Paper says imbalanced training data can match balanced subgroup accuracy
Original title: Guiding Data Allocation for Robust Subpopulation Generalization
A Gaussian feature model gives gradient directions for changing subgroup proportions to raise worst-group accuracy.
Researchers study whether balanced subgroup proportions are the only optimal training setup under subpopulation shift. Extending a Gaussian feature model, they argue several configurations, including substantially imbalanced ones, can match balanced performance. They derive gradient-based adjustments intended to improve adjusted and worst-group accuracy more efficiently than simple balancing. Code is published on GitHub.
- Full balance is described as costly, sometimes infeasible, and not always necessary.
- The analysis covers general training subgroup distributions, not only equal shares.
- The repository is https://github.com/zhaoyingpan/subpop-data-allocation.
From the original article
1 min read at arxiv.orgMachine learning models often achieve high average accuracy in the training data while performing poorly on test data due to subpopulation shift. A common data-centric solution is to enforce balanced subgroup proportions, but full balance can be costly, infeasible, and not always necessary.
By Zhaoying Pan · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.