Berlin 2024 – scientific programme
Parts | Days | Selection | Search | Updates | Downloads | Help
AKPIK: Arbeitskreis Physik, moderne Informationstechnologie und Künstliche Intelligenz
AKPIK 1: Reservoir Computing & Neural Networks
AKPIK 1.3: Talk
Tuesday, March 19, 2024, 10:00–10:15, MAR 0.002
Analyzing phase transitions in minimal reservoir computers — •Davide Prosperino1, Haochun Ma1, and Christoph Räth2 — 1Allianz Global Investors, risklab, Seidlstraße 24-24a, 80335, Munich, Germany — 2Deutsches Zentrum für Luft- und Raumfahrt (DLR), Institut für KI Sicherheit, Wilhelm-Runge-Straße 10, 89081 Ulm, Germany
Minimal reservoir computers are powerful machine learning algorithms that can accurately predict nonlinear systems [1]. They differ from traditional feedforward neural networks by not relying on randomness but instead utilizing linear optimization, which enables them to operate on small training datasets and requires minimal computational resources.
Additionally, they can make accurate predictions for over ten Lyapunov times with certain parameterizations. However, we discovered that for certain parametrizations, the prediction fails. With only a few parameters, the phase transition between various parameterizations can be analyzed to comprehend the reasons behind the success of a prediction. We do that by analyzing the reconstructed, underlying equations.
[1] H. Ma, D. Prosperino, et al., Sci. Rep., 13, 12970 (2023)
Keywords: Machine Learning; Reservoir Computing; Chaos; Nonlinear Dynamics