Sean Tull - Towards Compositional Interpretability for XAI - IPAM at UCLA
Presenter
November 6, 2024
Abstract
Recorded 06 November 2024. Sean Tull of Quantinuum presents "Towards Compositional Interpretability for XAI" at IPAM's Naturalistic Approaches to Artificial Intelligence Workshop.
Abstract: I will present a mathematical framework for defining AI models and their interpretability and use this to argue for the explainability benefits of models with interpretable compositional structure, referred to as ‘Compositionally Interpretable (CI)’ models. The approach uses the mathematics of category theory and its intuitive graphical language of string diagrams, which can capture deterministic, probabilistic, and even quantum AI models.
I will analyse the interpretability of a range of models, including neural networks and transformers, as well as CI models such as rule-based models, causal models, and ‘DisCo’ models in NLP. I will show how CI models allow for several forms of behavioural explanation not possible for unstructured black boxes, based on 'influence arguments', 'diagram surgery' and 'rewrite explanations'. Finally, I will mention some of the ongoing work on developing Compositional Intelligence at Quantinuum. This is joint work with Robin Lorenz, Stephen Clark, Ilyas Khan, and Bob Coecke.
Learn more online at: https://www.ipam.ucla.edu/programs/workshops/workshop-iii-naturalistic-approaches-to-artificial-intelligence/?tab=overview