SLANG: Fast Structured Covariance Approximations for Bayesian Deep Learning with Natural Gradient

Uncertainty estimation in large deep-learning models is a computationally challenging task, where it is difficult to form even a Gaussian approximation to the posterior distribution. In such situations, existing methods usually resort to a diagonal approximation of the covariance matrix despite the fact that these matrices are known to result in poor uncertainty estimates. To address this issue, we propose a new stochastic, low-rank, approximate natural-gradient (SLANG) method for variational inference in large, deep models. Our method estimates a "diagonal plus low-rank" structure based solely on back-propagated gradients of the network log-likelihood. This requires strictly less gradient computations than methods that compute the gradient of the whole variational objective. Empirical evaluations on standard benchmarks confirm that SLANG enables faster and more accurate estimation of uncertainty than mean-field methods, and performs comparably to state-of-the-art methods.

Mean-Field methods for Structured Deep-Learning in Computer Vision

Pierre Bruno Baqué

In recent years, Machine Learning based Computer Vision techniques made impressive progress. These algorithms proved particularly efficient for image classification or detection of isolated objects. From a probabilistic perspective, these methods can predict marginals, over single or multiple variables, independently, with high accuracy. However, in many tasks of practical interest, we need to predict jointly several correlated variables. Practical applications include people detection in crowded scenes, image segmentation, surface reconstruction, 3D pose estimation and others. A large part of the research effort in today's computer-vision community aims at finding task-specific solutions to these problems, while leveraging the power of Deep-Learning based classifiers. In this thesis, we present our journey towards a generic and practical solution based on mean-field (MF) inference. Mean-field is a Statistical Physics-inspired method which has long been used in Computer-Vision as a variational approximation to posterior distributions over complex Conditional Random Fields. Standard mean-field optimization is based on coordinate descent and in many situations can be impractical. We therefore propose a novel proximal gradient-based approach to optimizing the variational objective. It is naturally parallelizable and easy to implement. We prove its convergence, and then demonstrate that, in practice, it yields faster convergence and often finds better optima than more traditional mean-field optimization techniques. Then, we show that we can replace the fully factorized distribution of mean-field by a weighted mixture of such distributions, that similarly minimizes the KL-Divergence to the true posterior. Our extension of the clamping method proposed in previous works allows us to both produce a more descriptive approximation of the true posterior and, inspired by the diverse MAP paradigms, fit a mixture of mean-field approximations. We demonstrate that this positively impacts real-world algorithms that initially relied on mean-fields. One of the important properties of the mean-field inference algorithms is that the closed-form updates are fully differentiable operations. This naturally allows to do parameter learning by simply unrolling multiple iterations of the updates, the so-called back-mean-field algorithm. We derive a novel and efficient structured learning method for multi-modal posterior distribution based on the Multi-Modal Mean-Field approximation, which can be seamlessly combined to modern gradient-based learning methods such as CNNs. Finally, we explore in more details the specific problem of structured learning and prediction for multiple-people detection in crowded scenes. We then present a mean-field based structured deep-learning detection algorithm that provides state of the art results on this dataset.

EPFL2018

On Vacuous and Non-Vacuous Generalization Bounds for Deep Neural Networks.

Konstantinos Pitas

Deep neural networks have been empirically successful in a variety of tasks, however their theoretical understanding is still poor. In particular, modern deep neural networks have many more parameters than training data. Thus, in principle they should overfit the training samples and exhibit poor generalization to the complete data distribution. Counter intuitively however, they manage to achieve both high training accuracy and high testing accuracy. One can prove generalization using a validation set, however this can be difficult when training samples are limited and at the same time we do not obtain any information about why deep neural networks generalize well. Another approach is to estimate the complexity of the deep neural network. The hypothesis is that if a network with high training accuracy has high complexity it will have memorized the data, while if it has low complexity it will have learned generalizable patterns. In the first part of this thesis we explore Spectral Complexity, a measure of complexity that depends on combinations of norms of the weight matrices of the deep neural network. For a dataset that is difficult to classify, with no underlying model and/or no recurring pattern, for example one where the labels have been chosen randomly, spectral complexity has a large value, reflecting that the network needs to memorize the labels, and will not generalize well. Putting back the real labels, the spectral complexity becomes lower reflecting that some structure is present and the network has learned patterns that might generalize to unseen data. Spectral complexity results in vacuous estimates of the generalization error (the difference between the training and testing accuracy), and we show that it can lead to counterintuitive results when comparing the generalization error of different architectures. In the second part of the thesis we explore non-vacuous estimates of the generalization error. In Chapter 2 we analyze the case of PAC-Bayes where a posterior distribution over the weights of a deep neural network is learned using stochastic variational inference, and the generalization error is the KL divergence between this posterior and a prior distribution. We find that a common approximation where the posterior is constrained to be Gaussian with diagonal covariance, known as the mean-field approximation, limits significantly any gains in bound tightness. We find that, if we choose the prior mean to be the random network initialization, the generalization error estimate tightens significantly. In Chapter 3 we explore an existing approach to learning the prior mean, in PAC-Bayes, from the training set. Specifically, we explore differential privacy, which ensures that the training samples contribute only a limited amount of information to the prior, making it distribution and not training set dependent. In this way the prior should generalize well to unseen data (as it hasn't memorized individual samples) and at the same time any posterior distribution that is close to it in terms of the KL divergence will also exhibit good generalization.

EPFL2020

Ensembles for Sequence Learning

Christos Dimitrakakis

This thesis explores the application of ensemble methods to sequential learning tasks. The focus is on the development and the critical examination of new methods or novel applications of existing methods, with emphasis on supervised and reinforcement learning problems. In both types of problems, even after having observed a certain amount of data, we are often faced with uncertainty as to which hypothesis is correct among all the possible ones. However, in many methods for both supervised and for reinforcement learning problems this uncertainty is ignored, in the sense that there is a single solution selected out of the whole of the hypothesis space. Apart from the classical solution of analytical Bayesian formulations, ensemble methods offer an alternative approach to representing this uncertainty. This is done simply through maintaining a set of alternative hypotheses. The sequential supervised problem considered is that of automatic speech recognition using hidden Markov models. The application of ensemble methods to the problem represents a challenge in itself, since most such methods can not be readily adapted to sequential learning tasks. This thesis proposes a number of different approaches for applying ensemble methods to speech recognition and develops methods for effective training of phonetic mixtures with or without access to phonetic alignment data. Furthermore, the notion of expected loss is introduced for integrating probabilistic models with the boosting approach. In some cases substantial improvements over the baseline system are obtained. In reinforcement learning problems the goal is to act in such a way as to maximise future reward in a given environment. In such problems uncertainty becomes important since neither the environment nor the distribution of rewards that result from each action are known. This thesis presents novel algorithms for acting nearly optimally under uncertainty based on theoretical considerations. Some ensemble-based representations of uncertainty (including a fully Bayesian model) are developed and tested on a few simple tasks resulting in performance comparable with the state of the art. The thesis also draws some parallels between a proposed representation of uncertainty based on gradient-estimates and on ``prioritised sweeping'' and between the application of reinforcement learning to controlling an ensemble of classifiers and classical supervised ensemble learning methods.

École Polytechnique Fédérale de Lausanne2006