计算机科学

计算机科学 ›› 2024, Vol. 51 ›› Issue (9): 155-161.doi: 10.11896/jsjkx.230900109

• 计算机图形学&多媒体 • 上一篇    下一篇

一种语义引导的神经网络关键数据路由路径算法

朱富坤1, 滕臻2, 邵文泽1, 葛琦1, 孙玉宝3   

  1. 1 南京邮电大学通信与信息工程学院 南京 210003
    2 南京邮电大学贝尔英才学院 南京 210042
    3 南京信息工程大学教育部数字取证工程研究中心 南京 210044
  • 收稿日期:2023-09-19 修回日期:2024-03-06 出版日期:2024-09-15 发布日期:2024-09-10
  • 通讯作者: 邵文泽(shaowenze@njupt.edu.cn)
  • 作者简介:(zhufukun1999@163.com)
  • 基金资助:
    国家自然科学基金(61771250,61972213)

Semantic-guided Neural Network Critical Data Routing Path

ZHU Fukun1, TENG Zhen2, SHAO Wenze1, GE Qi1, SUN Yubao3   

  1. 1 School of Communications and Information Engineering,Nanjing University of Posts and Telecommunications,Nanjing 210003,China
    2 Bell Honors School,Nanjing University of Posts and Telecommunications,Nanjing 210042,China
    3 Engineering Research Center for Digital Forensics Ministry of Education,Nanjing University of Information Science and Technology,Nanjing 210044,China
  • Received:2023-09-19 Revised:2024-03-06 Online:2024-09-15 Published:2024-09-10
  • About author:ZHU Fukun,born in 1999,postgra-duate.His main research interests include interpretability and adversarial transferability of deep learning mo-dels.
    SHAO Wenze,born in 1981,Ph.D,professor.His main research interests include computational imaging,computer vision and machine learning.
  • Supported by:
    National Natural Science Foundation of China(61771250,61972213).

PDF (PC)

348

摘要： 近年来,由于人工智能在各领域的普及,研究神经网络的可解释方法及理解神经网络的运作机理已经成为一个愈发重要的话题。作为神经网络解释性方法的一个分支,网络的路径可解释性受到了越来越多的关注。文中特别探讨了关键数据路由路径(Critical Data Routing Path,CDRP)这一面向网络路径的可解释方法。首先,通过Score-CAM(Score-Class Activation Map)方法分析了CDRP在输入域上的路径可视化归因,指出CDRP方法在语义层面的潜在缺陷。然后,提出了一种语义引导的Score-CDRP方法,从方法机理上提升了CDRP与原始神经网络的语义一致性。最后,通过实验从路径热力图可视化以及相应的预测与定位精度等角度验证了Score-CDRP方法相较于CDRP的合理性、有效性和鲁棒性。

关键词: 计算机视觉, 深度神经网络, 神经网络可解释性, 特征可视化, 网络剪枝, 热力图

Abstract: In recent years,with the popularity of artificial intelligence in various fields,it has become an increasingly important topic to study the interpretable methods of neural networks and understand their running principles.As a subfield of neural network interpretability methods,the interpretability of network pathways garners increasing attention.This paper particularly focuses on the critical data routing path(CDRP),an interpretable method for network pathways.Firstly,the routing path visualization attribution of CDRP in the input domain is analyzed by use of the score-class activation map(Score-CAM) method,pointing out the inherent defects of the CDRP approach in terms of semantics.Then a channel semantic guided CDRP method termed as Score-CDRP is proposed,which improves the semantic consistency between the original deep neural network and its corresponding CDRP from the perspective of method mechanism.Lastly,experimental results demonstrate that the proposed Score-CDRP approach is more reasonable,effective and robust than CDRP in terms of visualization of the routing path heatmap as well as its corresponding prediction and localization accuracy.

Key words: Computer vision, Deep neural networks, Interpretability of neural networks, Feature visualization, Network pruning, Heatmap

中图分类号: 

  • TP183
[1]SUBAKAN C,RAVANELLI M,CORNELL S,et al.Attention is all you need in speech separation[C]//2021 IEEE International Conference on Acoustics,Speech and Signal Processing.(ICASSP 2021)IEEE,2021:21-25.
[2]ZHU X,LYU S,WANG X,et al.TPH-YOLOv5:ImprovedYOLOv5 based on transformer prediction head for object detection on drone-captured scenarios[C]//Proceedings of the IEEE/CVF International Conference on Computer Vision.2021:2778-2788.
[3]RICCARDO G,ANNA M,SALVATORE R,et al.A Survey Of Methods For Explaining Black Box Models[J].ACM Computing Surveys,2018,51(5):1-42.
[4]YOSINSKI J,CLUNE J,NGUYEN A,et al.Understandingneural networks through deep visualization[J].arXiv:1506.06579,2015.
[5]ERHAN D,BENGIO Y,COURVILLE A,et al.Visualizinghigher-layer features of a deep network[J].University of Montreal,2009,1341(3):1-13.
[6]ZHOU B,KHOSLA A,LAPEDRIZA A,et al.Learning deepfeatures for discriminative localization[C]//Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition.2016:2921-2929.
[7]BACH S,BINDER A,MONTAVON G,et al.On pixel-wise explanations for non-linear classifier decisions by layer-wise relevance propagation[J].PloS One,2015,10(7):e0130140.
[8]SPRINGENBERG J T,DOSOVITSKIY A,BROX T,et al.Striving for simplicity:The all convolutional net[J].arXiv:1412.6806,2014.
[9]WANG Y,SU H,ZHANG B,et al.Interpret neural networks byidentifying critical data routing paths[C]//proceedings of the IEEE Conference on Computer Vision and Pattern Recognition.2018:8906-8914.
[10]TIBSHIRANI R.Regression shrinkage and selection via the lasso[J].Journal of the Royal Statistical Society:Series B(Metho-dological),1996,58(1):267-288.
[11]HOEFLER T,ALISTARH D,BEN-NUN T,et al.Sparsity in deep learning:Pruning and growth for efficient inference and training in neural networks[J].The Journal of Machine Lear-ning Research,2021,22(1):10882-11005.
[12]HINTON G,VINYALS O,DEAN J.Distilling the Knowledgein a Neural Network[J].Computer Science,2015,14(7):38-39.
[13]SELVARAJU R R,COGSWELL M,DAS A,et al.Grad-cam:Visual explanations from deep networks via gradient-based localization[C]//Proceedings of the IEEE International Confe-rence on Computer Vision.2017:618-626.
[14]CHATTOPADHAY A,SARKAR A,HOWLADER P,et al.Grad-cam++:Generalized gradient-based visual explanations for deep convolutional networks[C]//2018 IEEE Winter Conference on Applications of Computer Vision.IEEE,2018:839-847.
[15]SATTARZADEH S,SUDHAKAR M,PLATANIOTIS K N,et al.Integrated grad-cam:Sensitivity-aware visual explanation of deep convolutional networks via integrated gradient-based scoring[C]//ICASSP 2021-2021 IEEE International Conference on Acoustics,Speech and Signal Processing.IEEE,2021:1775-1779.
[16]JIANG P,ZHANG C,HOU Q,et al.LayerCAM:Exploring Hi-erarchical Class Activation Maps for Localization.[J].IEEE Transactions on Image Processing:a Publication of the IEEE Signal Processing Society,2021,30:5875-5888.
[17]WANG H,WANG Z,DU M,et al.Score-CAM:Score-weighted visual explanations for convolutional neural networks[C]//Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops.2020:24-25.
[18]IBRAHIM R,SHAFIQ M O.Augmented Score-CAM:High reso-lution visual interpretations for deep neural networks[J].Knowledge-Based Systems,2022,252:109287.
[19]KHAKZAR A,BASELIZADEH S,KHANDUJA S,et al.Neural response interpretation through the lens of critical pathways[C]//Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition.2021:13528-13538.
[20]ZHANG Y,TIÑO P,LEONARDIS A,et al.A survey on neural network interpretability[J].IEEE Transactions on Emerging Topics in Computational Intelligence,2021,5(5):726-742.
[21]QIU Y,LENG J,GUO C,et al.Adversarial defense throughnetwork profiling based path extraction[C]//Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition.2019:4777-4786.
[22]SIMONYAN K,ZISSERMAN A.Very Deep Convolutional Networks for Large-Scale Image Recognition[J].arXiv:1409.1556,2014.
[23]RUSSAKOVSKY O,DENG J,SU H,et al.Imagenet large scale visual recognition challenge[J].International Journal of Computer Vision,2015,115:211-252.
[24]DESAI S,RAMASWAMY H G.Ablation-CAM:Visual Explanations for Deep Convolutional Network via Gradient-free Localization[C]//Workshop on Applications of Computer Vision.IEEE,2020:972-980.
Viewed
Full text


Abstract

Cited
  Shared   
  Discussed   
No Suggested Reading articles found!
渝ICP备16013121号-4
地址：重庆市渝北区洪湖西路18号    邮编：401121  
电话：023-63500828（编辑部） 023-67039612（会议/专题） 023-67039623（财务/发票）
E-mail：jsjkx12@163.com
本系统由北京玛格泰克科技发展有限公司设计开发