Abstract
Domain generalization is challenging due to the domain shift and the uncertainty caused by the inaccessibility of target domain data. In this pa- per, we address both challenges with a probabilis- tic framework based on variational Bayesian in- ference, by incorporating uncertainty into neu- ral network weights. We couple domain invari- ance in a probabilistic formula with the varia- tional Bayesian inference. This enables us to ex- plore domain-invariant learning in a principled way. Specifically, we derive domain-invariant rep- resentations and classifiers, which are jointly es- tablished in a two-layer Bayesian neural network. We empirically demonstrate the effectiveness of our proposal on four widely used cross-domain visual recognition benchmarks. Ablation studies validate the synergistic benefits of our Bayesian treatment when jointly learning domain-invariant representations and classifiers for domain general- ization. Further, our method consistently delivers state-of-the-art mean accuracy on all benchmarks.