Performance of Chaotic Time Series Prediction Using a Spiking Reservoir with an Excitation Layer and an Inhibition Layer

Abstract

In this study, a structural design is investigated to improve the stability of the autonomous prediction of chaotic time series using a spiking neural network (SNN) as a reservoir. A two-layer SNN reservoir composed of excitatory and inhibitory (E/I) neurons is constructed, and the prediction output is learned by updating only the readout weights online using FORCE. Using the Lorenz system as a benchmark, normalized root mean squared error (NRMSE) is evaluated by separating the training (teacher-forcing) and closed-loop autonomous segments; the results indicate that the two-layer architecture reproduces key waveform features and that the autonomous-segment prediction accuracy varies with the E/I ratio.