||DISI - Sala conferenze - 3° piano
||Adaptive Regularization Algorithms in Learning Theory
||Sergei V. Pereverzyev
||We investigate the problem of an adaptive parameter choice for regularization learning algorithms. In the theory of ill-posed problems there is a long history of choosing regularization parameters in optimal way without a priori knowledge of a smoothness of the element of interest. But known parameter choice rules cannot be applied directly in Learning Theory.
The point is that these rules are based on the estimation of the stability of
regularization algorithms measured in the norm of the space where unknown
element of interest should be recovered. But in the context of Learning
Theory this norm is determined by an unknown probability measure, and is not
accessible. In the talk we are going to discuss a new parameter choice
strategy adjusted to such a situation.