B1842
Title: Smoothed discrepancy principle as an early stopping rule in RKHS
Authors: Yaroslav Averyanov - Inria Lille-Nord Europe (France) [presenting]
Abstract: The focus is on the estimation of a regression function that belongs to a reproducing kernel Hilbert space (RKHS). We describe spectral filter framework for our estimator that allows us to deal with several iterative algorithms: gradient descent, kernel ridge regression, etc. The main goal is to propose a new early stopping rule by introducing smoothing parameter for empirical risk of the estimator in order to improve the previous results on discrepancy principle. We explain, as well, how to generalise our strategy to different learning algorithms such as kNN and kernel regressions for choosing the hyperparameters. Theoretical justifications as well as simulations experiments for the proposed rule are provided.
