Modeling Epistemic and Aleatoric Uncertainty with Bayesian Neural Networks and Latent Variables
Stefan Depeweg · 2019
In this thesis we develop a novel probabilistic model, the Bayesian neural network with latent variables (BNN+LV). This model class can describe complex stochastic patterns in the data via latent input variables, while, at the same time, account for epistemic uncertainty via a distribution over the network parameters. We investigate the applicability of BNN+LV for decision making, including regression, active learning and reinforcement learning.