Berlin 2018 – wissenschaftliches Programm

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

DY: Fachverband Dynamik und Statistische Physik

DY 26: Condensed Matter Simulations augmented by Advanced Statistical Methodologies I (joint session DY/CPP)

DY 26.4: Vortrag

Dienstag, 13. März 2018, 11:45–12:00, BH-N 334

Diagnostics of neural networks for machine learning phases and phase transitions — •Philippe L. Suchsland and Stefan Wessel — RWTH Aachen University, Aachen, Germany

Machine learning schemes based on neural networks have recently been proposed as new tools for classifying phases of matter as well as detecting phase transitions. One motivation behind such proposals is the ability of appropriately designed and trained neural networks to identify patterns from a large set of data without having to explicitly describe them. We apply both supervised and unsupervised machine learning schemes to basic models of statistical physics. In particular, we consider the 2D Ising model in the presence of extended domain-wall configurations as well as the 2D XY model that exhibits a Kosterlitz-Thouless transition. In both cases, we identify the physical properties of the models that are relevant for the classification task. This reveals how reliable these schemes are with respect to minimizing the amount of preprocessing of bare sample configurations before feeding them into the learning process.