ICML 2019 Papers

Knowledge distillation deals with the problem of training a smaller model (

We present Zeno, a technique to make distributed machine learning, particularly Stochastic Gradient Descent (SGD), tolerant to an arbitrary number of faulty workers. Zeno generalizes previous results that assumed a majority of non-faulty nodes; we need assume only one non-faulty worker. Our key idea is to suspect workers that are potentially defective. Since this is likely to lead to false positives, we use a ranking-based preference mechanism. We prove the convergence of SGD for non-convex problems under these scenarios. Experimental results show that Zeno outperforms existing approaches.

We prove that for an $L$-layer fully-connected linear neural network, if the width of every hidden layer is $\widetilde{\Omega}\left(L \cdot r \cdot d_{out} \cdot \kappa^3 \right)$, where $r$ and $\kappa$ are the rank and the condition number of the input data, and $d_{out}$ is the output dimension, then gradient descent with Gaussian random initialization converges to a global minimum at a linear rate. The number of iterations to find an $\epsilon$-suboptimal solution is $O(\kappa \log(\frac{1}{\epsilon}))$. Our polynomial upper bound on the total running time for wide deep linear networks and the $\exp\left(\Omega\left(L\right)\right)$ lower bound for narrow deep linear neural networks [Shamir, 2018] together demonstrate that wide layers are necessary for optimizing deep models.

Empirical evidence suggests that neural networks with ReLU activations generalize better with over-parameterization. However, there is currently no theoretical analysis that explains this observation. In this work, we provide theoretical and empirical evidence that, in certain cases, overparameterized convolutional networks generalize better than small networks because of an interplay between weight clustering and feature exploration at initialization. We demonstrate this theoretically for a 3-layer convolutional neural network with max-pooling, in a novel setting which extends the XOR problem. We show that this interplay implies that with overparamterization, gradient descent converges to global minima with better generalization performance compared to global minima of small networks. Empirically, we demonstrate these phenomena for a 3-layer convolutional neural network in the MNIST task.

Membership inference determines, given a sample and trained parameters of a machine learning model, whether the sample was part of the training set. In this paper, we derive the optimal strategy for membership inference with a few assumptions on the distribution of the parameters. We show that optimal attacks only depend on the loss function, and thus black-box attacks are as good as white-box attacks. As the optimal strategy is not tractable, we provide approximations of it leading to several inference methods, and show that existing membership inference methods are coarser approximations of this optimal strategy. Our membership attacks outperform the state of the art in various settings, ranging from a simple logistic regression to more complex architectures and datasets, such as ResNet-101 and Imagenet.

In standard classification problems, the assumption is that the entity making the decision (the

Importance-weighted risk minimization is a key ingredient in many machine learning algorithms for causal inference, domain adaptation, class imbalance, and off-policy reinforcement learning. While the effect of importance weighting is well-characterized for low-capacity misspecified models, little is known about how it impacts over-parameterized, deep neural networks. This work is inspired by recent theoretical results showing that on (linearly) separable data, deep linear networks optimized by SGD learn weight-agnostic solutions, prompting us to ask, for realistic deep networks, for which many practical datasets are separable, what is the effect of importance weighting? We present the surprising finding that while importance weighting impacts models early in training, its effect diminishes over successive epochs. Moreover, while L2 regularization and batch normalization (but not dropout), restore some of the impact of importance weighting, they express the effect via (seemingly) the wrong abstraction: why should practitioners tweak the L2 regularization, and by how much, to produce the correct weighting effect? Our experiments confirm these findings across a range of architectures and datasets.

We propose a novel model for temporal detection and localization which allows the training of deep neural networks using only counts of event occurrences as training labels. This powerful weakly-supervised framework alleviates the burden of the imprecise and time consuming process of annotating event locations in temporal data. Unlike existing methods, in which localization is explicitly achieved by design, our model learns localization implicitly as a byproduct of learning to count instances. This unique feature is a direct consequence of the model’s theoretical properties. We validate the effectiveness of our approach in a number of experiments (drum hit and piano onset detection in audio, digit detection in images) and demonstrate performance comparable to that of fully-supervised state-of-the-art methods, despite much weaker training requirements.

We consider the problem of detecting the presence of the signal in a rank-one signal-plus-noise data matrix. In case the signal-to-noise ratio is under the threshold below which a reliable detection is impossible, we propose a hypothesis test based on the linear spectral statistics of the data matrix. When the noise is Gaussian, the error of the proposed test is optimal as it matches the error of the likelihood ratio test that minimizes the sum of the Type-I and Type-II errors. The test is data-driven and does not depend on the distribution of the signal or the noise. If the density of the noise is known, it can be further improved by an entrywise transformation to lower the error of the test.

The Wasserstein distance serves as a loss function for unsupervised learning which depends on the choice of a ground metric on sample space. We propose to use the Wasserstein distance itself as the ground metric on the sample space of images. This ground metric is known as an effective distance for image retrieval, that correlates with human perception. We derive the Wasserstein ground metric on pixel space and define a Riemannian Wasserstein gradient penalty to be used in the Wasserstein Generative Adversarial Network (WGAN) framework. The new gradient penalty is computed efficiently via convolutions on the $L^2$ gradients with negligible additional computational cost. The new formulation is more robust to the natural variability of the data and provides for a more continuous discriminator in sample space.

A rapidly growing area of work has studied the existence of adversarial examples, datapoints which have been perturbed to fool a classifier, but the vast majority of these works have focused primarily on threat models defined by $\ell_p$ norm-bounded perturbations. In this paper, we propose a new threat model for adversarial attacks based on the Wasserstein distance. In the image classification setting, such distances measure the cost of moving pixel mass, which can naturally represent “standard” image manipulations such as scaling, rotation, translation, and distortion (and can potentially be applied to other settings as well). To generate Wasserstein adversarial examples, we develop a procedure for approximate projection onto the Wasserstein ball, based upon a modified version of the Sinkhorn iteration. The resulting algorithm can successfully attack image classification models, bringing traditional CIFAR10 models down to 3% accuracy within a Wasserstein ball with radius 0.1 (i.e., moving 10% of the image mass 1 pixel), and we demonstrate that PGD-based adversarial training can improve this adversarial accuracy to 76%. In total, this work opens up a new direction of study in adversarial robustness, more formally considering convex metrics that accurately capture the invariances that we typically believe should exist in classifiers, and code for all experiments in the paper is available at https://github.com/locuslab/projected_sinkhorn.

We investigate the feasibility of learning from both fully-labeled supervised data and contextual bandit data. We specifically consider settings in which the underlying learning signal may be different between these two data sources. Theoretically, we state and prove no-regret algorithms for learning that is robust to divergences between the two sources. Empirically, we evaluate some of these algorithms on a large selection of datasets, showing that our approaches are feasible, and helpful in practice.

Voronoi cell decompositions provide a classical avenue to classification. Typical approaches however only utilize point-wise cell-membership information by means of nearest neighbor queries and do not utilize further geometric information about Voronoi cells since the computation of Voronoi diagrams is prohibitively expensive in high dimensions. We propose a Monte-Carlo integration based approach that instead computes a weighted integral over the boundaries of Voronoi cells, thus incorporating additional information about the Voronoi cell structure. We demonstrate the scalability of our approach in up to 3072 dimensional spaces and analyze convergence based on the number of Monte Carlo samples and choice of weight functions. Experiments comparing our approach to Nearest Neighbors, SVM and Random Forests indicate that while our approach performs similarly to Random Forests for large data sizes, the algorithm exhibits non-trivial data-dependent performance characteristics for smaller datasets and can be analyzed in terms of a geometric confidence measure, thus adding to the repertoire of geometric approaches to classification while having the benefit of not requiring any model changes or retraining as new training samples or classes are added.

Bayesian nonparametric models provide a principled way to automatically adapt the complexity of a model to the amount of the data available, but computation in such models is difficult. Amortized variational approximations are appealing because of their computational efficiency, but current methods rely on a fixed finite truncation of the infinite model. This truncation level can be difficult to set, and also interacts poorly with amortized methods due to the over-pruning problem. Instead, we propose a new variational approximation, based on a method from statistical physics called Russian roulette sampling. This allows the variational distribution to adapt its complexity during inference, without relying on a fixed truncation level, and while still obtaining an unbiased estimate of the gradient of the original variational objective. We demonstrate this method on infinite sized variational auto-encoders using a Beta-Bernoulli (Indian buffet process) prior.

Variational autoencoders employ an amortized inference model to approximate the posterior of latent variables. However, such amortized variational inference faces two challenges: (1) the limited posterior expressiveness of fully-factorized Gaussian assumption and (2) the amortization error of the inference model. We present a novel approach that addresses both challenges. First, we focus on ReLU networks with Gaussian output and illustrate their connection to probabilistic PCA. Building on this observation, we derive an iterative algorithm that finds the mode of the posterior and apply fullcovariance Gaussian posterior approximation centered on the mode. Subsequently, we present a general framework named Variational Laplace Autoencoders (VLAEs) for training deep generative models. Based on the Laplace approximation of the latent variable posterior, VLAEs enhance the expressiveness of the posterior while reducing the amortization error. Empirical results on MNIST, Omniglot, Fashion-MNIST, SVHN and CIFAR10 show that the proposed approach significantly outperforms other recent amortized or iterative methods on the ReLU networks.

Networks provide a natural yet statistically grounded way to depict and understand how a set of entities interact. However, in many situations interactions are not directly observed and the network needs to be reconstructed based on observations collected for each entity. Our work focuses on the situation where these observations consist of counts. A typical example is the reconstruction of an ecological network based on abundance data. In this setting, the abundance of a set of species is collected in a series of samples and/or environments and we aim at inferring direct interactions between the species. The abundances at hand can be, for example, direct counts of individuals (ecology of macro-organisms) or read counts resulting from metagenomic sequencing (microbial ecology). Whatever the approach chosen to infer such a network, it has to account for the peculiaraties of the data at hand. The first, obvious one, is that the data are counts, i.e. non continuous. Also, the observed counts often vary over many orders of magnitude and are more dispersed than expected under a simple model, such as the Poisson distribution. The observed counts may also result from different sampling efforts in each sample and/or for each entity, which hampers direct comparison. Furthermore, because the network is supposed to reveal only direct interactions, it is highly desirable to account for covariates describing the environment to avoid spurious edges. Many methods of network reconstruction from count data have been proposed. In the context of microbial ecology, most methods (SparCC, REBACCA, SPIEC-EASI, gCODA, BanOCC) rely on a two-step strategy: transform the counts to pseudo Gaussian observations using simple transforms before moving back to the setting of Gaussian Graphical Models, for which state of the art methods exist to infer the network, but only in a Gaussian world. In this work, we consider instead a full-fledged probabilistic model with a latent layer where the counts follow Poisson distributions, conditional to latent (hidden) Gaussian correlated variables. In this model, known as Poisson log-normal (PLN), the dependency structure is completely captured by the latent layer and we model counts, rather than transformations thereof. To our knowledge, the PLN framework is quite new and has only been used by two other recent methods (Mint and plnDAG) to reconstruct networks from count data. In this work, we use the same mathematical framework but adopt a different optimization strategy which alleviates the whole optimization process. We also fully exploit the connection between the PLN framework and generalized linear models to account for the peculiarities of microbiological data sets. The network inference step is done as usual by adding sparsity inducing constraints on the inverse covariance matrix of the latent Gaussian vector to select only the most important interactions between species. Unlike the usual Gaussian setting, the penalized likelihood is generally not tractable in this framework. We resort instead to a variational approximation for parameter inference and solve the corresponding optimization problem by alternating a gradient descent on the variational parameters and a graphical-Lasso step on the covariance matrix. We also select the sparsity parameter using the resampling-based StARS procedure. We show that the sparse PLN approach has better performance than existing methods on simulated datasets and that it extracts relevant signal from microbial ecology datasets. We also show that the inference scales to datasets made up of hundred of species and samples, in line with other methods in the field. In short, our contributions to the field are the following: we extend the use of PLN distributions in network inference by (i) accounting for covariates and offset and thus removing some spurious edges induced by confounding factors, (ii) accounting for different sampling effort to integrate data sets from different sources and thus infer interactions between different types of organisms (e.g. bacteria - fungi), (iii) developing an inference procedure based on the iterative optimization of a well defined objective function. Our objective function is a provable lower bound of the observed likelihood and our procedure accounts for the uncertainty associated with the estimation of the latent variable, unlike the algorithm presented in Mint and plnDAG.

We introduce the implicit processes (IPs), a stochastic process that places implicitly defined multivariate distributions over any finite collections of random variables. IPs are therefore highly flexible implicit priors over

The generative adversarial network (GAN) has received considerable attention recently as a model for data synthesis, without an explicit specification of a likelihood function. There has been commensurate interest in leveraging likelihood estimates to improve GAN training. To enrich the understanding of this fast-growing yet almost exclusively heuristic-driven subject, we elucidate the theoretical roots of some of the empirical attempts to stabilize and improve GAN training with the introduction of likelihoods. We highlight new insights from variational theory of diffusion processes to derive a likelihood-based regularizing scheme for GAN training, and present a novel approach to train GANs with an unnormalized distribution instead of empirical samples. To substantiate our claims, we provide experimental evidence on how our theoretically-inspired new algorithms improve upon current practice.

The problem of estimating causal effects of treatments from observational data falls beyond the realm of supervised learning {—} because counterfactual data is inaccessible, we can never observe the true causal effects. In the absence of "supervision", how can we evaluate the performance of causal inference methods? In this paper, we use influence functions {—} the functional derivatives of a loss function {—} to develop a model validation procedure that estimates the estimation error of causal inference methods. Our procedure utilizes a Taylor-like expansion to approximate the loss function of a method on a given dataset in terms of the influence functions of its loss on a "synthesized", proximal dataset with known causal effects. Under minimal regularity assumptions, we show that our procedure is consistent and efficient. Experiments on 77 benchmark datasets show that using our procedure, we can accurately predict the comparative performances of state-of-the-art causal inference methods applied to a given observational study.