Learning, regularization and splines