Understanding deep learning requires rethinking generalization