Reproduce “Exploiting Domain-Specific Features to Enhance 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
Manh-Ha Bui, Toan Tran, Anh Tuan Tran, Dinh Phung
Researchers from VinAI Research and Monash University introduce the meta-Domain Specific-Domain Invariant (mDSDI) framework, which explicitly disentangles domain-invariant and label-informative domain-specific features, then applies meta-learning to the latter. This approach achieved the highest average classification accuracy of 67.9% across 7 benchmark datasets, demonstrating enhanced generalization by selectively adapting domain-specific knowledge.
Key insights
- Theoretical analysis, leveraging an information bottleneck perspective, formally demonstrates that label-correlated domain-specific information can enhance predictive performance beyond what is achievable with only domain-invariant features.
- Explicit disentanglement of domain-invariant and domain-specific features is crucial for preventing interference between them, maximizing their individual contributions to the classification task.
- Meta-learning is most effective in Domain Generalization when selectively applied to adaptable domain-specific features, as this preserves the stability of domain-invariant knowledge while enabling flexible adaptation of context-dependent cues.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.