Hannover 2020 – wissenschaftliches Programm
Die DPG-Frühjahrstagung in Hannover musste abgesagt werden! Lesen Sie mehr ...
Q 33.7: Vortrag
Mittwoch, 11. März 2020, 15:45–16:00, e001
Unsupervised phase discovery with deep anomaly detection — •Korbinian Kottmann1, Patrick Hümbeli1, Maciej Lewentein1,2, and Antonio Acin1,2 — 1ICFO, Avinguda Carl Friedrich Gauss, 3, 08860 Castelldefels — 2ICREA, Passeig de Lluís Companys, 23, 08010 Barcelona
We present a novel method for automated and unsupervised discovery of new and unknown phases in quantum many-body scenarios. Instead of supervised learning, where data is classified using labeled data, we perform anomaly detection, where the task is to differentiate a normal data set, composed of one or several classes, from anomalous data. We propose a scheme, employing deep neural networks, to map out the whole phase diagram. The method can be used completely unsupervised and automated to explore the entire phase diagram. As a paradigmatic example, we explore the phase diagram of the extended Bose Hubbard model in one dimension at integer filling. We compute the ground states using tensor networks and exemplarily use both unprocessed data like the central tensor and processed data like entanglement spectra that suffice to reproduce the phase diagram. The formulation of the method is independent of the nature of the data and could as well be used with physical observables, i.e. experimental data.