Towards Compositional Interpretability for XAI

Watch a 53-minute lecture from IPAM's Naturalistic Approaches to Artificial Intelligence Workshop where Sean Tull of Quantinuum explores a mathematical framework for AI model interpretability. Dive into how category theory and string diagrams can be used to analyze deterministic, probabilistic, and quantum AI models through a compositional lens. Examine the interpretability characteristics of various models including neural networks, transformers, rule-based systems, causal models, and 'DisCo' models in NLP. Learn about novel approaches to behavioral explanation like influence arguments, diagram surgery, and rewrite explanations that become possible with Compositionally Interpretable (CI) models. Discover ongoing research in Compositional Intelligence at Quantinuum, presented in collaboration with Robin Lorenz, Stephen Clark, Ilyas Khan, and Bob Coecke.