Reproduce “Guiding Data Allocation for Robust Subpopulation 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
Zhaoying Pan, Yipei Wang, Shenyu Lu, Xiaoqian Wang
The authors estimate how adjusted accuracy (AA) changes with subgroup proportions and use its gradient to guide training-data allocation, rather than simply balancing groups. In matched experiments, AA generally rises faster along this direction than toward balance. On Waterbirds-syn at step 15, worst-group accuracy is 73.24% versus 65.79% for balancing; a single-seed sweep also finds high-performing imbalanced mixes.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.