Exploiting Interclass Relationships for Improved Deep Neural Network Generalization in Open-world Environments

Abstract

Deep learning methods informing safety critical decisions must be trusted to make the correct decision in the face of uncertainty. The success of deep neural networks across a variety of academic benchmark tasks has led to their deployment in consequential decision making scenarios, such as autonomous driving. Performance on academic benchmarks, however, is not the only metric we must consider. Reliable uncertainty estimation and model interpretability are necessary for establishing trust in deep learning methods. This thesis develops tools for creating trusted deep learning classifiers in the computer vision domain. We seek to provide the tools and perspectives necessary for designing robust deep learning methods.Concretely, we design machine learning methods capable of detecting samples that the model has never seen before, or are outside of the training distribution. In real-world applications, machine learning systems will inevitably encounter data that is novel to them. The model must detect this data so that it can respond appropriately. This is the subject of out-of-distribution (OOD) detection and open-set recognition research. We introduce two perspectives for solving the OOD detection task. First, we reconsider the OOD detection problem from the hierarchical classification perspective. Second, we examine the problem from the perspective of Bayesian nonparametrics. In our first thrust, we introduce a method for training and calibrating hierarchical classifiers with deep convolutional neural network backbones. We develop a novel training objective for confidence calibration and demonstrate how hierarchical inference can tune the prediction depth based on the model's uncertainty. We show that by adopting a hierarchical classification perspective we convert the binary classification problem of inlier versus outlier to a problem of maximizing the amount of information we can provide while maintaining a certain level of confidence. By making predictions on a tree structure, we enable the analysis of intermediate decisions to interpret, explain, and validate the model's decision process.Our second thrust provides a Bayesian nonparametric perspective on the OOD detection problem. Nonparametric methods are a natural fit for outlier detection tasks as they grow to fit the complexity of the data by explicitly modeling the probability that a sample belongs to a novel cluster. However, they have largely been eschewed in favor of simpler methods. We show a formal relationship between Bayesian nonparametric models and popular existing methods that utilize the Mahalanobis distance as a confidence measure. Building on this relationship, we propose several hierarchical Bayesian extensions showcasing the utility of probabilistic generative modeling applied to deep learning. We conclude by outlining a path forward for both the hierarchical classification and the Bayesian nonparametric approaches. Our work lays the foundations for future studies in the related fields of category discovery which discovers new classes in the data without supervision, and continual learning in which a model must learn new tasks throughout its life. By exploiting the hierarchical relationships between classes and revisiting Bayesian nonparametric methods we provide interpretable and robust deep learning methods for the OOD detection problem.