Interpretable Representations and Neuro-symbolic Methods in Deep Learning

Video recording

Current state-of-the-art machine learning methods impress with their capabilities for prediction, classification, or when solving even complex analytical tasks in the case of large language models. However, these methods often appear as a "black box" from the outside which makes it hard to understand how a result was achieved.
In this talk I will discuss several approaches to interpretability in machine learning. First, I will describe a method for representation learning which leads to an interpretable latent representation. Second, I will present our work at the interface between symbolic and sub-symbolic representations, so called neuro-symbolic methods, which enable a direct interpretation of a model's intermediate output. The talk concludes with a discussion of the relationship between interpretability and causality.

Jan Stühmer studied Computer Science at the Technical University of Dresden and received his PhD in Computer Science in 2016 at the Technical University of Munich (TUM) in collaboration with the California Institute of Technology (Caltech) and the TUM Institute for Advanced Study. Subsequently he was a visiting researcher at the Computer Science and Artificial Intelligence Laboratory at the Massachusetts Institute of Technology (MIT CSAIL). He joined Microsoft as PostDoc Researcher at Microsoft Research Cambridge and became a Research Scientist at the MR&AI Lab in Zurich before joining the Samsung AI Center Cambridge as Senior Researcher in 2020. Since 2022 he is junior group leader at the Heidelberg Institute for Theoretical Studies and Juniorprofessor at KIT.