Making Explainable AI Understandable with Contextual Importance and Utility

The requirement for explainability in AI systems is as old as AI itself and lack of explainability has always been acknowledged as being a main hurdle for the success of AI in general.

The new rise of AI also lead to the launch of the eXplainable AI (XAI) concept and re-newed efforts for increased explainability.

In particular, so called black-box systems using neural networks and deep learning have made it necessary to develop new XAI methods.In the last few years, several so-called model-agnostic post-hoc XAI methods have emerged (LIME, Shapley values, ...) that belong to the family of Addiditve Feature Attribution (AFA) methods.

This project uses a different definition of "importance" than AFA methods and most other XAI methods.

The Contextual Importance and Utility (CIU) concept is a generalisation of importance and utility from the linear models of Decision Theory to the non-linear models of neural networks and AI in general.

CIU provides a solid mathematical framework that guarantees fidelity with the black-box model and can be used for producing textual, visual or graphica explanations with any level of abstraction.

CIU can also be used for validating the confidence, stability, robustness and reliability of the model.This project addresses the current concept confusion of the XAI domain, leverages the capabilities of CIU compared to the current state-of-the-art in XAI, and provides empirical results from the use of XAI and CIU in various domains.