基于云模型的极端峰态神经元概念表达  

Expressing Unit Concepts with Extreme Kurtosis Based on Cloud Model

作　　者：包桦楠 李帅[1] 王国胤[1] BAO Hua-nan;LI Shuai;WANG Guo-yin(Chongqing Key Laboratory of computational intelligence(Chongqing University of Posts and Telecommunications),Chongqing 400065,China)

机构地区：[1]计算智能重庆市重点实验室(重庆邮电大学),重庆400065

出　　处：《模糊系统与数学》2021年第6期134-142,共9页Fuzzy Systems and Mathematics

基　　金：国家自然科学基金资助项目(61936001)。

摘　　要：近年来,深度学习在很多领域获得了广泛的应用。但是,由于深度学习的不可解释性,人类不能信任它的处理结果。本文针对这个问题,尝试使用云模型来表达特征空间中单个神经元的概念。经过逆向云变换算法得到的概念内涵可以表达概念的成熟度。在卷积神经网络(CNN)中,峰态系数过高或者过低的概念数量很多。现有逆向云发生器在转换这些概念时会产生较大的估计误差。这种误差导致得到的概念内涵不准确,干扰了对深度学习的解释。我们分析了单步法和多步法产生误差的原因,并用实际概念对分析结果进行了验证。分析结果表明,这些逆向云发生器在两种极端峰态系数的条件下会因为不同的原因产生幅度不等的误差。最后我们挑选了一种合适的逆向云发生器用于CNN单元概念内涵的转换。In recent years,deep learning has been widely applied in many fields.However,due to the unexplained nature of the deep neural network,human beings cannot trust it.To solve this problem,the cloud model is used to express the individual unit concept in the feature space.The concept intension obtained by reverse cloud transformation method can express the maturity of the concept.There are a large number of concepts with too high or too low kurtosis in the Convolutional Neural Network(CNN).Existing backward cloud generators have a large error when they transform these concepts.This error leads to the incorrect concept intension and interferes with the interpretation results.We analyze the cause of the error of single-step backward cloud generators and multi-step backward cloud generators and use the actual data to verify the analysis results.The results show that in the case of extreme kurtosis,these backward cloud generators produce variable errors due to different reasons.Finally,we choose a suitable backward cloud generator to transform the CNN unit concept intension.

关 键 词：云模型 逆向云发生器 峰态系数 深度学习可解释性 单元概念 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]