Robust Training under Label Noise by Over-parameterization