A framework for causal concept-based model explanations.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Causal Inference

Background:

• Non-interpretable AI models pose challenges for understanding their decision-making processes.
• Post-hoc explainable AI (XAI) methods aim to provide insights into black-box models.
• Existing XAI approaches may lack sufficient fidelity or understandability.

Purpose of the Study:

• To propose a conceptual framework for causal concept-based post-hoc XAI.
• To ensure explanations are both understandable and faithful to the underlying AI model.
• To generate local and global explanations using concept interventions.

Main Methods:

• Developed a conceptual framework for causal concept-based XAI.
• Calculated the probability of sufficiency for concept interventions to generate explanations.
• Utilized a proof-of-concept model trained on the CelebA dataset for demonstration.

Main Results:

• Generated example local and global explanations based on concept interventions.
• Demonstrated understandability through a clear, causally interpretable concept vocabulary.
• Addressed fidelity by outlining framework assumptions and the importance of contextual alignment.

Conclusions:

• The proposed framework offers a method for creating understandable and faithful explanations in XAI.
• Causal concept interventions provide a robust mechanism for generating local and global insights.
• Aligning explanation context with generation context is crucial for reliable XAI.