MCPNet++: Interpretable Classification Models via Multi-Level Concept Prototypes.

Area of Science:

• Artificial Intelligence
• Machine Learning Interpretability
• Computer Vision

Background:

• Current AI interpretability methods often focus on high-level semantics, limiting understanding of model decision-making.
• Explanations lacking lower- and mid-level semantic insights are not fully faithful or useful.
• A need exists for holistic interpretation methods that capture multi-level feature information.

Purpose of the Study:

• To develop a novel method for more comprehensive AI model interpretability.
• To autonomously discover meaningful concepts from multiple feature map levels.
• To bridge the gap between AI-discovered concepts and human perception.

Main Methods:

• Introduction of the Multi-Level Concept Prototypes Classifier (MCPNet) for multi-level interpretation.
• Development of MCPNet++ for application to both CNN and transformer architectures.
• Integration of a large language model (LLM) to align learned concepts with human understanding.

Main Results:

• MCPNet++ provides more comprehensive AI explanations by utilizing multi-level concept information.
• The method autonomously discovers meaningful concepts from feature maps.
• Discovered concepts show strong alignment with human perception and understanding.
• Model performance is maintained while enhancing interpretability.

Conclusions:

• MCPNet++ offers a more holistic and faithful approach to AI model interpretability.
• Autonomous concept discovery from multi-level features enhances understanding of AI decision-making.
• LLM integration effectively bridges the gap between AI concepts and human perception.