Reproduce “Cross-Domain Diffusion with Progressive Alignment for Efficient Adaptive Retrieval” 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
Junyu Luo, Yusheng Zhao, Xiao Luo, Zhiping Xiao, Wei Ju, Li Shen, Dacheng Tao, Ming Zhang
A new method called COUPLE enhances unsupervised domain adaptive retrieval by simulating cross-domain transfer dynamics with graph diffusion and progressively aligning domains at multiple levels. This approach improves mean Average Precision by over 3% on Office-Home datasets and achieves a 26x to 31x speedup compared to dense vector retrieval.
Key insights
- Simulating mass transfer dynamics on a cross-domain graph effectively uncovers intrinsic structures and identifies low-noise 'early adopter' samples in the target domain, leading to more reliable pseudo-labeling for adaptation.
- Addressing domain discrepancies hierarchically at both pixel and manifold levels through progressive mixup is more effective than sole high-level feature alignment, enabling a comprehensive and gradual adaptation.
- Utilizing cross-domain random walks on the relationship graph to sample mixup pairs ensures that the generated intermediate samples effectively bridge the source and target domains, facilitating smooth knowledge transfer.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.