作者投稿和查稿 主编审稿 专家审稿 编委审稿 远程编辑

计算机工程 ›› 2020, Vol. 46 ›› Issue (9): 1-15. doi: 10.19678/j.issn.1000-3428.0057951

• 热点与综述 • 上一篇    下一篇

深度神经网络解释方法综述

苏炯铭, 刘鸿福, 项凤涛, 吴建宅, 袁兴生   

  1. 国防科技大学 智能科学学院, 长沙 410073
  • 收稿日期:2020-04-02 修回日期:2020-05-26 发布日期:2020-05-29
  • 作者简介:苏炯铭(1984-),男,副研究员、博士,主研方向为可解释的人工智能、智能对抗;刘鸿福,副研究员、博士;项凤涛、吴建宅,讲师、博士;袁兴生,副研究员、博士。
  • 基金资助:
    国家自然科学基金(61806212,61603403,U1734208,61603402,61703417);湖南省自然科学基金(2019JJ50724)。

Survey of Interpretation Methods for Deep Neural Networks

SU Jiongming, LIU Hongfu, XIANG Fengtao, WU Jianzhai, YUAN Xingsheng   

  1. College of Intelligence Science and Technology, National University of Defense Technology, Changsha 410073, China
  • Received:2020-04-02 Revised:2020-05-26 Published:2020-05-29

摘要: 深度神经网络具有非线性非凸、多层隐藏结构、特征矢量化、海量模型参数等特点,但弱解释性是限制其理论发展和实际应用的巨大障碍,因此,深度神经网络解释方法成为当前人工智能领域研究的前沿热点。针对军事、金融、医药、交通等高风险决策领域对深度神经网络可解释性提出的强烈要求,对卷积神经网络、循环神经网络、生成对抗网络等典型网络的解释方法进行分析梳理,总结并比较现有的解释方法,同时结合目前深度神经网络的发展趋势,对其解释方法的未来研究方向进行展望。

关键词: 可解释的人工智能, 深度神经网络, 卷积神经网络, 循环神经网络, 生成对抗网络

Abstract: Deep Neural Networks(DNN) are featured by non-linear non-convex properties,multiple hidden layers,feature vectorization,massive model parameters,etc.However,its weak interpretability has strangled their theory development and practical applications,so the interpretation methods for DNN have attracted attention from artificial intelligence researchers.In view of the strong requirements for the interpretability of DNN in high-risk decision-making fields such as military,finance,medicine,and transportation,this paper comprehensively combs and analyses typical network interpretation methods for typical networks such as Convolutional Neural Networks(CNN),Recurrent Neural Networks(RNN),and Generative Adversarial Networks(GAN).It also summarizes and compares existing interpretation methods.Then,based on the current development trend of DNN,the future research directions of interpretation methods are prospected.

Key words: eXplainable Artificial Intelligence(XAI), Deep Neural Networks(DNN), Convolutional Neural Networks(CNN), Recurrent Neural Networks(RNN), Generative Adversarial Networks(GAN)

中图分类号: