卷积神经网络表征可视化研究综述

近年来, 深度学习在图像分类、目标检测及场景识别等任务上取得了突破性进展, 这些任务多以卷积神经网络为基础搭建识别模型, 训练后的模型拥有优异的自动特征提取和预测性能, 能够为用户提供“输入–输出”形式的端到端解决方案. 然而, 由于分布式的特征编码和越来越复杂的模型结构, 人们始终无法准确理解卷积神经网络模型内部知识表示, 以及促使其做出特定决策的潜在原因. 另一方面, 卷积神经网络模型在一些高风险领域的应用, 也要求对其决策原因进行充分了解, 方能获取用户信任. 因此, 卷积神经网络的可解释性问题逐渐受到关注. 研究人员针对性地提出了一系列用于理解和解释卷积神经网络的方法, 包括事后解释方法和构建自解释的模型等, 这些方法各有侧重和优势, 从多方面对卷积神经网络进行特征分析和决策解释. 表征可视化是其中一种重要的卷积神经网络可解释性方法, 能够对卷积神经网络所学特征及输入–输出之间的相关关系以视觉的方式呈现, 从而快速获取对卷积神经网络内部特征和决策的理解, 具有过程简单和效果直观的特点. 对近年来卷积神经网络表征可视化领域的相关文献进行了综合性回顾, 按照以下几个方面组织内容: 表征可视化研究的提起、相关概念及内容、可视化方法、可视化的效果评估及可视化的应用, 重点关注了表征可视化方法的分类及算法的具体过程. 最后是总结和对该领域仍存在的难点及未来研究趋势进行了展望.

Representation Visualization of Convolutional Neural Networks: A Survey

In recent years, deep learning has made breakthrough progress on image classification, object detection, and scene recognition tasks. These tasks mostly build recognition models based on the convolutional neural network (CNN). The trained models have excellent automatic feature extraction and prediction performance, which is able to provide users with “input-output” end-to-end solutions. However, due to the distributed feature coding and the increasingly complex model structure, users cannot yet accurately understand the internal knowledge representation of the model as well as the potential reasons for a specific decision. On the other hand, the application of the CNN models in some high-risk areas also requires a full understanding of the reason for their decisions, so as to get user＇s trust. Therefore, the interpreting ability of CNN has gradually attracted attention. Researchers have proposed a serious of methods for understanding and interpreting CNN, including post-hoc interpretation methods and building self-explainable models. These methods have their respective focuses and advantages, performing feature analysis and decision interpretation of CNN from various aspects. As one of the important CNN interpreting ability methods, representation visualization can visually present the features learned by CNN and the correlation between the input and output. In this way, a straightforward understanding of CNN internal features and decision-making can be obtained in a simple and intuitive way. This paper gives a comprehensive review of the related literatures on CNN representation visualization research in recent years, and organizes the content according to the following aspects: the introduction of representation visualization research, related concepts and contents, visualization methods, visualization effect evaluation, and the application of visualization. The classification of the representation visualization methods and the specific algorithms are our focus. Finally, the difficulties and future trends in the field are prospected, and the full text is summarized.