A Rate-Distortion Approach to Domain Generalization

Domain generalization deals with the difference in the distribution between the training and testing datasets, i.e., the domain shift problem, by extracting domain-invariant features. In this paper, we propose an information-theoretic approach for domain generalization. We first establish the domain transformation model, mapping a domain-free latent image into a domain. Then, we cast the domain generalization as a rate-distortion problem, and use the information bottleneck penalty to measure how well the domain-free latent image is reconstructed from a compressed representation of a domain-specific image compared to its direct prediction from the domain-specific image itself. We prove that the information bottleneck penalty guarantees that domain-invariant features can be learned. Lastly, we draw links of our proposed method with self-supervised contrastive learning without negative data pairs. Our empirical study on two different tasks verifies the improvement over recent baselines.