基于局部注意力图互迁移的可解释性优化方法  

Interpretability Optimization Method Based on Mutual Transfer of Local Attention Map

作　　者：成科扬[1] 王宁 崔宏纲 詹永照[1] CHENG Ke-yang;WANG Ning;CUI Hong-gang;ZHAN Yong-zhao(School of Computer Science and Communication Engineering,Jiangsu University,Zhenjiang,Jiangsu 212013,China)

机构地区：[1]江苏大学计算机科学与通信工程学院,江苏镇江212013

出　　处：《计算机科学》2022年第5期64-70,共7页Computer Science

基　　金：国家自然科学基金(61972183,61672268);社会安全风险感知与防控大数据应用国家工程实验室主任基金项目。

摘　　要：目前,深度学习模型已被广泛部署于各个工业领域。然而,深度学习模型具有的复杂性与不可解释性已成为其应用于高风险领域最主要的瓶颈。在深度学习模型可解释性方法中,最重要的方法是可视化解释方法,其中注意力图是可视化解释方法的主要表现方式,可通过对样本图像中的决策区域进行标注,来直观地展示模型决策依据。目前已有的基于注意力图的可视化解释方法中,单一模型注意力图存在标注区域易出现标注错误而造成可视化可解释性置信度不足的问题。针对上述问题,文中提出了一种基于局部注意力图互迁移的可解释性优化方法,用于提升模型注意力图的标注准确度,展示出精准的决策区域,加强视觉层面对模型决策依据的可解释性。具体表现为:采用轻量模型构建互迁移网络结构,于单一模型层间提取特征图并进行叠加,对全局注意力图进行局部划分,使用皮尔逊相关系数对模型间对应的局部注意力图进行相似度度量,随后将局部注意力图进行正则化并结合交叉熵函数对模型注意力图进行迁移。实验结果表明,所提算法显著提升了模型注意力图标注的准确性,并分别实现了28.2%的平均下降率和29.5%的平均增长率,与最先进的算法相比,其在平均下降率方面实现了3.3%的提升。实验结果表明,所提算法能成功地找出样本图像中预测标签最相关区域,而不局限于视觉可视化区域;与现有的同类方法相比,所提方法能更准确地揭示原始CNN模型的决策依据。At present,deep learning models have been widely deployed in various industrial fields.However,the complexity and inexplicability of deep learning model have become the main bottleneck of its application in high-risk fields.The most important method is the visual interpretation,in which the attention map is the main representation of the visual interpretation method.The decision area in the sample image can be marked to visually display the decision basis of the model.In the existing visual interpretation methods based on attention map,the single model attention map has the problem of insufficient confidence of visualization interpretability due to the annotation error easily appearing in the annotated region.To solve the above problems,this paper proposes an interpretable optimization method based on the mutual transfer of local attention map,aiming at improving the annotation accuracy of the model attention map and displaying the precise decision area,so as to strengthen the visual interpretable ability for the model decision basis.Specifically,the structure of the intermigration network is constructed by using the lightweight model,the feature maps are extracted and superimposed between the layers of the single model,and the global attention map is divided locally.Pearson correlation coefficient is used to measure the similarity of the corresponding local attention map between the models,and then the local attention map is regularized and transferred combined with the cross-entropy function..Experimental results show that the proposed algorithm significantly improves the accuracy of the model attention map label accuracy.The proposed algorithm achieves an average drop rate of 28.2%and an average increase rate of 29.5%,respectively,and achieves an increase of 3.3%in the average decline rate compared with the most advanced algorithm.The above experiments show that the proposed algorithm can successfully find out the most responsive region in the sample image,rather than being limited to the vi-sual visualization

关 键 词：可解释性 注意力图 区域划分 相似度 互迁移 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]