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 model, its interpretability is poor, and it is often referred to as “black box” model, which can't 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 briefly introduced. Then, CiteSpace visualization software was used to retrieve, analyze and visualize relevant literature in Web of Science database, and it was found that interpretability is a hot research direction of computer science and artificial intelligence in recent years. Around the interpretability of deep learning, the status quo was analyzed from eight aspects, including hidden layer visualization, class activation mapping, sensitivity analysis, frequency principle, robust disturbance test, information theory, interpretable module, optimization method, and show its application in network security, recommendation system, medical and social networking. Finally, the existing problems and future development direction of deep learning interpretability research were discussed.

Key words: deep learning, interpretability, visualization, class activation mapping, frequency principle, interpretable module, information theory