基于脉冲序列核的递归脉冲神经网络突触权值-延迟学习算法

打开文本图片集

关键词：递归脉冲神经网络；监督学习；突触延迟学习；脉冲序列核中图分类号：TP183 文献标志码：A 文章编号：1001-3695（2025）12-011-3611-08doi：10.19734/j.issn.1001-3695.2025.04.0134

Abstract：Recurent spiking neural networks（RSNNs）areaclassofbrain-likeintellgentcomputationalmodelswithfeedback loops thathave powefulcapabilities toleancomplex spatiotemporal pattrs，butitisstillchallengingtoconstructther eficientspiketrain-levelsupervisedlearingalgorithms.Synapticweightsanddelaysplayanimportantroleininfomation transmission of neurons，but most of the existing studies focus onlearning synaptic weights，andlearning andoptimizationof synapticdelaysarerelativelyinsuficient.Inresponse tothissituation，thispaperproposedasupervisedlearmingalgorithm basedonspiketrainkemelforsynapticweight-delayinRSNNstoimprovethelearningperformanceof thenetworkbysimultaneouslyoptimizing synaptic weightsand delays.Theresultsofspiketrain learning and UCI dataset clasification experments showthatthe dynamic synaptic delay learning algorithmhas higherlearning accuracyand requires fewer leaming epochs than the staticsynapticdelaylearning algorithm，indicating thatsynapticdelaylearningcansignificantlyimprovetheauracyof network training and accelerate network convergence.

Key words：recurrent spiking neural network；supervised learning；synaptic delay learning；spike train kernel

0 引言

脉冲神经网络（spikingneural network，SNN）是一种受生物神经系统启发的类脑计算模型，其核心特征在于使用离散的脉冲信号作为神经元之间信息传递的基本单元。（剩余21282字）