Learning Algorithm For Deep Recurrent Spiking Neural Networks

By mimicking the hierarchical structure of human brain,Deep Recurrent Spiking Neural Networks(DRSNNs)can gradually extract features from low-level to high-level,and improving the performance for the processing of spatio-temporal information.However,due to the complex hierarchical structure and inherent non-linear mechanism of the recurrent spiking neural network,it is very difficult to construct an efficient deep learning method based on spike trains coding.Based on the above description,this paper uses the Recurrent Spiking Neural Machine(RSNM)module to construct a layered structure of DRSNNs,studies the unsupervised learning algorithm and the supervised learning algorithm based on the feedback alignment mechanism,and designs a real-time calculation model of epilepsy electroencephalography data based on spiking neural networks.The main research contents of this paper are as follows:First,based on spike trains coding,a new unsupervised multi-spike learning rule is proposed.We use this rule to train RSNM,which can realize the task of learning the complex spatio-temporal pattern of spike trains.The spike signal propagates forward first and then reconstructs it backward,and adjusts the synapse weight according to the reconstruction error.This unsupervised learning algorithm has been successfully applied to spike trains learning experiments.In the experiment,we analyzed the influence of different parameters such as the learning rate and the number of neurons in RSNM on the learning results.In addition,a layer-wise pre-training algorithm for DRSNNs is proposed,and the experimental results show that the algorithm has a lower reconstruction error.Secondly,according to the error back propagation mechanism of the spike trains,a deep recurrent spiking neural network supervised learning algorithm based on the feedback alignment mechanism is proposed.We use the feedback alignment mechanism,that is,in the process of error back propagation,a fixed random matrix is used to replace the synaptic weight in the feedforward structure to transmit the error and adjust the synaptic weight.Then the spiking trains learning process is used to verify the effectiveness of the algorithm,and the influence of different learning rate,connectivity,upper limit of random matrix value and input spike firing frequency on the learning result is analyzed.Experimental results show that the algorithm has good learning performance on spike trains learning tasks.Finally,a classification model of epilepsy electroencephalography data based on spiking neural networks is designed,and experiments are carried out on the CHB-MIT dataset.We preprocessed the original data,extracted the data needed in this research,and then used the Bens Spiker Algorithm to encode it into the spike trains,combined with the unsupervised learning algorithm and the supervised learning algorithm based on the feedback alignment mechanism.Trained on the test set,a classification result with a sensitivity of 94.17%,a specificity of 92.23%,and an accuracy of 93.20% is obtained.Compared with other classifiers,this model has better classification capabilities.