Reproduce “Upcycling Models under Domain and Category Shift” 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
Sanqing Qu, Tianpei Zou, Florian Roehrbein, Cewu Lu, Guang Chen, Dacheng Tao, Changjun Jiang
Researchers from Tongji University and collaborators introduce Global and Local Clustering (GLC), a finetuning-free approach for Universal Domain Adaptation (UniDA) that operates without source data. This method adapts a standard pre-trained closed-set model to various category shifts, achieving a 14.8% H-score improvement over prior methods on the VisDA dataset for open-partial-set adaptation.
Key insights
- Universal Domain Adaptation can be achieved without source data by adaptively generating pseudo-labels, distinguishing known and unknown classes, and dynamically estimating target categories.
- A combination of global one-vs-all clustering and local k-NN consistency is effective in mitigating negative transfer and managing diverse category shifts.
- Utilizing a standard pre-trained closed-set source model is sufficient for SF-UniDA, eliminating the need for specialized architectures.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.