When faced with distribution shift at test time, deep neural networks often make inaccurate predictions with unreliable uncertainty estimates. While improving the robustness of neural networks is one promising approach to mitigate this issue, an appealing alternative to robustifying networks against all possible test-time shifts is to instead adapt them to unlabeled test data. In this paper, we propose a Bayesian model relating unlabeled data from the shifted test inputs to the classifier parame…