Learning, Regularization and Ill-Posed Inverse Problems