WB-LRP: Layer-wise relevance propagation with weight-dependent baseline

Yanshan Li, Huajie Liang, Lirong Zheng

Abstract: Highlights•We discover only perturbations parallel to the weight-input plane affect outputs.•We extend DeepLift theory to cover various LRP algorithms with baseline derivations.•We propose the WB-LRP algorithm based on the perturbation parallelism constraint.•The WB-LRP algorithm provides instance-level explanations for CNNs and GNNs.•We generalize WB-LRP theory to create a framework for nearly all LRP algorithms.