Implementation of Learning Mechanisms on a Cat-Scale Cerebellar Model and Its Simulation

Abstract

We have built a large-scale spiking network model of the cerebellum with 1 billion neurons on a supercomputer previously. The model, however, did not incorporate synaptic plasticity such as long-term depression and potentiation at parallel fiber-Purkinje cell synapses. In this study, we implemented them on the model. To test the learning capability, as a benchmark, we carried out simulation of eye movement reflex called gain adaptation of optokinetic response (OKR). The present model successfully reproduced the increase of firing rate modulation of a Purkinje cell during simulated OKR training, resulting in the increase of OKR gain. The model completed a 6 s simulation within 4.4 s, suggesting realtime simulation even with the learning mechanisms. These results suggest that the present cerebellar model can now perform reservoir computing, a supervised learning machine for spatiotemporal signals, with very large reservoir composed of 1 billion neurons.