Abstract

Estimating uncertainty is important to improving the safety of AI systems. Current methods do not explicitly model predictive uncertainty arising from distributional mismatch between training and test data, either conflating distributional uncertainty with data uncertainty or implicitly modeling it via model uncertainty. Thus, a new framework for modeling predictive uncertainty, called Prior Networks (PNs), is proposed. Prior Networks allow each source of predictive uncertainty – data, distributional and model uncertainty, to be modeled explicitly. This work examines predictive uncertainty in the context of classification. Prior Networks are evaluated on the tasks of identifying out-of-distribution (OOD) samples and detecting misclassification on the MNIST dataset, where they are found to outperform previous methods.