Reparametrization invariance in representation learning