ODSL Forum: Interpretable neural network architectures and their application to quantum many-body systems

by Prof. Annabelle Bohrdt (University of Regensburg)

Friday 7 Jul 2023, 15:00 → 16:00 Europe/Berlin

Basement Seminar Room (Origins Building)

Abstract: In the past decade, machine learning tools, and in particular neural networks, have been firmly established as tools to analyze data in many different research communities. In this talk, I will provide an introduction to typical problem settings and data types in quantum many-body physics. I will introduce a fully interpretable neural network architecture, the correlator convolutional neural network (CCNN), which allows to directly extract the correlations most relevant for the classification task at hand. In this network architecture, the order of correlations considered serves as a hyperparameter, such that a systematic exploration of higher-order correlations is straightforwardly possible. I will then show how the CCNN can be applied to a variety of condensed-matter problems, such as distinguishing different theory approaches, studying non-equilibrium dynamics, and searching for relevant thermodynamic properties.