深度学习可解释性综述  

Review on Interpretability of Deep Learning

作　　者：陈冲 陈杰[1] 张慧[1] 蔡磊[1] 薛亚茹[1] CHEN Chong;CHEN Jie;ZHANG Hui;CAI Lei;XUE Yaru(College of Information Science and Engineering,China University of Petroleum(Beijing),Beijing 102249,China)

机构地区：[1]中国石油大学(北京)信息科学与工程学院,北京102249

出　　处：《计算机科学》2023年第5期52-63,共12页Computer Science

基　　金：国家自然科学基金(62006247);国家重点研发计划(2019YFC1510501,2022YFC2803704)。

摘　　要：随着数据量呈爆发式增长,深度学习理论与技术取得突破性进展,深度学习模型在众多分类与预测任务(图像、文本、语音和视频数据等)中表现出色,促进了深度学习的规模化与产业化应用。然而,深度学习模型的高度非线性导致其内部逻辑不明晰,并常常被视为“黑箱”模型,这也限制了其在关键领域(如医疗、金融和自动驾驶等)的应用。因此,研究深度学习的可解释性是非常必要的。首先对深度学习的现状进行简要概述,阐述深度学习可解释性的定义及必要性;其次对深度学习可解释性的研究现状进行分析,从内在可解释模型、基于归因的解释和基于非归因的解释3个角度对解释方法进行概述;然后介绍深度学习可解释性的定性和定量评估指标;最后讨论深度学习可解释性的应用以及未来发展方向。With the explosive growth of data volume and the breakthrough of deep learning theory and technology,deep learning models perform well enough in many classification and prediction tasks(image,text,voice and video data,etc.),which promotes the large-scale and industrialized application of deep learning.However,due to the high nonlinearity of the deep learning model with undefined internal logic,it is often regarded as a“black box”model which restricts further applications in key fields(such as medical treatment,finance,autonomous driving).Therefore,it is necessary to study the interpretability of deep learning.Firstly,recent studies on deep learning,the definition and necessity of explaining deep learning models are overviewed and described.Secondly,recent studies on interpretation methods of deep learning,and its classifications from the perspective of intrinsic interpretable model and attribution-based/non-attribution-based interpretation are analyzed and summarized.Then,the qualitative and quantitative performance criteria of the interpretability of deep learning are introduced.Finally,the applications of deep learning interpretability and future research directions are discussed and recommended.

关 键 词：深度学习 可解释性 归因解释 非归因解释 评估方法 

分 类 号：TP181[自动化与计算机技术—控制理论与控制工程]