《计算机应用》唯一官方网站 ›› 2022, Vol. 42 ›› Issue (12): 3639-3650.DOI: 10.11772/j.issn.1001-9081.2021091649

• 人工智能 •    

深度学习的可解释性研究综述

李凌敏, 侯梦然, 陈琨, 刘军民()   

  1. 西安交通大学 数学与统计学院,西安 710049
  • 收稿日期:2021-09-22 修回日期:2022-01-15 接受日期:2022-01-20 发布日期:2022-12-21 出版日期:2022-12-10
  • 通讯作者: 刘军民
  • 作者简介:李凌敏(1998—),女,山西晋城人,硕士研究生,主要研究方向:神经网络可解释性
    侯梦然(1999—),女,山西临汾人,硕士研究生,主要研究方向:神经网络可解释性
    陈琨(1984—),男,陕西咸阳人,博士研究生,主要研究方向:人工智能、知识图谱及其应用
    刘军民(1982—),男,陕西西安人,副教授,博士,CCF会员,主要研究方向:机器学习、图像处理、数据分析。
  • 基金资助:
    国家重点研发计划项目(2020AAA0105601)

Survey on interpretability research of deep learning

Lingmin LI, Mengran HOU, Kun CHEN, Junmin LIU()   

  1. School of Mathematics and Statistics,Xi’an Jiaotong University,Xi’an Shaanxi 710049,China
  • Received:2021-09-22 Revised:2022-01-15 Accepted:2022-01-20 Online:2022-12-21 Published:2022-12-10
  • Contact: Junmin LIU
  • About author:LI Lingmin, born in 1998, M. S. candidate. Her research interests include interpretability of neural networks.
    HOU Mengran, born in 1999, M. S. candidate. Her research interests include interpretability of neural networks.
    CHEN Kun,born in 1984, Ph. D. candidate. His research interests include artificial intelligence, knowledge graph and its application.
    LIU Junmin,born in 1982, Ph. D., associate professor. His research interests include machine learning, image processing, data analysis.3650
  • Supported by:
    National key Research and Development Program of China(2020AAA0105601)

摘要:

近年来,深度学习在很多领域得到广泛应用;然而,由于深度神经网络模型的高度非线性操作,导致其可解释性较差,并常常被称为“黑箱”模型,无法应用于一些对性能要求较高的关键领域;因此,对深度学习的可解释性开展研究是很有必要的。首先,简单介绍了深度学习;然后,围绕深度学习的可解释性,从隐层可视化、类激活映射(CAM)、敏感性分析、频率原理、鲁棒性扰动测试、信息论、可解释模块和优化方法这8个方面对现有研究工作进行分析;同时,展示了深度学习在网络安全、推荐系统、医疗和社交网络领域的应用;最后,讨论了深度学习可解释性研究存在的问题及未来的发展方向。

关键词: 深度学习, 可解释性, 隐层可视化, 类激活映射, 频率原理, 可解释模块, 信息论

Abstract:

In recent years, deep learning has been widely used in many fields. However, due to the highly nonlinear operation of deep neural network models, the interpretability of these models is poor, these models are often referred to as “black box” models, and cannot be applied to some key fields with high performance requirements. Therefore, it is very necessary to study the interpretability of deep learning. Firstly, deep learning was introduced briefly. Then, around the interpretability of deep learning, the existing research work was analyzed from eight aspects, including hidden layer visualization, Class Activation Mapping (CAM), sensitivity analysis, frequency principle, robust disturbance test, information theory, interpretable module and optimization method. At the same time, the applications of deep learning in the fields of network security, recommender system, medical and social networks were demonstrated. Finally, the existing problems and future development directions of deep learning interpretability research were discussed.

Key words: deep learning, interpretability, hidden layer visualization, Class Activation Mapping (CAM), frequency principle, interpretable module, information theory

中图分类号: