Title: Lepskii principle in supervised learning
Authors: Nicole Muecke - Institute for Stochastics and Applications (Germany) [presenting]
Abstract: A statistical supervised learning problem is investigated under random design. In particular, we analyze reproducing kernel-based estimators arising from a fairly large class of spectral regularization methods. We derive a fully adaptive data driven estimator via a slightly modified version of Lepskiis principle, giving oracle properties both in reproducing kernel norm and prediction norm.