Peer-to-Peer AI Network Meeting on “Small Data”

One of the key themes of the meeting was the interpretability of models, presented by Tom Splittgerber from the Applied Statistics working group at the Institute of Statistics at the University of Bremen, led by Prof. Dr. Werner Brannath. While models built for big data are often highly complex and difficult to interpret, small data allows for the development of models that remain understandable to humans. Splittgerber demonstrated how to design models that are more expressive than classical linear models (GLMs) while being more interpretable than neural networks (NNs).

In the second presentation, Prof. Dr. Klaus Eickel introduced new research initiatives at Fraunhofer Institute for Digital Medicine MEVIS. The focus was on an innovative approach to decomposing raw MRI data into scan-specific and scan-agnostic representations to enable improved data harmonization—a practical example of applying small-data strategies in the field of medical imaging.

To conclude, participants reflected on the recent activities of the Small Data working group and outlined next steps. These include organizing a scoping workshop to sharpen the group's thematic focus and initiating coordinated, externally funded research projects.