CPKD: Concepts-Prober-Guided Knowledge Distillation for Fine-Grained CNN Explanation

Explainable artificial intelligence (XAI) for CNN has received increasing research attentions, especially in security-critical domains. Some researches exploit knowledge distillation based technique to transform the sophisticated deep black-box model to a simple yet explainable model. However, traditional distillation-based methods suffer from two typical dilemmas during the explanation: (1) they can hardly be self-explanatory; (2) they are unable to provide concept-based fine-grained explanation, which causes that the explanation result is not straightforward and human-understandable. In this work, we forgo these dilemmas by proposing Concepts-Prober-guided Knowledge Distillation (CPKD), a novel distillation framework for general and fine-grained CNN explanation. CPKD employs techniques of concepts prober and distillation to generate a self-explanatory Concept-related Decision Tree (CDT) with semantic concepts explicitly organized at the tree nodes, thus bringing a concept-based explanation for CNN at a fine-grained and high-fidelity level. The comparison experiments and qualitative analysis result show that CPKD improves 3% of student model's accuracy than that of state-of-the-art concept-based explanation and is more effective in terms of interpretability than that of distillation-based explanation methods.