English

Implicit Regularization of Random Feature Models

Machine Learning 2020-09-24 v2 Machine Learning

Abstract

Random Feature (RF) models are used as efficient parametric approximations of kernel methods. We investigate, by means of random matrix theory, the connection between Gaussian RF models and Kernel Ridge Regression (KRR). For a Gaussian RF model with PP features, NN data points, and a ridge λ\lambda, we show that the average (i.e. expected) RF predictor is close to a KRR predictor with an effective ridge λ~\tilde{\lambda}. We show that λ~>λ\tilde{\lambda} > \lambda and λ~λ\tilde{\lambda} \searrow \lambda monotonically as PP grows, thus revealing the implicit regularization effect of finite RF sampling. We then compare the risk (i.e. test error) of the λ~\tilde{\lambda}-KRR predictor with the average risk of the λ\lambda-RF predictor and obtain a precise and explicit bound on their difference. Finally, we empirically find an extremely good agreement between the test errors of the average λ\lambda-RF predictor and λ~\tilde{\lambda}-KRR predictor.

Keywords

Cite

@article{arxiv.2002.08404,
  title  = {Implicit Regularization of Random Feature Models},
  author = {Arthur Jacot and Berfin Şimşek and Francesco Spadaro and Clément Hongler and Franck Gabriel},
  journal= {arXiv preprint arXiv:2002.08404},
  year   = {2020}
}