基于可解释基拆解和知识图谱的深度神经网络可视化

Recently, owing to the advantages of deep-layered learning and unlabeled learning, etc., deep learning models represented by convolutional neural network, deep neural network, recurrent neural network, have gained increasing applications in various fields, such as image recognition, video, and natural language processing. To achieve the high transparency and security assurance of deep learning models, the interpretability research of deep neural networks is of great theoretical significance and industrial application value and recently gains increasingly attention. However, because of the intrinsic black-box characteristics of the deep learning models, the interpretation of its internal structure and the running mechanism is still of great challenges, including the rigorous theoretical results originated from the manual observations of large-scale training and testing set, and scarce appropriate explanation of the learning results based on the human understanding. Moreover, most of the existing researches analyzing the decision-making process of deep learning models only from a local perspective and lacks a graphical representation based on the overall understanding. On the other hand, the interpretable basis decomposition (IBD) model has the advantages that its interpretation result is not only a strict corresponding relation from scene to feature, but also is a kind of semi-structured data which can facilitate IBD based knowledge map construction from it. Aiming at the problem that existing deep neural network visualization researches lacks the interpretability based on the knowledge map and the well-suited knowledge map representability of IBD, we propose a deep neural network visualization approach based on interpretable basis decomposition and knowledge map, which fully takes the advantage of map construction ability of interpretable basis decomposition. Firstly, we propose a knowledge map construction method based on the feature decomposition structure of IBD, which constructs the map information, such as the interpretation relationship and juxtaposition relationship, between the scene and the interpretable feature. Then, a similarity clustering algorithm between scenes using Jaccard coefficient based on the interpretation relation network of scenes and features is proposed. Based on a scene discriminant feature extraction method, the decomposed features that can distinguish this class from other classes are extracted from each type of sample, namely discriminant features. Meanwhile, we quantify the accuracy of discriminant feature extraction by means of manual evaluation by exploring the difference between different models' understanding of the recognition target and that of human beings. Furthermore, a fidelity test method for deep network has been proposed to solve the problem that existing research lacks fidelity test. We combine the multi-feature thermal spectrograms into a comprehensive characteristic thermal spectrogram, and then use the Hadmag product to refuse the comprehensive characteristic thermal spectrogram with the original image to obtain the characteristic fusion spectrogram. The luminance labeled depth neural network classification model of feature fusion map was used to identify the target location pixel area, and the target location ability of thermal spectrum map was measured by comparing the deviation of input original map and feature fusion map to the model classification ability, so as to obtain the fidelity of the interpretable basis decomposition model. Both the fidelity test and the human confidence test show that the proposed method can achieve excellent results.