Reproduce “Universal Domain Adaptation via Compressive Attention Matching” 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
Didi Zhu, Yincuan Li, Junkun Yuan, Zexi Li, Kun Kuang, Chao Wu
UniAM is a framework for Universal Domain Adaptation that uses Vision Transformer (ViT) attention and Compressive Attention Matching to robustly identify common and target-private classes without prior knowledge of label space relationships. The method achieved state-of-the-art H-scores on four benchmarks, including 95.95% on Office-31 and 81.68% on Office-Home.
Key insights
- Vision Transformer (ViT) self-attention, with its inherent shape bias, can be directly and effectively harnessed for classification and commonness detection in Universal Domain Adaptation.
- A compressive attention matching strategy, representing target attention sparsely using source prototypes, helps overcome attention mismatch caused by domain shifts.
- A holistic commonness assessment, combining residual-based metrics from both attention and global features, yields a more accurate and robust distinction between common and target-private classes.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.