Bonn 2025 – scientific programme
Parts | Days | Selection | Search | Updates | Downloads | Help
Q: Fachverband Quantenoptik und Photonik
Q 21: Quantum Optomechanics II
Q 21.7: Talk
Tuesday, March 11, 2025, 12:30–12:45, HS I
Training of neuromorphic systems based on coupled phase oscillators via equilibrium propagation: effects of network architecture — •Qingshan Wang1, Clara Wanjura1, and Florian Marquardt1,2 — 1Max Planck Institute for the Science of Light, Staudtstrasse 2, Erlangen, Germany — 2Department of Physics, University of Erlangen-Nuremberg, 91058 Erlangen, Germany
The increasing scale and resource demands of machine learning applications have driven research into developing more efficient learning machines that align more closely with the fundamental laws of physics. A key question in this field is whether both inference and training can exploit physical dynamics to achieve greater parallelism and acceleration. Equilibrium propagation, a learning mechanism for energy-based models, has shown promising results in physical systems with energy functions more complex than Hopfield-like models.
In this study, we focus on equilibrium propagation training of coupled phase oscillator systems.We investigate the influence of different experimentally feasible network architectures on the training performance. We analyze lattice structures, convolutional networks, and autoencoders, examining the effects of network size and other hyperparameters. Our findings lay the ground work for future experimental implementations of energy-based neuromorphic systems for machine learning, encompassing systems such as coupled laser arrays, CMOS oscillators, Josephson junction arrays, coupled mechanical oscillators, and magnetic systems.
Keywords: Equilibrium propagation; neuromorphic computing; XY model; neural network; coupled oscillator