Stability and Generalization of Bilevel Programming in Hyperparameter Optimization

Author name not available (Why is that?)

Publication date: 8 June 2021

Abstract: The (gradient-based) bilevel programming framework is widely used in hyperparameter optimization and has achieved excellent performance empirically. Previous theoretical work mainly focuses on its optimization properties, while leaving the analysis on generalization largely open. This paper attempts to address the issue by presenting an expectation bound w.r.t. the validation set based on uniform stability. Our results can explain some mysterious behaviours of the bilevel programming in practice, for instance, overfitting to the validation set. We also present an expectation bound for the classical cross-validation algorithm. Our results suggest that gradient-based algorithms can be better than cross-validation under certain conditions in a theoretical perspective. Furthermore, we prove that regularization terms in both the outer and inner levels can relieve the overfitting problem in gradient-based algorithms. In experiments on feature learning and data reweighting for noisy labels, we corroborate our theoretical findings.

Has companion code repository: https://github.com/baofff/stability_ho

This page was built for publication: Stability and Generalization of Bilevel Programming in Hyperparameter Optimization

Report a bug (only for logged in users!)Click here to report a bug for this page (MaRDI item Q6369740)