a range of different GAN architectures (DCGAN, LSGAN, WGAN-GP).