Supervised Learning Algorithm For Spiking Neural Networks Based On Nonlinear Synaptic Interaction

In biology,the transmission of information between synapses is the direct or indirect transmission of stimulating signal from one cell to another through various neurotransmitters,resulting in certain changes in the information received by the receptor,and finally the input of each signal in the synapse reflects the characteristics of non-linear interaction.Therefore,non-linear synaptic interaction has become an important part of biological neural networks.It is also an important consideration in the simulation of biological computation and the design of learning algorithms.At present,the structure of most spiking neural networks basically contains input layer,hidden layer and output layer.Synapses are simplified to a linear structure in signal transmission between layers.In fact,the input signals to synapses in the biological nervous system show complex non-linear interaction characteristics.Aiming at the precisely timed spike trains,the synaptic transfer function is used to process the neural signals with temporal encoding,and the supervised learning algorithm of spiking neural network based on non-linear synaptic interaction is studied in this paper.Firstly,the synaptic plasticity mechanism of the biological nervous system and the expression of the spike trains kernel function of neurons are introduced.The spike train kernel function converts the discrete spike train into a continuous signal on the synapse through convolution operation.At the same time,the linear spike train kernel and non-linear spike train kernel and the synaptic learning rules of corresponding spike neurons are analyzed.Secondly,a supervised learning algorithm based on non-linear synaptic interaction for spiking neural networks is proposed.The non-linear synaptic interaction model based on the neural engineering framework is introduced combining with the characteristics of the spiking neural networks with temporal encoding.Through defining the error function of spike trains,the corresponding adjustment rules of synaptic weights are given.In the simulation,the supervised learning algorithm based on non-linear synaptic interaction is applied to the task of spike trains learning.The learning process of temporal and spatial patterns of the spike trains is analyzed,and the learning performance of different types of non-linear synaptic kernel is compared.The simulation results show that the non-linear synaptic kernel has higher learning accuracy and efficiency.In addition,the effects of different parameters such as the number of non-linear synapses,learning rate and simulation duration on learning outcomes are analyzed.Finally,combining with the characteristics of brain spatiotemporal data,the proposed supervised learning algorithm based on non-linear synaptic interaction is applied to the epilepsy classification,and the corresponding classification model is designed.The electroencephalogram signals collected from human brain cortex are encoded,and the signals of each channel of multiple samples are converted.The data are encoded into spike trains to show the changes of data,and the encoded spike trains are transmitted to the classification model as input information for effect test.In the epilepsy classification experiment,the proposed algorithm can achieve is higher of the average of classification accuracy than the existing learning methods.The experimental results of the spike trains learning task and the epilepsy classification show that it is great significance to design supervised learning algorithm of spiking neural network.It can also achieve better classification results in classification problems,and will be further applied to solve more practical problems.