A0739
Title: Benign overfitting in stochastic regression
Authors: Shogo Nakakita - The University of Tokyo (Japan) [presenting]
Masaaki Imaizumi - The University of Tokyo (Japan)
Abstract: The excess risks of overparameterized stochastic regression, that is, linear regression with covariates being stochastic processes and whose number is greater than that of samples, are considered. Recent studies show that overparameterized linear regression with i.i.d. samples can predict well even if they have fewer samples than parameters and no sparsity. We examine how time series structure can affect the performance of overparameterized regression without sparsity. One of the results is that even if the covariates have long-range dependence, the sufficiently fast decay of eigenvalues of the covariance operator can make the excess risk converge to zero.
