Improving liquid state machine with hybrid plasticity

Abstract

Liquid state machine (LSM) is a recently developed computational network model based on spiking neural network (SNN), whose elements and structure are highly inspired from real biological neural network. However, the reservoir in most existing LSM is essentially a recurrent spiking neural network with a random synaptic weights and almost identical neural elements, which is in contrast to the behavior of the real biological network. In this paper, we propose a novel approach to develop the reservoir for LSM by combining two types of plasticity, which forming a self-organizing network (SON) composed of heterogeneous neurons with different behaviors and different degrees of excitability. The SON is constructed by refining synaptic connectivity and neuronal intrinsic based on two kinds of biological learning rules on cortical coding. The connectivity among neurons is self-achieved by spike-timing-dependent plasticity learning (STDP), where we consider two different types of STDP: e-STDP for excitatory synapse and i-STDP for inhibitory synapse. Our study shows that the proposed model can carry out temporal pattern classification tasks with high accuracy. Furthermore, we compare two different training methods of readout: linear regression and Fisher linear discriminant. It is shown that Fisher discriminant can be a substitute of linear regression especially for LSM with small size of network.