Domain generalization via domain uncertainty shrinkage

Ensuring model robustness against distributional shifts still presents a significant challenge in many machine learning applications. To address this issue, a wide range of domain generalization (DG) methods have been developed. However, these approaches mainly focus on invariant representations by leveraging multiple source domain data, which ignore the uncertainty presented from different domains. In this paper, we establish a novel DG framework in form of evidential deep learning (EDL-DG). To reach DG objective under finite given domains, we propose a new Domain Uncertainty Shrinkage (DUS) regularization scheme on the output Dirichlet distribution parameters, which achieves better generalization across unseen domains without introducing additional structures. Theoretically, we analyze the convergence of EDL-DG, and provide a generalization bound in the framework of PAC-Bayesian learning. We show that our proposed method reduce the PAC-Bayesian bound under certain conditions, and thus achieve better generalization across unseen domains. In our experiments, we validate the effectiveness our proposed method on DomainBed benchmark in multiple real-world datasets.