Reproduce “Partial Adversarial Domain Adaptation” 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

Zhangjie Cao, Lijia Ma, Mingsheng Long, Jianmin Wang

Partial Adversarial Domain Adaptation (PADA), developed by researchers at Tsinghua University, extends the DANN framework to effectively handle scenarios where the target label space is a subset of the source. The method substantially outperforms existing domain adaptation techniques across various benchmarks by explicitly mitigating negative transfer from irrelevant source classes.

Key insights

  • Negative transfer from source classes not present in the target domain is the primary challenge for effectively applying domain adaptation in partial label space scenarios.
  • The output probabilities from a source classifier on unlabeled target data can serve as an implicit indicator of source class relevance to the target domain.
  • Modifying both the classification and adversarial alignment objectives with these dynamic relevance weights allows for simultaneous mitigation of negative transfer and promotion of positive transfer within shared classes.

Summary by alphaXiv.

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