DEUP: Direct Epistemic Uncertainty Prediction

Moksh Jain, Salem Lahlou, Hadi Nekoei, Victor I Butoi, Paul Bertin, Jarrid Rector-Brooks, Maksym Korablyov, Yoshua Bengio

Keywords: deep learning, uncertainty estimation

Abstract: Epistemic uncertainty is the part of out-of-sample prediction error due to the lack of knowledge of the learner. Whereas previous work was focusing on model variance, we propose a principled approach for directly estimating epistemic uncertainty by learning to predict generalization error and subtracting an estimate of aleatoric uncertainty, i.e., intrinsic unpredictability. This estimator of epistemic uncertainty includes the effect of model bias (or misspecification) and is useful in interactive learning environments arising in active learning or reinforcement learning. In addition to discussing these properties of Direct Epistemic Uncertainty Prediction (DEUP), we illustrate its advantage against existing methods for uncertainty estimation on downstream tasks including sequential model optimization and reinforcement learning. We also evaluate the quality of uncertainty estimates from DEUP for probabilistic classification of images and for estimating uncertainty about synergistic drug combinations.

One-sentence Summary: We propose training a separate predictor to learn the out-of-sample error at a given input, providing a more precise estimate of epistemic uncertainty, robust to the existence of misspecification.

Community Implementations: [ 3 code implementations](https://www.catalyzex.com/paper/arxiv:2102.08501/code)