国家重点研发计划(2022YFA1005101, 2021ZD014030); 国家自然科学基金(91948303-1, 61836005, 61872371, 6203202)
随着智能信息时代的发展, 深度神经网络在人类社会众多领域中的应用, 尤其是在自动驾驶、军事国防等安全攸关系统中的部署, 引起了学术界和工业界对神经网络模型可能表现出的错误行为的担忧. 虽然神经网络验证和神经网络测试可以提供关于错误行为的定性或者定量结论, 但这种事后分析并不能防止错误行为的发生, 如何修复表现出错误行为的预训练神经网络模型依然是极具挑战性的问题. 为此, 深度神经网络修复这一领域应运而生, 旨在消除有缺陷的神经网络产生的错误预测, 使得神经网络满足特定的规约性质. 目前为止, 典型的神经网络修复范式有3种, 重训练、无错误定位的微调和包含错误定位的微调. 介绍深度神经网络的发展和神经网络修复的必要性; 厘清相近概念; 明确神经网络修复的挑战; 详尽地调研目前已有的神经网络修复策略, 并对内在联系与区别进行分析和比较; 调研整理神经网络修复策略常用的评价指标和基准测试; 展望未来神经网络修复领域研究中需要重点关注的可行方向.
With the development of the intelligent information era, applications of deep neural networks in various fields of human society, especially deployments in safety-critical systems such as automatic driving and military defense, have aroused concern from academic and industrial communities on the erroneous behaviors that deep neural networks may exhibit. Although neural network verification and neural network testing can provide qualitative or quantitative conclusions about erroneous behaviors, such post-analysis cannot prevent their occurrence. How to repair the pre-trained neural networks that feature wrong behavior is still a very challenging problem. To this end, deep neural network repair/patching comes into being, aiming at eliminating the unexpected predictions generated by defective neural networks and making the neural networks meet certain specification properties. So far, there are three typical neural network repair paradigms: retraining, fine tuning without fault localization, and fine tuning with fault localization. This study introduces the development of deep neural networks and the necessity of deep neural network repair, clarifies some similar concepts, and identifies the challenges of deep neural network repair. In addition, it investigates the existing neural network repair strategies in detail and compares the internal relationships and differences among these strategies. Moreover, the study explores and sorts out the evaluation metrics and benchmark tests commonly used in neural network repair strategies. Finally, it forecasts the feasible research directions that should be paid attention to in the future development of neural network repair strategies.