Kim Daehee ; Park Seunghyun ; Kim Jinkyu ; Lee Jaekoo

Abstract

In general, an experimental environment for deep learning assumes that thetraining and the test dataset are sampled from the same distribution. However,in real-world situations, a difference in the distribution between twodatasets, domain shift, may occur, which becomes a major factor impeding thegeneralization performance of the model. The research field to solve thisproblem is called domain generalization, and it alleviates the domain shiftproblem by extracting domain-invariant features explicitly or implicitly. Inrecent studies, contrastive learning-based domain generalization approacheshave been proposed and achieved high performance. These approaches requiresampling of the negative data pair. However, the performance of contrastivelearning fundamentally depends on quality and quantity of negative data pairs.To address this issue, we propose a new regularization method for domaingeneralization based on contrastive learning, self-supervised contrastiveregularization (SelfReg). The proposed approach use only positive data pairs,thus it resolves various problems caused by negative pair sampling. Moreover,we propose a class-specific domain perturbation layer (CDPL), which makes itpossible to effectively apply mixup augmentation even when only positive datapairs are used. The experimental results show that the techniques incorporatedby SelfReg contributed to the performance in a compatible manner. In the recentbenchmark, DomainBed, the proposed method shows comparable performance to theconventional state-of-the-art alternatives. Codes are available athttps://github.com/dnap512/SelfReg.