Multivariate Bayesian Last Layer for Regression with Uncertainty Quantification and Decomposition

TL;DR

This paper introduces Bayesian last layer models for multivariate regression that quantify uncertainty and distinguish between aleatoric and epistemic sources, improving neural network interpretability and robustness.

Contribution

It proposes novel Bayesian last layer models with EM algorithms for multivariate regression, enabling uncertainty quantification and uncertainty decomposition in neural networks.

Findings

Effective uncertainty quantification with a single forward pass

Disentangles aleatoric and epistemic uncertainty

Enhances deep neural networks with uncertainty-aware capabilities

Abstract

We present new Bayesian Last Layer neural network models in the setting of multivariate regression under heteroscedastic noise, and propose EM algorithms for parameter learning. Bayesian modeling of a neural network's final layer has the attractive property of uncertainty quantification with a single forward pass. The proposed framework is capable of disentangling the aleatoric and epistemic uncertainty, and can be used to enhance a canonically trained deep neural network with uncertainty-aware capabilities.