Dresden 2026 – scientific programme
Parts | Days | Selection | Search | Updates | Downloads | Help
SOE: Fachverband Physik sozio-ökonomischer Systeme
SOE 17: Focus Session: Physics of AI II (joint session SOE/DY)
SOE 17.7: Talk
Friday, March 13, 2026, 11:45–12:00, GÖR/0226
Understanding Generative Models via Interactions — •Claudia Merger1,2,3, Alexandre Rene2,4, Kirsten Fischer2,3, Peter Bouss2,3, Sandra Nestler2,3, David Dahmen2, Carsten Honerkamp3, Moritz Helias2,3, and Sebastian Goldt1 — 1SISSA, Trieste, Italy — 2Jülich Research Centre, Jülich, Germany — 3RWTH Aachen University, Aachen, Germany — 4University of Ottawa, Ottawa, Canada
Generative models have become remarkably powerful at reproducing complex data distributions. They can infer the characteristic statistics of a system from comparatively small datasets and even generate new, realistic samples. Yet, our understanding of what these models learn remains limited: which statistics do they capture, and how accurately? To address the first question, we translate the statistics learned by generative models into a central concept of statistical physics: interactions between degrees of freedom that describe how pairs, triplets, and higher-order groups coact to produce the observed statistics of a system. Using invertible neural networks, we extract these interactions directly from trained models, providing a microscopic description of their learned data structure. To assess how accurately these interactions are learned, we use an analytic theory of diffusion models that predicts the precision with which pairwise interactions can be inferred from finite datasets, quantifying how generalization depends on sample size, data hierarchy, and regularization. Together, these results provide a framework grounded in statistical physics to interpret and predict the behavior of modern generative models.
Keywords: Diffusion; Generalization; Interpretable AI; Interactions; Normalizing Flows