Theoretical guarantees for explainable machine learning

• Applicant: Professorin Dr. Ulrike von Luxburg
• Subject Area: Methods in Artificial Intelligence and Machine Learning
• Term: since 2025
• Project identifier: Deutsche Forschungsgemeinschaft (DFG) - Project number 560788681

Machine learning models are increasingly used to support decision-making in social con- texts as well as for applications in science. At the same time, machine learning models have become extremely complex. Many users of machine learning systems do not trust predictions from "black box" functions, and even machine learning experts have a hard time discovering critical failure modes in their machine learning models. Explainable machine learning is often being suggested as a partial remedy: while decision functions might be highly complex and hard to describe globally, it might still be possible to partially explain or reason about their decisions locally after the decision has been made. The goal of this proposal is to develop and prove rigorous mathematical guarantees for local posthoc explanations, in particular feature attribution methods, and to investigate which assumptions are necessary to achieve guarantees. Theoretical guarantees are particularly important when it comes to high-risk applications in society where transparency is or will be required by law, or in applications of machine learning in science where the aim is to discover "the underlying truth" and thus we need to evaluate our tools critically to make scientifically valid claims.

• DFG Programme: Research Grants