摘要
首先,对人工智能模型算法的可解释性风险和算法可解释性难题进行了描述;其次,对当前人工智能模型算法的几种解释方法的优缺点进行了分析,包括人工智能自动化解释法、等价解释法和局部解释法;最后,提出了一种人工智能模型解释方法的新思路,即通过对人工智能芯片电磁场的监测分析,实现对算法程序物理运行逻辑的复现,从而实现对算法的解释。
Firstly,the interpretability risks and algorithm interpretability problems of artificial intelligence(AI)model algorithms are described.Secondly,the advantages and disadvantages of several interpretation methods of current AI model algorithms are analyzed,including the AI automated explanation method,equivalent interpretation method and local explanation method.Finally,a new idea of AI model interpretation method is proposed,that is,through the monitoring and analysis of the electromagnetic field of the AI chip,the reproduction of the physical operation logic of the algorithm program is realized,so as to realize the interpretation of the algorithm.
作者
刘滨
栗向龙
古文刚
李乃鑫
黄创绵
LIU Bin;LI Xianglong;GU Wengang;LI Naixin;HUANG Chuangmian(CEPREI,Guangzhou 511370,China)
出处
《电子产品可靠性与环境试验》
2024年第3期71-74,共4页
Electronic Product Reliability and Environmental Testing
关键词
人工智能
算法
可解释性
电磁场
artificial intelligence
algorithm
explainability
electromagnetic field