An Investigation of Why Overparameterization Exacerbates Spurious Correlations
Shiori Sagawa, Aditi Raghunathan, Pang Wei Koh, Percy Liang
Abstract
We study why overparameterization—increasing model size well beyond the point of zero training error—can hurt test error on minority groups despite improving average test error when there are spurious correlations in the data. Through simulations and experiments on two image datasets, we identify two key properties of the training data that drive this behavior: the proportions of majority versus minority groups, and the signal-to-noise ratio of the spurious correlations. We then analyze a linear setting and theoretically show how the inductive bias of models towards “memorizing” fewer examples can cause overparameterization to hurt. Our analysis leads to a counterintuitive approach of subsampling the majority group, which empirically achieves low minority error in the overparameterized regime, even though the standard approach of upweighting the minority fails. Overall, our results suggest a tension between using overparameterized models versus using all the training data for achieving low worst-group error.
BibTeX
@InProceedings{pmlr-v119-sagawa20a,
title = {An Investigation of Why Overparameterization Exacerbates Spurious Correlations},
author = {Sagawa, Shiori and Raghunathan, Aditi and Koh, Pang Wei and Liang, Percy},
booktitle = {Proceedings of the 37th International Conference on Machine Learning},
pages = {8346--8356},
year = {2020},
editor = {III, Hal Daumé and Singh, Aarti},
volume = {119},
series = {Proceedings of Machine Learning Research},
month = {13--18 Jul},
publisher = {PMLR},
pdf = {http://proceedings.mlr.press/v119/sagawa20a/sagawa20a.pdf},
url = {https://proceedings.mlr.press/v119/sagawa20a.html},
abstract = {We study why overparameterization—increasing model size well beyond the point of zero training error—can hurt test error on minority groups despite improving average test error when there are spurious correlations in the data. Through simulations and experiments on two image datasets, we identify two key properties of the training data that drive this behavior: the proportions of majority versus minority groups, and the signal-to-noise ratio of the spurious correlations. We then analyze a linear setting and theoretically show how the inductive bias of models towards “memorizing” fewer examples can cause overparameterization to hurt. Our analysis leads to a counterintuitive approach of subsampling the majority group, which empirically achieves low minority error in the overparameterized regime, even though the standard approach of upweighting the minority fails. Overall, our results suggest a tension between using overparameterized models versus using all the training data for achieving low worst-group error.}
}