Introduction to Deep Evidential Regression for Uncertainty Quantification

Deep Evidential Regression (DER) is a method for uncertainty quantification in neural networks that enables single-pass estimation of both epistemic and aleatoric uncertainty. Unlike deep ensembles or variational inference, DER trains a network to output parameters of the Normal Inverse Gamma (NIG) distribution, from which uncertainty estimates are derived analytically. The post covers the theory behind epistemic vs. aleatoric uncertainty, the NIG distribution, the evidential loss function, and a PyTorch implementation approximating a cubic function. Limitations include difficulty fully disentangling uncertainty types and sensitivity to the regularization hyperparameter.