Learning noise-induced transitions by multi-scaling reservoir computing

This paper explores the use of reservoir computing, a machine learning model, to learn noise-induced transitions in dynamical systems. The authors propose a multi-scaling approach that leverages a key hyperparameter to control the timescale of the reservoir dynamics, allowing it to effectively capture slow transitions amidst fast-scale noise. The method is demonstrated on various systems (bistable with white/colored noise, multistable) and experimental protein folding data, showing superior performance compared to conventional techniques like SINDy and recurrent neural networks in capturing stochastic transition statistics and even specific transition times for colored noise. The approach demonstrates potential for analyzing noisy time series beyond simply noise reduction.

Zequn Lin, Zhaofan Lu, Zengru Di, Ying Tang