Cascading Class Activation Mapping: A Counterfactual Reasoning-Based Explainable Method for Comprehensive Feature Discovery

Key Points

• The research aims to develop a method for improving the interpretability of CNN classification by discovering comprehensive features.
• Introduced cascading class activation mapping technique for feature extraction.
• Applied counterfactual reasoning to enhance model interpretability.
• Evaluated the effectiveness against standard CNN interpretation methods.
• Demonstrated that the method reveals previously unnoticed features.
• Showed improved clarity in classification reasoning compared to traditional methods.
• Provided insights into the evidence utilized by CNNs for decision making.