Enhancing Practical Domain Generalization via Partially Shared Features.

Abstract

Domain Generalization (DG) aims to enhance the performance of models trained on source domains when applied to out-of-distribution (OOD) domains. Traditional DG approaches typically rely on the causal model that assumes the observed data in each domain is generated by shared semantic variables, along with domain-specific non-semantic variables. This principle suggests that strong DG performance can be achieved by maximizing model dependence on semantic variables while minimizing that on non-semantic ones, guiding numerous prior works centered on learning entirely shared features (ESFs), which are features common across all source domains. However, this well-established causal model may inadequately represent real-world scenarios, as many domains often exhibit non-shared semantic variables-where each domain captures only a partial aspect of the overall semantics. For instance, a photorealistic painting might emphasize the color of an animal, while a sketch primarily focuses on its outline and shape. Hence, overly emphasizing ESFs risks suboptimal generalization, as the maximum dependency between ESFs and semantic variables can be notably compromised when semantic disparities across source domains grow. To this end, we first formalize a refined causal graph that characterizes such non-shared semantic variables. Then, we introduce a new DG framework that focuses on learning partially shared features (PSFs)-features shared among subsets of source domains, which contain and generalize ESFs. To implement this concept, we propose a framework to extend previous methods that learn ESFs by adopting an inverse meta-learning mechanism to learn PSFs. Experimental results show that our framework enhances the performance of various existing methods and outperforms many state-of-the-art methods, highlighting the advantages of further exploring PSFs to advance practical DG.