Neural networks provide state-of-the-art performance on a variety of tasks. However, they are often overconfident when making predictions. This inability to properly account for uncertainty limits their application to high-risk decision making, active learning and Bayesian optimisation