Robustness to corruption in pre-trained Bayesian neural networks

TL;DR

This paper introduces ShiftMatch, a novel likelihood method for Bayesian neural networks that enhances robustness to corruption by aligning test-time spatial correlations with training data, leveraging pre-trained samples.

Contribution

ShiftMatch is a new training-data-dependent likelihood that improves robustness in Bayesian neural networks without altering training likelihood, enabling use with pre-trained models.

Findings

ShiftMatch outperforms EmpCov priors on CIFAR-10-C.

ShiftMatch surpasses plain deep ensembles in robustness.

ShiftMatch effectively uses pre-trained BNN samples for improved performance.

Abstract

We develop ShiftMatch, a new training-data-dependent likelihood for robustness to corruption in Bayesian neural networks (BNNs). ShiftMatch is inspired by the training-data-dependent "EmpCov" priors from Izmailov et al. (2021a), and efficiently matches test-time spatial correlations to those at training time. Critically, ShiftMatch is designed to leave the neural network's training time likelihood unchanged, allowing it to use publicly available samples from pre-trained BNNs. Using pre-trained HMC samples, ShiftMatch gives strong performance improvements on CIFAR-10-C, outperforms EmpCov priors (though ShiftMatch uses extra information from a minibatch of corrupted test points), and is perhaps the first Bayesian method capable of convincingly outperforming plain deep ensembles.