Learning location-invariant orthographic representations for printed words