Paper Overview
Field: Computer Vision Authors: Zhaoying Pan, Yipei Wang, Shenyu Lu, Xiaoqian Wang Published: 2026-11-09 arXiv: 2611.12884
Abstract
Machine 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 remedy is to enforce a balanced proportion across subgroups, but full balance can be costly, infeasible, and is not always necessary. This work asks: is balanced training data the unique optimal configuration for robust subpopulation generalization?
The authors extend the Gaussian feature model framework to theoretically characterize the performance landscape under general training subpopulation distributions. Their results show that multiple configurations—including substantially imbalanced ones—can perform comparably to the balanced one.
They further derive gradient-based directions for adjusting subpopulation ratios to more efficiently improve both adjusted accuracy and worst-group accuracy. These findings indicate that strategic data allocation provides a more flexible and principled path to robust performance than simple balancing.
Code: https://github.com/zhaoyingpan/subpop-data-allocation
Key Takeaways
- Balanced training data is not the only effective configuration for handling subpopulation shift.
- Severely imbalanced training distributions can match balanced ones in robust generalization.
- Gradient-based guidance identifies which subpopulation ratios to adjust for the best worst-group accuracy gains.