Understanding deep learning requires rethinking generalization
2016·Arxiv