Reproduce “Minimal Semantic Sufficiency Meets Unsupervised Domain Generalization” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Tan Pan, Kaiyu Guo, Dongli Xu, Zhaorui Tan, Chen Jiang, Deshu Chen, Xin Guo, Brian C. Lovell and 3 more

MS-UDG presents a method for Unsupervised Domain Generalization (UDG) that learns minimal sufficient semantic representations without requiring domain labels. It achieves state-of-the-art performance on benchmarks like PACS, improving average accuracy by up to +6.62% for linear evaluation and +5.48% for full fine-tuning compared to prior methods.

Key insights

  • Optimizing for a minimal sufficient semantic representation, which strictly captures only essential shared semantic information, theoretically leads to reduced out-of-distribution risk.
  • Effective disentanglement in UDG without domain labels can be achieved by jointly minimizing the mutual information between semantic and variation factors and ensuring the variation component captures all input information not contained in semantics.
  • A reconstruction objective, requiring decoding the original input from cross-view semantic and variation components, effectively ensures that the variation representation (v) captures domain-specific information, promoting minimality in the semantic representation (s).

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.