SMN:Spatiotemporal Merge Network

As a powerful processing tool,Artificial neural networks(ANNs)have been used in many areas such as pattern recognition,control,robotics and bioinformatics.Their wide applicability has inspired researchers greatly to improve the performance of ANNs by studying the biological brain.The biologically more reasonable Spike Neural Network(SNN)makes more faithful use of biological characteristics to provide higher processing power.In recent years,SNN has gained greater momentum in the development of neuromorphic lowpower systems.Most of the existing spike supervised learning methods have achieved better performance and achieved great success.However,the existing SNN model is limited to be used only for processing spike time series feature data,while the ANN model is more used for processing spatial feature data.Based on the processing mechanism of biological neurons for spatio-temporal signals,we propose a model for simultaneous multi-scale spatio-temporal information learning in a network,the Spatiotemporal Merge Network(SMN).Specifically,the SMN model contains two branches: the SNN branch and the ANN branch.The SNN branch is used to extract temporal information,and the ANN branch is used to extract spatial information.In the training process,in order to integrate temporal and spatial information,the activation value of the hidden layer output of the ANN and the impulse response value of the hidden layer of the SNN are respectively scaled and accumulated.In addition,we also derive the SMN learning algorithm.This article tests the SMN model on the following UCI datasets: Iris,Breast Cancer Wisconsin,Liver Disorders,Statlog Landsat and Pima Indians Diabetes.In terms of the accuracy and iteration steps evaluation indicators,the experimental results found that the SMN model uses fewer iteration steps and achieves the best accuracy on most data sets;in order to analyze the convergence of the model,the results of MSE and generalization gap are compared and analyzed,and it is found that SMN has achieved the fastest convergence and the smallest generalization gap;in order to analyze the misclassification of SMN,BP and SNN models on each data set,we use confusion matrix,Macro-Recall,Macro-Precision and Macro-F1 as evaluation indicators,and it is found that the precision and recall of SMN on each data set are better than other models.This work,on the one hand,provides a new perspective for studying high-performance spiking neural networks with future brain-like computing paradigms,and effectively compensates for the shortcomings of SNN and ANN that only pay attention to information in a certain aspect of time or space.On the other hand,combining the advantages of spike neural networks and wide networks,this work provides new ideas for deep networks to prevent overfitting,reduce complexity,and improve generalization.