Explainable AI: Emmy Noether funding for LMU scientist

Enlarge

Today, artificial intelligence achieves impressive results in image and text analysis – yet how modern models arrive at their decisions often remains a black box. This is precisely where Johannes Paetzold's research comes in. He has been awarded an Emmy Noether grant by the German Research Foundation (DFG) to establish a group at LMU. Including the programme allowance, the funding amounts to 2.1 million euros over a period of six years.

At present, Vision-Language Models (VLMs) often lack an understanding of underlying physical relationships. Particularly in sensitive domains such as medicine, this lack of transparency and explainability is a central obstacle to their reliable use. Paetzold's goal is to advance modern Vision-Language Models so that they truly understand structures and relationships in images – and make their decisions intelligible to humans.

To this end, he combines methods from geometry, topology, and graph theory with current AI architectures. The underlying idea: geometric concepts – such as the shortest route on a map, or recognizing an object independently of its orientation – are intuitive to human cognition. When such concepts are built into AI systems, the resulting models should become more robust, more transparent, and easier to interpret.

The focus is on medical applications such as the early detection of eye and vascular diseases as well as cancer. The methods are broadly applicable, however – for example to satellite imagery. "In the long term, we hope to develop AI tools that provide doctors with reliable support in treatment decisions while remaining understandable to patients," says Paetzold.