On learning with shift-invariant structures