What Uncertainties Do We Need in Bayesian Deep Learning for Computer Vision?

June 2019

tl;dr: A new DL framework to deal with two uncertainties (aleatoric and epistemic) simultaneously. It learns the mapping from the input data to aleatoric uncertainty from the input data, without the need for explicit “uncertainty labels”.

Overall impression

The paper summarizes nicely existing literature in the field about aleatoric and epistemic uncertainty. This is indeed an important area that can contribute to explainable AI.

Key ideas

Technical details