An attribution graph-based interpretable method for CNNs

Xiangwei Zheng, Lifeng Zhang, Chunyan Xu, Xuanchi Chen, Zhen Cui

Abstract: Highlights•A graph construction method is proposed to model CNNs as attribution graphs (At-GCs).•A classifier is built to reveal distinct kernel activation patterns across image categories in CNN.•A scoring network is built to assess node importance in At-GCs, mirroring kernel roles in CNN.