Speaker: Nils Strodthoff (Professor for eHealth/AI4Health, Oldenburg University)

Title: Towards XAI 2.0: From feature interaction relevances to concept-based explanations"

Location: Click here to join the meeting (Microsoft Teams meeting)

Summary: In this talk I will talk about extension in two directions that go beyond conventional explainable AI methods, namely going beyond single-feature attributions through feature interactions and going beyond single-example explanations through concept-based explanations. In the first part of the talk, I will talk about our recent work on PredDiff [1], a perturbation-based attribution method that is firmly rooted in probability theory which share close connections to Shapley values. I will review the formalism as such, our extensions towards including feature interactions and implications for Shapley values. In the second part, I will discuss SSCCD [2], our recently proposed approach for concept-based attributions, where we put forward a new definition of concepts in terms of low-dimensional subspaces of the feature space as well as a constructive way of identifying them. The results will be illustrated on different image classification datasets.

[1] Stefan Blücher, Johanna Vielhaben, Nils Strodthoff. PredDiff: Explanations and Interactions from Conditional Expectations. https://arxiv.org/abs/2102.13519
[2] Johanna Vielhaben, Stefan Blücher, Nils Strodthoff. Sparse Subspace Clustering for Concept Discovery (SSCCD). https://arxiv.org/abs/2203.06043