Theoretical guarantees for explainable machine learning

01/10/2025 to 30/09/2028

Explainable machine learning is often advertised as a universal tool where local post-hoc explanation algorithms can be applied to complex machine learning models in order to establish trust or reliability. This is the rationale behind attempts in legislation and EU regulation to achieve transparency through explainability. Similarly, one can observe that non-expert users of machine learning systems, for example, scientists of other domains than computer science, use explainable machine learning to “confirm” that their trained machine learning models are trustworthy. However, it is unclear whether and under which assumptions explainable machine learning can reliably be used for such purposes. The overall objective of this proposal is to develop and prove rigorous mathematical guarantees for feature attribution methods as provided by explainable AI, and to investigate which assumptions are necessary to achieve such guarantees.