Understanding deep learning requires rethinking generalization

   Abstract