Group Robustness under the Microscope: How Class Imbalance Shapes the Effectiveness of Last-Layer Retraining
Author(s)
Zhang, Xinchen
Advisor(s)
Editor(s)
Collections
Supplementary to:
Permanent Link
Abstract
Machine learning models trained with empirical risk minimization (ERM) often rely on spurious correlations, leading to severe performance degradation on underrepresented subpopulations. This thesis provides a comprehensive examination of group robustness in fine-tuned neural networks, focusing on how class imbalance and data composition shape the effectiveness of last-layer retraining (LLR) and related balancing techniques. Through extensive experiments on four benchmark datasets spanning vision and language domains—Waterbirds, CelebA, CivilComments, and MultiNLI—we uncover several surprising and previously underexplored behaviors in worst-group accuracy (WGA).
We first diagnose failure modes in widely used class-balancing strategies. While subsetting, upsampling, and upweighting are intended to improve robustness, we show that upsampling and upweighting frequently cause catastrophic WGA collapse in imbalanced datasets, whereas subsetting may remove structurally important minority-subgroup examples. To address these limitations, we introduce mixture balancing, a hybrid method that combines subsetting with controlled resampling. Mixture balancing consistently improves WGA and mitigates the instability exhibited by prior methods.
We further analyze why LLR—despite modifying only the final classification layer—achieves disproportionate gains in robustness. Contrary to hypotheses based on neural collapse or maximum-margin convergence, our results demonstrate that LLR’s effectiveness is primarily driven by improved group balance in the held-out retraining set, even when group labels are unavailable. Finally, we identify a structural phenomenon, spectral imbalance, in which minority-group feature covariances systematically exhibit larger top eigenvalues, revealing deeper geometric causes of group disparity.
Together, these findings offer a unified perspective on data balance, model scaling, and representation geometry, and provide practical guidance for designing robust models under spurious correlations.
Sponsor
Date
2025-12
Extent
Resource Type
Text
Resource Subtype
Thesis (Masters Degree)