Generative Posterior Networks for Approximately Bayesian Epistemic Uncertainty Estimation

In many real-world problems, there is a limited set of training data, but an abundance of unlabeled data. We propose a new method, Generative Posterior Networks (GPNs), that uses unlabeled data to estimate epistemic uncertainty in high-dimensional problems.