Understanding Individual Decisions of CNNs via Contrastive Backpropagation

  • 2018-12-05 16:43:04
  • Jindong Gu, Yinchong Yang, Volker Tresp
  • 6


A number of backpropagation-based approaches such as DeConvNets, vanillaGradient Visualization and Guided Backpropagation have been proposed to betterunderstand individual decisions of deep convolutional neural networks. Thesaliency maps produced by them are proven to be non-discriminative. Recently,the Layer-wise Relevance Propagation (LRP) approach was proposed to explain theclassification decisions of rectifier neural networks. In this work, weevaluate the discriminativeness of the generated explanations and analyze thetheoretical foundation of LRP, i.e. Deep Taylor Decomposition. The experimentsand analysis conclude that the explanations generated by LRP are notclass-discriminative. Based on LRP, we propose Contrastive Layer-wise RelevancePropagation (CLRP), which is capable of producing instance-specific,class-discriminative, pixel-wise explanations. In the experiments, we use theCLRP to explain the decisions and understand the difference between neurons inindividual classification decisions. We also evaluate the explanationsquantitatively with a Pointing Game and an ablation study. Both qualitative andquantitative evaluations show that the CLRP generates better explanations thanthe LRP.


Introduction (beta)



Conclusion (beta)