Towards counterfactual and contrastive explainability and transparency of DCNN image classifiers

作者：

Highlights：

• Proposed method extracts knowledge learned by DCNNs in the form of filters.

• Identifies features and concepts essential for the decision-making process.

• Using filters, contrastive and counterfactual reasoning behind decisions is provided.

• Does not alter inputs and keeps model integrity intact in provided explanations.

摘要

•Proposed method extracts knowledge learned by DCNNs in the form of filters.•Identifies features and concepts essential for the decision-making process.•Using filters, contrastive and counterfactual reasoning behind decisions is provided.•Does not alter inputs and keeps model integrity intact in provided explanations.

论文关键词：Explainable AI,Interpretable DL,Counterfactual explanation,Contrastive explanation,Image classification,DCNN

论文评审过程：Received 10 June 2021, Revised 5 September 2022, Accepted 13 September 2022, Available online 17 September 2022, Version of Record 30 September 2022.

论文官网地址：https://doi.org/10.1016/j.knosys.2022.109901