Study On The Influence Of Dynamic Characteristic Of Spiking Neural Network On Liquid State Machine

In recent years,the Spiking Neural Network(SNN)has aroused widespread concern and attention of a large number of scholars owing to the fact that it makes the approximate biological neuron model as the information processing unit.Compared with traditional artificial neural networks based on analog signals,SNN utilizes the neuron spike time encoding method to represent and transmit information.Consequently,it can exhibit the richness dynamic behavior of real biological system.Liquid State Machine(LSM)is a neural computation model based on SNN.Its liquid layer(hidden layer)is a recurrent neural network with the neurons which are all-to-all connected.The results of the existing studies show that the LSM is more prominent in terms of computational speed and accuracy,which is due to the rich dynamic characteristics of SNN.Therefore,the study of the dynamic characteristics of SNN about how to affect the computation performance of the LSM is of great significance.In this paper,we have studied the influence of the dynamic characteristics of SNN on the computational performance of liquid state machine from two aspects: network firing patterns and self-organized critical dynamic behavior.The real biological neurons have many firing patterns.Both the spiking and bursting are the most typical firing patterns.Spike firing is a repetitive single firing state,while bursting is two or more spikes followed by a period of quiescence.In this paper,we have studied the influence of different firing patterns on the computational performance of liquid state machine.By constructing a simulation platform based on MATLAB,two benchmark tasks are designed: bionic signal reconstruction and classification of jittered spike trains.Our results show that neural networks with bursting activity have much better computational performance than those with spike firings.Self-organized critical dynamics is an important phenomenon of complex networks.The network with self-organized critical state has the characteristics of power-law distribution of avalanche size,entropy maximization and robustness,and so on.We have studied the relationship between different network(such as random network,network after STDP learning,network after STDP+IP learning)and the self-organized criticality(SOC)dynamic.Meanwhile,we have further explored the influence of self-organized criticality dynamic on computation capability of LSM.And we have found that the network shows the best computational performance when it is subjected to critical dynamical states.Moreover,network learning can result in network activity closer to the critical state and enhance the robustness of the critical state.These results can greatly promote the exploration and research on the relationship between the dynamic behavior and the performance of neural network.