Quantifying Uncertainty using Bayesian Deep Learning and Deep Ensembles
Rishit Mohan Ahuja, Maxime Alos, Alex McQuilkin, Anudeep Venapally · 2023
Deep Learning algorithms have achieved success in predictive accuracy, but they still have shortcomings that restrict their use in high-risk industries like medicine, finance, etc. Quantifying uncertainty in these situations can help users properly account for the risk of incorrect predictions, which is especially important in high-risk situations. Both individual Bayesian Deep Learning (BDL) models and ensembles of standard deep neural networks have been used to estimate uncertainty. In this project, Bayesian deep learning models were trained and ensembles of those models were implemented that outperform the individual models in both accuracy and uncertainty quantification. Models were tested on out-of-distribution datasets obtained by rotating the existing datasets. This approach was chosen since, in realworld situations, most of the predictions are to be made on data that does not match the distribution of the training data.