Abstract
We develop ShiftMatch, a new training-data-dependent likelihood for out ofdistribution (OOD) robustness in Bayesian neural networks (BNNs). ShiftMatch isinspired by the training-data-dependent "EmpCov" priors from Izmailov et al.(2021a) and efficiently matches test-time spatial correlations to those attraining time. Critically, ShiftMatch is designed to leave neural networktraining unchanged, allowing it to use publically available samples frompretrained BNNs. Using pre-trained HMC samples, ShiftMatch gives strongperformance improvements on CIFAR-10-C, outperforms EmpCov priors, and isperhaps the first Bayesian method capable of convincingly outperforming plaindeep ensembles. ShiftMatch can be integrated with non-Bayesian methods likedeep ensembles, where it offers smaller, but still considerable, performanceimprovements. Overall, Bayesian ShiftMatch gave slightly better accuracy thanensembles with ShiftMatch, though they both had very similar log-likelihoods.