Please join us for CIRES HDR, Eslam Zaher’s, PhD confirmation milestone seminar.

Robust Axiomatic Explainability for Black Box Models.

Speaker: Eslam Zaher

Abstract: As deep learning models become more complex, ensuring their trustworthiness and transparency through Explainability becomes paramount. Gradient-based feature attribution methods, while computationally efficient, often produce noisy saliency maps and are vulnerable to adversarial attributional attacks. This talk addresses these limitations from a data-manifold perspective. In the context of black-box vision models, I will present Manifold Integrated Gradients (MIG), a robust, path-based feature attribution method that aligns with the Riemannian structure of data manifolds. MIG overcomes traditional limitations, producing clear, perceptually aligned feature visualisations and showing robustness against targeted attributional attacks. The discussion will cover the development of the method, its empirical advantages, and future directions for enhancing the expressiveness and robustness of explanations.