【作者】 王宁；

【导师】 成科扬；

【作者基本信息】 江苏大学 ， 计算机技术（专业学位）， 2021， 硕士

【摘要】 深度神经网络可解释性是深度学习领域中非常重要的课题,能够突破模型本身复杂性与不可解释性而造成其难以应用于各个工业领域的瓶颈。在诸多深度学习神经网络解释方法中,最为重要的是基于注意力图的可视化解释。由于注意力图能够对样本图像中决策区域进行标注,直观展示模型决策依据,因此注意力图对决策区域进行标注的准确度直接影响可视化可解释性的置信度。在实际工作当中,存在注意力图标注区域易出现偏差而导致模型决策依据难以直观理解,模型可视化解释难以验证的问题。为解决上述问题,本文在深入研究深度学习可视化解释技术的基础上,提出了注意力图迁移和可解释性验证方法,同时设计并实现了基于内部决策解释的物品分类可解释性系统。本文主要研究工作包括以下内容:(1)提出一种基于局部注意力图互迁移的可解释性优化方法。目前存在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偏差甚至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,本节提出多模型间注意力图互迁移算法,致力于提升模型注意力图标注准确度,以此加强视觉层面对于模型决策依据的可解释性。具体为采用轻量模型构建互迁移网络结构,在单一模型层间提取特征图并进行叠加,并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。实验表明,该算法有效提升了注意力图标注区域的准确度,可视化层面可解释性显著提高。(2)提出一种基于深度决策树节点可视化的可解释性验证方法。该方法突破上节方法基于结果进行解释的局限性,主要针对模型内部进行逐层决策解释并对可解释性置信度进行验证。首先于全连接层前提取模型特征图并输入至深度支持决策树根节点,与节点内部权重做内积并进行逐层推理,提出优化类激活映射算法对节点输出进行可视化解释。而后通过交并比算法对骨干网络中卷积核单元与类别间的相关度进行度量并对特定卷积核单元置零,提取重构后的骨干网络特征图并重新输入决策树,从模型内部验证决策的可靠性,并对置信度进行主观与量化分析。实验结果表明,该算法有效提升了决策树节点注意力图标注准确度,同时,所提出的检验算法能够有效提升可解释性置信度。(3)采用功能模块化的设计,开发并实现了基于内部决策解释的物品分类可解释性系统。主要系统功能模块包括图像输入模块、图像分类模块、注意力图可视化模块、决策过程可视化模块、可解释性验证模块和存储模块。实验表明,该系统具有较为良好的易用性和有效性,适用于包括移动拍摄平台等多种设备中,具有较高的应用价值和前景。更多还原

【Abstract】 At present,deep learning models have been widely deployed in various industrial fields.However,the complexity and inexplicable of deep learning model have become the main bottleneck of its application in high-risk fields.Among many deep learning neural network interpretation methods,visual interpretation based on attention diagram is the most important one.Since attention diagram can intuitively show the basis of model decision by marking decision areas in sample images,the accuracy of decision areas annotation by attention diagram directly affects the confidence of visualization interpretability.In practical work,there is a problem that the annotated area of attention diagram is prone to deviation,which makes it difficult to intuitively understand the basis of the model decision and difficult to verify the visual interpretation of the model.In order to solve this problem,the thesis proposes the improvement of attention diagram migration and interpretability verification method based on in-depth study of deep learning visual interpretation technology,and designed a interpretability system of item classification based on internal decision interpretation.The main research work of this paper includes the following contents:(1)A local mutual migration method based on attention map is proposed.Among the existing visual interpretation methods based on attention diagram,the single model attention diagram is prone to deviation or even labeling error in the annotated region,which leads to the problem of insufficient confidence in visual interpretability.In view of the above problems,this section proposes an algorithm of mutual transfer of attention diagram among multiple models,aiming to improve the annotation accuracy of model attention diagram and display accurate decision area,so as to strengthen the explanability of visual level for the basis of model decision.The lightweight model was used to construct the network structure,and migrate feature maps between layers in a single model,the global attention diagram is partitioned locally.We use Pearson correlation coefficient between the model corresponding to the parts of their attention to a similarity measure,then will be regularized and combined with local attention to cross entropy function note is trying to migrate to the model.The Experiment results show that the algorithm can improve the accuracy of the annotated area of the attention map,and the interpretability of the visualization level is significantly improved.(2)A new method for verifying interpretability based on visualization of deep decision tree nodes is proposed.This method breaks through the limitation of the method based on results in the previous section,and mainly interprets the decision process within the model and verifies the interpretability confidence.Firstly,the feature map of the model was extracted before the full connection layer and send to the root node of the neural backed decision tree.Then,the inner product with the internal weight of the node was made and calculate layer by layer.Then,we proposed an optimized class activation mapping algorithm to visually explain the output of the node.Then,the correlation between convolutional kernel units and categories in the backbone network was measured by the crossover ratio algorithm IOU and the specific convolutional kernel units were dropped.The reconstructed backbone network feature graphs were extracted and re-entered into the decision tree.The reliability of the decision was verified from inside the model,and the subjective and quantitative analysis of the confidence degree was carried out.The experimental results show that the proposed algorithm can effectively improve the annotation accuracy of attention map in decision tree nodes,and the proposed test algorithm can effectively improve the confidence of interpretability.(3)The classification interpretability system based on internal decision interpretation is developed and implemented by functional modularization design.The main functional modules of the system include image input module,image classification module,attention force visualization module,decision-making process visualization module,interpretability verification module and storage module.The experiment shows that the system has good usability and effectiveness,and it is suitable for many kinds of equipment,including mobile camera platform,and has high application value and prospect.更多还原