Output-targeted baseline for neuron attribution calculation
作者：
Highlights：
• Discussion of the potential of attribution baselines to improve attribution scores.
• Proposition of two primal baseline properties for Aumann-Shapley attributions.
• A general objective function for calculating an optimization baseline.
• A fast baseline calculating method by quadratic approximating.
• Neuron attribution applications on network pruning and adversarial defense.
摘要
•Discussion of the potential of attribution baselines to improve attribution scores.•Proposition of two primal baseline properties for Aumann-Shapley attributions.•A general objective function for calculating an optimization baseline.•A fast baseline calculating method by quadratic approximating.•Neuron attribution applications on network pruning and adversarial defense.
论文关键词：Convolutional neural networks,Network interpretability,Attribution methods,Shapley values
论文评审过程：Received 26 May 2022, Accepted 28 June 2022, Available online 1 July 2022, Version of Record 9 July 2022.
论文官网地址：https://doi.org/10.1016/j.imavis.2022.104516
