Explore the influential samples in domain generalization

Abstract

Domain Generalization (DG) aims to learn a model that generalizes in testing domains unseen from training. All DG methods assume that the domain-invariant features can be learned by discarding the domain-specific ones. However, in practice, the learned invariant features usually contain "spurious invariance" that is only invariant across training domains but still variant to testing ones. We point out that this is because the contribution of the minority training samples without such spurious invariance is outgunned. Therefore, we are motivated to split these samples out of the original domains to form a new one, to which the spurious invariance is no longer invariant and thus removed. We present a cross-domain influence-based method, Domain+, to obtain the new domain. Specifically, for each sample per training domain, we estimate its influence by up-weighting it and then calculating how much the invariance loss of the other training domains changes---the more it changes, the higher the influence, and the more likely the sample belongs to the new domain. Then, with the split domains, we can deploy any off-the-shelf DG methods to achieve better generalization. We benchmark Domain+ on DomainBed and show that it helps existing SOTA methods achieve new SOTAs.