Dresden 2026 – wissenschaftliches Programm
Bereiche | Tage | Auswahl | Suche | Aktualisierungen | Downloads | Hilfe
DY: Fachverband Dynamik und Statistische Physik
DY 14: Machine Learning in Dynamics and Statistical Physics II
DY 14.3: Vortrag
Montag, 9. März 2026, 15:30–15:45, HÜL/S186
Autoencoder Learning Dynamics on MCMC Ising Dataset — •Max Weinmann1,2,3 and Miriam Klopotek1,3 — 1University of Stuttgart, Stuttgart Center for Simulation Science, SimTech Cluster of Excellence EXC 2075, Stuttgart, Germany — 2University of Stuttgart, Interchange Forum for Reflecting on Intelligent Systems, IRIS3D, Stuttgart, Germany — 3Heidelberger Akademie der Wissenschaften, WIN-Kolleg, Heidelberg, Germany.
While consistent and abstract descriptions of learning dynamics in neural networks remain rare, they have become omnipresent and are used in many branches of science. As a result, predicting dynamics under diverse choices of ML model parameters can fail catastrophically and it remains difficult to mitigate these failures. Reliable control requires a deep understanding of the relevant mechanisms and conditions for learning particular kinds of datasets. Our study focuses on autoencoder architectures that perform well if they encode the dataset into a compressed representation that reflects core physical concepts of the data it is trained on, succeeding at a self-learned inverse model to decode this representation to reconstruct the original input (of physical origin). Some physical concepts are learned in a particular order, which depend on theoretical complexity of the representation and that of the ML architecture. We measure generalization ability against hard theoretical baselines and investigate the information geometry, stability, and physical interpretability of latent space over training time.
Keywords: Machine learning; Statistical Physics; Dynamical systems; Non-equilibrium; Generalization