Thermodynamics-inspired explanations of artificial intelligence

Area of Science:

• Artificial Intelligence
• Machine Learning
• Scientific Computing

Background:

• Predictive machine learning models are increasingly used in science but often function as black boxes.
• Establishing trust in these models requires understanding their decision-making processes.
• Assessing the human interpretability of AI explanations is a significant challenge.

Purpose of the Study:

• To introduce a universal metric, interpretation entropy, for evaluating the human interpretability of linear models.
• To develop a model-agnostic method for generating optimally human-interpretable explanations for black-box AI models.
• To demonstrate the broad applicability of the proposed explanation method across various scientific domains.

Main Methods:

• Introduction of interpretation entropy as a universal measure of interpretability.
• Development of Thermodynamics-inspired Explainable Representations of AI (TIER-AI), a novel model-agnostic explanation technique.
• Application of TIER-AI to explain predictions from diverse black-box models.

Main Results:

• Interpretation entropy provides a quantifiable measure for assessing explanation interpretability.
• TIER-AI successfully generates optimally human-interpretable explanations for various black-box models.
• The method's effectiveness is demonstrated across molecular simulations, text, and image classification tasks.

Conclusions:

• Interpretation entropy offers a standardized approach to evaluating AI explanation interpretability.
• TIER-AI provides a powerful, model-agnostic tool for enhancing trust and understanding in complex AI systems.
• The developed methods have broad implications for the reliable deployment of AI in scientific research.