Does data interpolation contradict statistical optimality?

摘要

We show that classical learning methods interpolating the training data can achieve optimal rates for the problems of nonparametric regression and prediction with square loss.