DomCLP: Domain-wise Contrastive Learning with Prototype Mixup for Unsupervised Domain Generalization

Jin-Seop Lee, Noo-Ri Kim, Jee-Hyong LeeAAAI 2025 · DOI

Introduction: The Domain Gap in Self-Supervised Learning

Unsupervised Domain Generalization (UDG) is a critical frontier in computer vision, aiming to develop models that can generalize to unseen environments without the benefit of labeled data. The core challenge is learning "domain-invariant" representations—ensuring that a model trained on photographs of dogs can still recognize a dog when presented as a sketch, a cartoon, or an oil painting.

Overview of the DomCLP framework showing the separation of domain-irrelevant and domain-relevant features alongside t-SNE visualizations.

Currently, Self-Supervised Learning (SSL) via the InfoNCE loss is the standard approach for learning without labels. However, standard SSL often relies on "instance discrimination," which inadvertently encourages models to focus on "domain-relevant" features—such as background, texture, or lighting—to distinguish between images. This leads to a failure in generalization because the model ignores the "essence" of the object in favor of its stylistic context.

The DomCLP Framework

To address these limitations, the authors introduce DomCLP, a framework designed to isolate domain-irrelevant features (like shape and components) from domain-relevant ones (like style and background). The framework consists of two primary novel components:

1. Domain-wise Contrastive Learning (DCon)

In standard contrastive learning, images from different domains are often treated as negative samples. The authors demonstrate that this forces the encoder to amplify domain-specific differences to minimize loss. DCon addresses this by restricting negative samples to the same domain as the anchor image. By removing cross-domain negatives, the model is prevented from "cheating" and must instead find intrinsic differences between objects within a single style.

2. Prototype Mixup Learning (PMix)

To prepare the model for entirely unseen domains, DomCLP utilizes Prototype Mixup (PMix). This technique creates synthetic training signals by interpolating between known domain representations:

  • Prototypical Discovery: The model uses k-means clustering within each domain to find centroids, or "prototypes," which represent shared semantic features across images.
  • Cross-Domain Interpolation: Mixup is applied to these prototypes across different domains.
  • Feature Mapping: By training the model to map a mixed input image to its corresponding mixed prototype, the framework encourages the model to navigate the "inter-manifold space"—the gaps between known domains that represent potential unseen environments.

Experimental Results

The framework was evaluated on the PACS and DomainNet benchmarks, two of the most rigorous tests for UDG. DomCLP demonstrated significant performance gains over previous state-of-the-art (SOTA) methods like DN2A and BrAD:

  • PACS Dataset: Achieved an 11.3% average accuracy improvement at a 1% label fraction.
  • DomainNet Dataset: Showed a remarkable 12.32% improvement at the 1% label fraction.

These results, visualized in t-SNE plots, show that DomCLP successfully clusters features by class rather than by domain, proving that the learned representations are inherently more robust and useful for downstream tasks.

Conclusion

DomCLP rethinks the application of contrastive learning in multi-domain settings. By focusing on domain-wise negatives and simulating intermediate styles through prototype mixup, it successfully extracts the domain-irrelevant features crucial for real-world AI deployment in diverse, unpredictable environments.

Summary updated

BibTeX
@inproceedings{lee2025domclp,
  author = {Jin-Seop Lee and Noo-Ri Kim and Jee-Hyong Lee},
  title = {{DomCLP: Domain-wise Contrastive Learning with Prototype Mixup for Unsupervised Domain Generalization}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2025},
  doi = {10.1609/aaai.v39i17.33993},
}
Figure