Liquid state machines Gaussian process.

Abstract

This study explores the integration of liquid state machines with Gaussian process regression, proposing a novel algorithm termed liquid state machine Gaussian process. Liquid state machines process information using spiking neurons in their reservoir, exhibiting complex dynamic responses that are well-suited for capturing intricate features in nonlinear time series. However, traditional readout layers in liquid state machines typically rely on linear regression or simple classifiers. This paper enhances the readout layer using the Bayesian framework of Gaussian processes, enabling output prediction and providing relevant predictive distribution information. Experimental evaluations on chaotic time series, Japanese vowel data classification, skeleton action recognition, and UR5 teaching tasks demonstrate that the liquid state machine Gaussian process method exhibits accuracy and robustness.