An Investigation of Why Overparameterization Exacerbates Spurious Correlations

TL;DR

This paper investigates why increasing model size beyond perfect training error can worsen minority group test error due to spurious correlations, highlighting data properties and model biases.

Contribution

It identifies data proportions and signal-to-noise ratio as key factors and proposes subsampling the majority group to improve worst-group performance in overparameterized models.

Findings

Overparameterization can increase minority error despite better average test accuracy.

Subsampling the majority group reduces worst-group error in overparameterized models.

Standard upweighting of minority groups may fail to improve minority error.

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.