| 本文已被：浏览 5199次 下载 8968次 |
码上扫一扫！ |
|
|
| 卷积神经网络的可解释性研究综述 |
|
窦慧1,2, 张凌茗1,2, 韩峰1,2, 申富饶1,3, 赵健4
|
|
1.计算机软件新技术国家重点实验室(南京大学), 江苏 南京 210023;2.南京大学 计算机科学与技术系, 江苏 南京 210023;3.南京大学 人工智能学院, 江苏 南京 210023;4.南京大学 电子科学与工程学院, 江苏 南京 210023
|
|
| 摘要: |
| 神经网络模型性能日益强大, 被广泛应用于解决各类计算机相关任务, 并表现出非常优秀的能力, 但人类对神经网络模型的运行机制却并不完全理解. 针对神经网络可解释性的研究进行了梳理和汇总, 就模型可解释性研究的定义、必要性、分类、评估等方面进行了详细的讨论. 从解释算法的关注点出发, 提出一种神经网络可解释算法的新型分类方法, 为理解神经网络提供一个全新的视角. 根据提出的新型分类方法对当前卷积神经网络的可解释方法进行梳理, 并对不同类别解释算法的特点进行分析和比较. 同时, 介绍了常见可解释算法的评估原则和评估方法. 对可解释神经网络的研究方向与应用进行概述. 就可解释神经网络面临的挑战进行阐述, 并针对这些挑战给出可能的解决方向. |
| 关键词: 神经网络 可解释性 分类 深度学习 |
| DOI：10.13328/j.cnki.jos.006758 |
| 分类号: |
| 基金项目:科技部科技创新2030重大项目(2021ZD0201300); 国家自然科学基金(61876076) |
|
| Survey on Convolutional Neural Network Interpretability |
|
DOU Hui1,2, ZHANG Ling-Ming1,2, HAN Feng1,2, SHEN Fu-Rao1,3, ZHAO Jian4
|
|
1.State Key Laboratory for Novel Software Technology (Nanjing University), Nanjing 210023, China;2.Department of Computer Science and Technology, Nanjing University, Nanjing 210023, China;3.School of Artificial Intelligence, Nanjing University, Nanjing 210023, China;4.School of Electronic Science and Engineering, Nanjing University, Nanjing 210023, China
|
| Abstract: |
| With the increasingly powerful performance of neural network models, they are widely used to solve various computer-related tasks and show excellent capabilities. However, a clear understanding of the operation mechanism of neural network models is lacking. Therefore, this study reviews and summarizes the current research on the interpretability of neural networks. A detailed discussion is rendered on the definition, necessity, classification, and evaluation of research on model interpretability. With the emphasis on the focus of interpretable algorithms, a new classification method for the interpretable algorithms of neural networks is proposed, which provides a novel perspective for the understanding of neural networks. According to the proposed method, this study sorts out the current interpretable methods for convolutional neural networks and comparatively analyzes the characteristics of interpretable algorithms falling within different categories. Moreover, it introduces the evaluation principles and methods of common interpretable algorithms and expounds on the research directions and applications of interpretable neural networks. Finally, the problems confronted in this regard are discussed, and possible solutions to these problems are given. |
| Key words: neural network interpretability taxonomy deep learning |
