NeurIPS 2019 Papers

We propose SWA-Gaussian (SWAG), a simple, scalable, and general purpose approach for uncertainty representation and calibration in deep learning. Stochastic Weight Averaging (SWA), which computes the first moment of stochastic gradient descent (SGD) iterates with a modified learning rate schedule, has recently been shown to improve generalization in deep learning. With SWAG, we fit a Gaussian using the SWA solution as the first moment and a low rank plus diagonal covariance also derived from the SGD iterates, forming an approximate posterior distribution over neural network weights; we then sample from this Gaussian distribution to perform Bayesian model averaging. We empirically find that SWAG approximates the shape of the true posterior, in accordance with results describing the stationary distribution of SGD iterates. Moreover, we demonstrate that SWAG performs well on a wide variety of tasks, including out of sample detection, calibration, and transfer learning, in comparison to many popular alternatives including variational inference, MC dropout, KFAC Laplace, and temperature scaling.

Learning to detect content-independent transformations from data is one of the central problems in biological and artificial intelligence. An example of such problem is unsupervised learning of a visual motion detector from pairs of consecutive video frames. Rao and Ruderman formulated this problem in terms of learning infinitesimal transformation operators (Lie group generators) via minimizing image reconstruction error. Unfortunately, it is difficult to map their model onto a biologically plausible neural network (NN) with local learning rules. Here we propose a biologically plausible model of motion detection. We also adopt the transformation-operator approach but, instead of reconstruction-error minimization, start with a similarity-preserving objective function. An online algorithm that optimizes such an objective function naturally maps onto an NN with biologically plausible learning rules. The trained NN recapitulates major features of the well-studied motion detector in the fly. In particular, it is consistent with the experimental observation that local motion detectors combine information from at least three adjacent pixels, something that contradicts the celebrated Hassenstein-Reichardt model.

Due to the foveated nature of the human vision system, people can focus their visual attention on a small region of their visual field at a time, which usually contains only a single object. Estimating this object of attention in first-person (egocentric) videos is useful for many human-centered real-world applications such as augmented reality applications and driver assistance systems. A straightforward solution for this problem is to pick the object whose bounding box is hit by the gaze, where eye gaze point estimation is obtained from a traditional eye gaze estimator and object candidates are generated from an off-the-shelf object detector. However, such an approach can fail because it addresses the where and the what problems separately, despite that they are highly related, chicken-and-egg problems. In this paper, we propose a novel unified model that incorporates both spatial and temporal evidence in identifying as well as locating the attended object in firstperson videos. It introduces a novel Self Validation Module that enforces and leverages consistency of the where and the what concepts. We evaluate on two public datasets, demonstrating that Self Validation Module significantly benefits both training and testing and that our model outperforms the state-of-the-art.

Dynamic mechanisms offer powerful techniques to improve on both revenue and efficiency by linking sequential auctions using state information, but these techniques rely on exact distributional information of the buyers’ valuations (present and future), which limits their use in learning settings. In this paper, we consider the problem of contextual auctions where the seller gradually learns a model of the buyer's valuation as a function of the context (e.g., item features) and seeks a pricing policy that optimizes revenue. Building on the concept of a bank account mechanism---a special class of dynamic mechanisms that is known to be revenue-optimal---we develop a non-clairvoyant dynamic mechanism that is robust to both estimation errors in the buyer's value distribution and strategic behavior on the part of the buyer. We then tailor its structure to achieve a policy with provably low regret against a constant approximation of the optimal dynamic mechanism in contextual auctions. Our result substantially improves on previous results that only provide revenue guarantees against static benchmarks.

We propose and study a general framework for regularized Markov decision processes (MDPs) where the goal is to find an optimal policy that maximizes the expected discounted total reward plus a policy regularization term. The extant entropy-regularized MDPs can be cast into our framework. Moreover, under our framework, many regularization terms can bring multi-modality and sparsity, which are potentially useful in reinforcement learning. In particular, we present sufficient and necessary conditions that induce a sparse optimal policy. We also conduct a full mathematical analysis of the proposed regularized MDPs, including the optimality condition, performance error, and sparseness control. We provide a generic method to devise regularization forms and propose off-policy actor critic algorithms in complex environment settings. We empirically analyze the numerical properties of optimal policies and compare the performance of different sparse regularization forms in discrete and continuous environments.

In this paper, we formulate the forest representation learning approach called \textsc{CasDF} as an additive model which boosts the augmented feature instead of the prediction. We substantially improve the upper bound of the generalization gap from $\mathcal{O}(\sqrt{\ln m/m})$ to $\mathcal{O}(\ln m/m)$, while the margin ratio of the margin standard deviation to the margin mean is sufficiently small. This tighter upper bound inspires us to optimize the ratio. Therefore, we design a margin distribution reweighting approach for deep forest to achieve a small margin ratio by boosting the augmented feature. Experiments confirm the correlation between the margin distribution and generalization performance. We remark that this study offers a novel understanding of \textsc{CasDF} from the perspective of the margin theory and further guides the layer-by-layer forest representation learning.

Generative models produce realistic objects in many domains, including text, image, video, and audio synthesis. Most popular models—Generative Adversarial Networks (GANs) and Variational Autoencoders (VAEs)—usually employ a standard Gaussian distribution as a prior. Previous works show that the richer family of prior distributions may help to avoid the mode collapse problem in GANs and to improve the evidence lower bound in VAEs. We propose a new family of prior distributions—Tensor Ring Induced Prior (TRIP)—that packs an exponential number of Gaussians into a high-dimensional lattice with a relatively small number of parameters. We show that these priors improve Fréchet Inception Distance for GANs and Evidence Lower Bound for VAEs. We also study generative models with TRIP in the conditional generation setup with missing conditions. Altogether, we propose a novel plug-and-play framework for generative models that can be utilized in any GAN and VAE-like architectures.

Since the introduction of Generative Adversarial Networks (GANs) and Variational Autoencoders (VAE), the literature on generative modelling has witnessed an overwhelming resurgence. The impressive, yet elusive empirical performance of GANs has lead to the rise of many GAN-VAE hybrids, with the hopes of GAN level performance and additional benefits of VAE, such as an encoder for feature reduction, which is not offered by GANs. Recently, the Wasserstein Autoencoder (WAE) was proposed, achieving performance similar to that of GANs, yet it is still unclear whether the two are fundamentally different or can be further improved into a unified model. In this work, we study the $f$-GAN and WAE models and make two main discoveries. First, we find that the $f$-GAN and WAE objectives partake in a primal-dual relationship and are equivalent under some assumptions, which then allows us to explicate the success of WAE. Second, the equivalence result allows us to, for the first time, prove generalization bounds for Autoencoder models, which is a pertinent problem when it comes to theoretical analyses of generative models. Furthermore, we show that the WAE objective is related to other statistical quantities such as the $f$-divergence and in particular, upper bounded by the Wasserstein distance, which then allows us to tap into existing efficient (regularized) optimal transport solvers. Our findings thus present the first primal-dual relationship between GANs and Autoencoder models, comment on generalization abilities and make a step towards unifying these models.

For several problems of interest, there are natural constraints which exist over the output label space. For example, for the joint task of NER and POS labeling, these constraints might specify that the NER label ‘organization’ is consistent only with the POS labels ‘noun’ and ‘preposition’. These constraints can be a great way of injecting prior knowledge into a deep learning model, thereby improving overall performance. In this paper, we present a constrained optimization formulation for training a deep network with a given set of hard constraints on output labels. Our novel approach first converts the label constraints into soft logic constraints over probability distributions outputted by the network. It then converts the constrained optimization problem into an alternating min-max optimization with Lagrangian variables defined for each constraint. Since the constraints are independent of the target labels, our framework easily generalizes to semi-supervised setting. We experiment on the tasks of Semantic Role Labeling (SRL), Named Entity Recognition (NER) tagging, and fine-grained entity typing and show that our constraints not only significantly reduce the number of constraint violations, but can also result in state-of-the-art performance

We consider the problem of computing the maximum likelihood multivariate log-concave distribution for a set of points. Specifically, we present an algorithm which, given $n$ points in $\mathbb{R}^d$ and an accuracy parameter $\eps>0$, runs in time $\poly(n,d,1/\eps),$ and returns a log-concave distribution which, with high probability, has the property that the likelihood of the $n$ points under the returned distribution is at most an additive $\eps$ less than the maximum likelihood that could be achieved via any log-concave distribution. This is the first computationally efficient (polynomial time) algorithm for this fundamental and practically important task. Our algorithm rests on a novel connection with exponential families: the maximum likelihood log-concave distribution belongs to a class of structured distributions which, while not an exponential family, ``locally'' possesses key properties of exponential families. This connection then allows the problem of computing the log-concave maximum likelihood distribution to be formulated as a convex optimization problem, and solved via an approximate first-order method. Efficiently approximating the (sub) gradients of the objective function of this optimization problem is quite delicate, and is the main technical challenge in this work.

This study provides a normative theory for how Bayesian causal inference can be implemented in neural circuits. In both cognitive processes such as causal reasoning and perceptual inference such as cue integration, the nervous systems need to choose different models representing the underlying causal structures when making inferences on external stimuli. In multisensory processing, for example, the nervous system has to choose whether to integrate or segregate inputs from different sensory modalities to infer the sensory stimuli, based on whether the inputs are from the same or different sources. Making this choice is a model selection problem requiring the computation of Bayes factor, the ratio of likelihoods between the integration and the segregation models. In this paper, we consider the causal inference in multisensory processing and propose a novel generative model based on neural population code that takes into account both stimulus feature and stimulus reliability in the inference. In the case of circular variables such as heading direction, our normative theory yields an analytical solution for computing the Bayes factor, with a clear geometric interpretation, which can be implemented by simple additive mechanisms with neural population code. Numerical simulation shows that the tunings of the neurons computing Bayes factor are consistent with the "opposite neurons" discovered in dorsal medial superior temporal (MSTd) and the ventral intraparietal (VIP) areas for visual-vestibular processing. This study illuminates a potential neural mechanism for causal inference in the brain.

In many scientific settings there is a need for adaptive experimental design to guide the process of identifying regions of the search space that contain as many true positives as possible subject to a low rate of false discoveries (i.e. false alarms). Such regions of the search space could differ drastically from a predicted set that minimizes 0/1 error and accurate identification could require very different sampling strategies. Like active learning for binary classification, this experimental design cannot be optimally chosen a priori, but rather the data must be taken sequentially and adaptively in a closed loop. However, unlike classification with 0/1 error, collecting data adaptively to find a set with high true positive rate and low false discovery rate (FDR) is not as well understood. In this paper, we provide the first provably sample efficient adaptive algorithm for this problem. Along the way, we highlight connections between classification, combinatorial bandits, and FDR control making contributions to each.

We construct a new distribution for the simplex using the Kumaraswamy distribution and an ordered stick-breaking process. We explore and develop the theoretical properties of this new distribution and prove that it exhibits symmetry (exchangeability) under the same conditions as the well-known Dirichlet. Like the Dirichlet, the new distribution is adept at capturing sparsity but, unlike the Dirichlet, has an exact and closed form reparameterization--making it well suited for deep variational Bayesian modeling. We demonstrate the distribution's utility in a variety of semi-supervised auto-encoding tasks. In all cases, the resulting models achieve competitive performance commensurate with their simplicity, use of explicit probability models, and abstinence from adversarial training.

Natural images are virtually surrounded by low-density misclassified regions that can be efficiently discovered by gradient-guided search --- enabling the generation of adversarial images. While many techniques for detecting these attacks have been proposed, they are easily bypassed when the adversary has full knowledge of the detection mechanism and adapts the attack strategy accordingly. In this paper, we adopt a novel perspective and regard the omnipresence of adversarial perturbations as a strength rather than a weakness. We postulate that if an image has been tampered with, these adversarial directions either become harder to find with gradient methods or have substantially higher density than for natural images. We develop a practical test for this signature characteristic to successfully detect adversarial attacks, achieving unprecedented accuracy under the white-box setting where the adversary is given full knowledge of our detection mechanism.

We introduce a new notion of the stability of computations, which holds under post-processing and adaptive composition. We show that the notion is both necessary and sufficient to ensure generalization in the face of adaptivity, for any computations that respond to bounded-sensitivity linear queries while providing accuracy with respect to the data sample set. The stability notion is based on quantifying the effect of observing a computation's outputs on the posterior over the data sample elements. We show a separation between this stability notion and previously studied notion and observe that all differentially private algorithms also satisfy this notion.

Reinforcement learning is effective in optimizing policies for recommender systems. Current solutions mostly focus on model-free approaches, which require frequent interactions with a real environment, and thus are expensive in model learning. Offline evaluation methods, such as importance sampling, can alleviate such limitations, but usually request a large amount of logged data and do not work well when the action space is large. In this work, we propose a model-based reinforcement learning solution which models the user-agent interaction for offline policy learning via a generative adversarial network. To reduce bias in the learnt policy, we use the discriminator to evaluate the quality of generated sequences and rescale the generated rewards. Our theoretical analysis and empirical evaluations demonstrate the effectiveness of our solution in identifying patterns from given offline data and learning policies based on the offline and generated data.

Deep generative models are able to suggest new organic molecules by generating strings, trees, and graphs representing their structure. While such models allow one to generate molecules with desirable properties, they give no guarantees that the molecules can actually be synthesized in practice. We propose a new molecule generation model, mirroring a more realistic real-world process, where (a) reactants are selected, and (b) combined to form more complex molecules. More specifically, our generative model proposes a bag of initial reactants (selected from a pool of commercially-available molecules) and uses a reaction model to predict how they react together to generate new molecules. We first show that the model can generate diverse, valid and unique molecules due to the useful inductive biases of modeling reactions. Furthermore, our model allows chemists to interrogate not only the properties of the generated molecules but also the feasibility of the synthesis routes. We conclude by using our model to solve retrosynthesis problems, predicting a set of reactants that can produce a target product.

In this paper we consider the problem of how a reinforcement learning agent that is tasked with solving a sequence of reinforcement learning problems (a sequence of Markov decision processes) can use knowledge acquired early in its lifetime to improve its ability to solve new problems. We argue that previous experience with similar problems can provide an agent with information about how it should explore when facing a new but related problem. We show that the search for an optimal exploration strategy can be formulated as a reinforcement learning problem itself and demonstrate that such strategy can leverage patterns found in the structure of related problems. We conclude with experiments that show the benefits of optimizing an exploration strategy using our proposed framework.

We conduct the first large meta-analysis of overfitting due to test set reuse in the machine learning community. Our analysis is based on over one hundred machine learning competitions hosted on the Kaggle platform over the course of several years. In each competition, numerous practitioners repeatedly evaluated their progress against a holdout set that forms the basis of a public ranking available throughout the competition. Performance on a separate test set used only once determined the final ranking. By systematically comparing the public ranking with the final ranking, we assess how much participants adapted to the holdout set over the course of a competition. Our study shows, somewhat surprisingly, little evidence of substantial overfitting. These findings speak to the robustness of the holdout method across different data domains, loss functions, model classes, and human analysts.