Bayesian Deep Learning and Uncertainty in Computer Vision

Abstract

Visual data contains rich information about the operating environment of an intelligent robotic system. Extracting this information allows intelligent systems to reason and decide their future actions. Erroneous visual information, therefore, can lead to poor decisions, causing accidents and casualties, especially in a safety-critical application such as automated driving. One way to prevent this is by measuring the level of uncertainty in the visual information interpretation, so that the system knows the reliability degree of the extracted information.

Deep neural networks are now being used in many vision tasks due to their superior accuracy compared to traditional machine learning methods. However, their estimated uncertainties have been shown to be unreliable. To mitigate this issue, researchers have developed methods and tools to apply Bayesian modeling to deep neural networks. This results in a class of models known as Bayesian neural networks, whose uncertainty estimates are more reliable and informative. In this thesis, we make the following contributions in the context of Bayesian Neural Network applied to vision tasks. In particular: - We improve the understanding of visual uncertainty estimates from Bayesian deep models. Specifically, we study the behavior of Bayesian deep models applied to road-scene image segmentation under different factors, such as varying weather, depth, and occlusion levels. - We show the importance of model calibration technique in the context of autonomous driving, which strengthens the reliability of the estimated uncertainty. We demonstrate its effectiveness in a simple object localization task. - We address the high run-time cost of the current Bayesian deep learning techniques. We develop a distillation technique based on the Dirichlet distribution, which allows us to estimate the uncertainties in real-time.