Optimization Of Liquid State Machine Based On Synaptic Plasticity

Spiking neural network(SNN)is often called third generation artificial neural network.Compared with traditional neural network,spiking neural network adopts spiking neuron model which is more in line with biological network,and more complex synaptic models for information transmission.The use of spiking neurons makes the network have the ability of spiking coding which is considered as the basis for the ability of human brain to deal with complex information quickly.therefore,compared to frequency coding of traditional artificial neural network spiking code can improve the information processing capability of the network.Liquid state machine(liquid state machine,LSM)is a special kind of spiking neural network.Its network is recurrently connected,which is called reservoir.Model with similar connection structure is refer to reservoir computing.The reservoir helps the LSM to calculate in real time,so the network does not need the converge of the network state to get the ideal output.Spiking neural networks are connected by synapses,and synaptic plasticity refers to changes in the function of synapses,usually changes in the connection weights.In this paper,the liquid state machine model and synaptic plasticity were studied,and the optimization of the liquid state machine model was realized.Synaptic integration is the basis of information transmission of spiking neural networks,which includes spatial information integration and time information integration.Based on synaptic integration,this paper studies how to improve the liquid state machine in time pattern classification.From the perspective of spatial integration,spike-timing-dependent plasticity(STDP)is used to construct self-organized networks.Experiments show that the self-organized network can effectively improve the performance of the network.From the perspective of time integration,this paper studies the effect of synaptic integration parameters on network performance.The results show that larger synaptic integration parameters can improve the performance of the network for simple time pattern classification.The biological neural system consists of 80% of the excitatory neurons and 20% of the inhibitory neurons.Excitatory neurons play an important role in the transmission of information,and the inhibitory neurons are also an essential part of the biological network that helps the network to proceed smoothly.Based on the above study,the inhibitory neurons are added to LSM.In this paper,excitatory synaptic plasticity and inhibitory synaptic plasticity are combined to construct the self-organizing network.The experimental results showed that network with mixed synaptic plasticity could effectively inhibit the excitatory neurons to reduce unnecessary noise,which can improve the performance of LSM in the spike trains classification.