Learning to Estimate Epistemic Uncertainty in Neural Networks

Katherine Elizabeth Brown, Doug Talbert

Keywords: uncertainty quantification, regression, machine learning, deep learning

Abstract: Epistemic uncertainty quantification provides useful insight into a deep neural network's understanding of the relationship between its training distribution and unseen instances. A Bayesian-based approaches have been shown to quantify this relationship better than softmax probabilities. Unfortunately, however, those approaches to uncertainty quantification require multiple Monte-Carlo samples of a neural network, augmenting the neural network to learn distributions for its weights, or utilizing an ensemble of neural networks. Such extra calculations are problematic in time-critical, resource-limited scenarios such as trauma triage. In this work, we propose a technique that allows epistemic uncertainty to be estimated using learned regression algorithms. We find that this technique, once trained, allows epistemic uncertainty to be effectively and efficiently predicted.

One-sentence Summary: This work proposes a technique that allows epistemic uncertainty to be estimated using learned regression algorithms.