Dropout as a Bayesian Approximation: Representing Model Uncertainty in Deep Learning

Deep learning tools have gained tremendous attention in applied machine. However such tools for regression and classification do not capture uncertainty. In comparison, Bayesian models offer a mathematically framework to reason about model uncertainty, but usually come with a computational cost. In this paper we develop a new theoretical casting dropout training in deep neural networks (NNs) as approximate inference in deep Gaussian processes. A direct result of this theory us tools to model uncertainty with dropout NNs -- extracting information existing models that has been thrown away so far. This mitigates the of representing uncertainty in deep learning without sacrificing either complexity or test accuracy. We perform an extensive study of the of dropout's uncertainty. Various network architectures and-linearities are assessed on tasks of regression and classification, using as an example. We show a considerable improvement in predictive-likelihood and RMSE compared to existing state-of-the-art methods, and by using dropout's uncertainty in deep reinforcement learning.