Interpolation can hurt robust generalization even when there is no noise

Abstract

Numerous recent works show that overparameterization implicitly reduces variance for min-norm interpolators and max-margin classifiers. These findings suggest that ridge regularization has vanishing benefits in high dimensions. We challenge this narrative by showing that, even in the absence of noise, avoiding interpolation through ridge regularization can significantly improve generalization. We prove this phenomenon for the robust risk of both linear regression and classification and hence provide the first theoretical result on robust overfitting.

References in corpus (6)

• Classification vs regression in overparameterized regimes: Does the loss function matter?
• Do Wider Neural Networks Really Help Adversarial Robustness?
• On the Optimal Weighted ${\ell }_2$ Regularization in Overparameterized Linear Regression
• Domain adaptation under structural causal models
• Data Quality Matters For Adversarial Training: An Empirical Study
• Provable Robustness of Adversarial Training for Learning Halfspaces with Noise