Improving predictions of Bayesian neural networks via local linearization

In this paper we argue that in Bayesian deep learning, the frequently utilized generalized Gauss-Newton (GGN) approximation should be understood as a modification of the underlying probabilistic model and should be considered separately from further approximate inference techniques. Applying the GGN approximation turns a BNN into a locally linearized generalized linear model or, equivalently, a Gaussian process. Because we then use this linearized model for inference, we should also predict using this modified likelihood rather than the original BNN likelihood. This formulation extends previous results to general likelihoods and alleviates underfitting behaviour observed e.g. by Ritter et al. (2018). We demonstrate our approach on several UCI classification datasets as well as CIFAR10.