Riemannian Integrated Gradients: A Geometric View of Explainable AI
arXiv:2503.00892v12 citationsh-index: 4GSI
Originality Synthesis-oriented
AI Analysis
This provides a geometric framework for explainable AI, but it is incremental as it extends an existing method to a new mathematical setting.
The paper tackled the problem of extending Integrated Gradients to Riemannian manifolds for explainable AI, resulting in a method that restricts to the original in Euclidean space and frames feature attribution as an eigenvalue problem.
We introduce Riemannian Integrated Gradients (RIG); an extension of Integrated Gradients (IG) to Riemannian manifolds. We demonstrate that RIG restricts to IG when the Riemannian manifold is Euclidean space. We show that feature attribution can be phrased as an eigenvalue problem where attributions correspond to eigenvalues of a symmetric endomorphism.