Reproduce “Simple and Fast Group Robustness by Automatic Feature Reweighting” 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
Shikai Qiu, Andres Potapczynski, Pavel Izmailov, Andrew Gordon Wilson
Automatic Feature Reweighting (AFR) is a two-stage method that significantly enhances model robustness to spurious correlations by automatically reweighting examples for last-layer retraining. AFR achieves state-of-the-art or highly competitive performance on various benchmarks with only 6% computational overhead compared to standard training and minimal reliance on explicit group labels.
Key insights
- The inherent weakness of an ERM model on minority groups within a held-out dataset can be leveraged to automatically identify and prioritize these examples for targeted intervention.
- Retraining only the last classification layer of a pretrained ERM model with an automatically weighted loss is a highly efficient and effective strategy for improving group robustness.
- A simple exponential reweighting mechanism, based on the first-stage model's predicted probabilities, effectively balances the training signal to mitigate spurious correlations.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.