Interpolation can hurt robust generalization even when there is no noise

08/05/2021
by   Konstantin Donhauser, et al.
0

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.

READ FULL TEXT

Please sign up or login with your details

Forgot password? Click here to reset