Low-Cost Bayesian Methods for Fixing Neural Networks' Overconfidence

Well-calibrated predictive uncertainty of neural networks—essentially making them know when they do not know—is paramount in safety-critical applications. However, deep neural networks are overconfident in the region both far away and near the training data. In this thesis, we study Bayesian neural networks and their extensions to mitigate this issue. First, we show that being Bayesian, even just at the last layer and in a post-hoc manner via Laplace approximations, helps mitigate overconfidence in deep ReLU classifiers. Then, we provide a cost-effective Gaussian-process extension to ReLU Bayesian neural networks that provides a guarantee that ReLU nets will never be overconfident in the region far from the data. Furthermore, we propose three ways of improving the calibration of general Bayesian neural networks in the regions near the data by (i) refining parametric approximations to the Bayesian neural networks’ posteriors with normalizing flows, (ii) training the uncertainty of Laplace approximations, and (iii) leveraging out-of-distribution data during training. We provide an easy-to-use library, laplace-torch, to facilitate the modern arts of Laplace approximations in deep learning. It gives users a way to turn a standard pre-trained deep net into a Bayesian neural network in a cost-efficient manner