Deep Learning Under Uncertainty
Title: Deep Learning Under Uncertainty
DNr: Berzelius-2021-72
Project Type: LiU Berzelius
Principal Investigator: Hossein Azizpour <azizpour@kth.se>
Affiliation: Kungliga Tekniska högskolan
Duration: 2021-10-30 – 2022-03-05
Classification: 10207
Keywords:

Abstract

Deep networks are deployed in many industrial and societal sectors. It is important to increase the trustworthiness of such deployed models. One key aspect of such trust is enabled through modeling and robustness to data and/or predictive uncertainties. Within this project we would like to explore different proof-of-concept techniques (with regards to ensemble techniques, distribution distillation, and robust loss functions and learning algorithms) to equip deep networks with robustness to aleatoric heteroscedastic noise as well as calibrated estimation of epistemic uncertainty due to model misspecification and/or data scarcity.