Learning by Choice of Internal Representations