Zoom link: https://uio.zoom.us/j/67438761292
Speaker: Poushali Sengupta from IFI, UiO.
Abstract:
Explainability of Artificial Intelligence(AI) models is an important topic that can have a significant impact in all domains and applications from autonomous driving to healthcare. The existing approaches to explainable AI (XAI) are mainly limited to simple machine learning(ML) algorithms, and the research regarding the explainability-accuracy tradeoff is still in its infancy- especially for complex machine learning techniques. In this work, we introduce- Explainibility through Correlation Impact Ratio (ExCIR), a new approach aimed at achieving a balance between model explainability and accuracy in complex AI systems, for both scenarios of independent features and dependent features. We construct a lightweight data space derived from the original data to train the XAI model, establishing an environment wherein the accuracy of the lightweight model is the same as that of the original model. We introduce a new metric- Correlation Impact Ratio(CIR), such that the impact of the features on the model is explained while also incorporating the uncertainty associated with the actual contribution of those features on the model. We also offer an upper-bound analysis of the computational complexity of ExCIR, particularly concerning dependent features, revealing its efficiency in high-dimensional feature spaces with limited data. ExCIR distinguishes itself from existing approaches by demonstrating robust performance for complex models in scenarios characterized by limited data and high-dimensional feature spaces, while concurrently exhibiting satisfactory computational efficiency even in the presence of substantial feature interdependencies.