Self-supervised learning (SSL) methods based on the instance discrimination tasks with InfoNCE have achieved remarkable success. Despite their success, SSL models often struggle to generate effective representations for unseen-domain data. To address this issue, research on unsupervised domain generalization (UDG), which aims to develop SSL models that can generate domain-irrelevant features, has been conducted. Most UDG approaches utilize contrastive learning with InfoNCE to generate representations, and perform feature alignment based on strong assumptions to generalize domain-irrelevant common features from multi-source domains. However, existing methods that rely on instance discrimination tasks are not effective at extracting domain-irrelevant common features. This leads to the suppression of domain-irrelevant common features and the amplification of domain-relevant features, thereby hindering domain generalization. Furthermore, strong assumptions underlying feature alignment can lead to biased feature learning, reducing the diversity of common features. In this paper, we propose a novel approach, DomCLP, Domain-wise Contrastive Learning with Prototype Mixup. We explore how InfoNCE suppresses domain-irrelevant common features and amplifies domain-relevant features. Based on this analysis, we propose Domain-wise Contrastive Learning (DCon) to enhance domain-irrelevant common features. We also propose Prototype Mixup Learning (PMix) to generalize domain-irrelevant common features across multiple domains without relying on strong assumptions. The proposed method consistently outperforms state-of-the-art methods on the PACS and DomainNet datasets across various label fractions, showing significant improvements.
Most UDG approaches rely on instance discrimination tasks, which are not well-suited for domain generalization. In contrastive learning with InfoNCE, representations are learned to distinguish between instances. If the model attempts to differentiate between instances, deep neural networks tend to learn features that are useful to discriminate instances. In UDG environments, they are easy to capture domain-relevant features rather than domain-irrelevant common features, because domain-relevant features are more helpful to distinguish instances across various domains. As a result, the instance discrimination task suppresses domain-irrelevant features and amplifies domain-relevant features, thereby hindering domain generalization. Another limitation is that previous approaches heavily depend on feature alignment strategies across multi-domain, which reduce the diversity of domain-irrelevant common features.
To address these limitations, we propose a novel approach, DomCLP, Domain-wise Contrastive Learning with Prototype Mixup for unsupervised domain generalization. First, we theoretically and experimentally demonstrate that some negative terms in InfoNCE can suppress domain-irrelevant common features and amplifies domain-relevant features. Building on this insight, we introduce the Domain-wise Contrastive Learning (DCon) to enhance domain-irrelevant common features while representation learning. Second, to effectively generalize diverse domain-irrelevant common features across multi-domain, we propose the Prototype Mixup Learning (PMix). In PMix, to generalize common features from multi-domain, we interpolate common features in each domain utilizing mixup. We extract prototypes of features by k-means clustering, and train the model with mixed prototypes by mixup. It allows the model to effectively learn feature representations for unseen inter-manifold spaces while retaining diverse common feature information. Through our proposed method, DomCLP, the model effectively enhances and generalizes diverse common features.