Regensburg 2025 – scientific programme
Parts | Days | Selection | Search | Updates | Downloads | Help
DY: Fachverband Dynamik und Statistische Physik
DY 8: Focus Session: Nonlinear Dynamics and Stochastic Processes – Advances in Theory and Applications II
DY 8.11: Talk
Monday, March 17, 2025, 18:15–18:30, H43
Weak generalized synchronization in random neural networks and its impact on time series forecasting — •hiromichi suetani1,2 and ulrich parlitz3,4 — 1Faculty of Science and Technology, Oita University, Oita, Japan — 2International Research Center for Neurointelligence , The University of Tokyo, Tokyo, Japan — 3Max Planck Institute for Dynamics and Self-Organization, Göttingen, Germany — 4Institute for the Dynamics of Complex Systems, Universität Göttingen, Göttingen, Germany
Time series forecasting is one of the important issues in data science, and approaches based on reservoir computing (RC), have been attracting attention. Previous studies have often suggested that the hyperparameter region at the so-called ``edge of chaos," provides optimal performance in time series forecasting. But this concept is problematic because generally a reservoir is a non-autonomous dynamical system driven by external inputs, it should be referred to as the ``edge of conditional stability" rather than the edge of chaos.
In this study, we argue that this is not just a matter of terminology, and that the edge of conditional stability does not provide optimal performance. For this purpose, we clarify the relevance of the concept of ``weak generalized synchronization (W-GS)." This study demonstrates that random neural networks driven by chaotic inputs exhibit W-GS and shows that the fractal nature of the GS function affects forecasting ability. We quantitatively compare the relationship between the characteristics of GS and those of RC, such as the information processing capacity, to elucidate the role of GS in RC.
Keywords: time series forecasting; random neural network; generalized synchronization; reservoir computing