Dresden 2026 – wissenschaftliches Programm

Bereiche | Tage | Auswahl | Suche | Aktualisierungen | Downloads | Hilfe

AKPIK: Arbeitskreis Physik, moderne Informationstechnologie und Künstliche Intelligenz

AKPIK 5: Poster

AKPIK 5.9: Poster

Donnerstag, 12. März 2026, 15:00–16:30, P5

Understanding phase transitions in information processing systems using Geometric Thermodynamics — •Jonas Maximilian Müller, Ibrahim Talha Ersoy, and Karoline Wiesner — University of Potsdam

Phase transitions are well understood phenomena which arise in many fields of physics. Near the transition point very different systems show identical behaviour in accordance to their universality class. The framework of Geometric Thermodynamics allows for a more abstract approach to the transitions. We have shown, that neural networks undergo phase transitions related to accuracy hierarchies. However, the full extent of the analogy is unclear and there is no direct mapping of the critical phenomena described for DNNs to a thermodynamical framework. Luckily, both systems can at least locally be described using information geometry. In this study we characterise the transition phenomenology of well known thermodynamical systems using Geometric Thermodynamics. Using the Fisher metric we then construct a precise mapping between information processing systems, specifically DNNs, and thermodynamic systems in the proximity of the transition point. This mapping will in turn help us explore the full extent of the phase transition analogy for DNNs and better understand how they process information by leveraging the knowledge and techniques of Thermodynamics and Statistical Physics.

Keywords: Neural Networks; Phase transitions; Fisher information; Complexity science; Mathematical physics