基于多子网络预训练的脉冲神经网络分类模型  

Spiking Neural Network Classification Model Based on Multi-subnetwork Pre-training

作　　者：卓明松 莫凌飞[1] ZHUO Mingsong;MO Lingfei(School of Instrument Science and Engineering,Southeast University,Nanjing 210096,China)

机构地区：[1]东南大学仪器科学与工程学院,南京210096

出　　处：《计算机科学》2024年第S02期33-38,共6页Computer Science

基　　金：江苏省高校青蓝工程。

摘　　要：脉冲神经网络(Spiking Neural Network,SNN)被认为是最符合生物大脑机制的类脑计算模型,凭借其事件驱动、高能效、可解释等特点吸引了越来越多的研究关注。然而,由于脉冲的二值输出与不可微分性,SNN的训练方法仍存在一定空缺。于是借鉴皮层记忆单元通过局部网络存储记忆信息的方式,提出一种基于多子网络预训练的脉冲神经网络分类方法。该方法使用样本标签信息优化了脉冲序列特征提取过程,采用改进的脉冲时间依赖可塑性学习规则预训练多个单类别特征提取子网络,并将预训练后的子网络进行无监督特征融合,有效提高了网络的特征分类能力。此外,在权重可视化与t-SNE可视化工具的帮助下,分析了方法的有效性。所提方法在MNIST与Fashion-MNIST数据集上分别取得了97.40%与88.81%的分类准确度。Spiking neural network(SNN)is widely regarded as the most biologically plausible model,aligning closely with the mechanisms of the biological brain.It has garnered increasing research attention due to its event-driven nature,high energy efficiency,and interpretability.However,the training methods of SNN still have some limitations due to the binary output and non-differentiability of the spike.This paper proposes a SNN classification method based on multi-subnetwork pre-training,which draws inspiration from the way cortical memory units store memory information through local networks.This approach leverages sample label information to optimize the feature extraction process,employs enhanced spike-timing-dependent-plasticity learning rules for pre-training multiple single-class feature extraction subnetworks,and conducts unsupervised feature fusion on these pre-trained subnetworks to effectively enhance the network’s feature classification capability.Furthermore,the effectiveness of this method is analyzed through weight visualization and t-SNE visualization tools.Finally,the classification accuracy of 97.40% and 88.81% is achieved on MNIST and Fashion MNIST datasets,respectively.

关 键 词：脉冲神经网络 脉冲时间依赖可塑性 单类别特征提取子网络 无监督特征融合 类脑计算 

分 类 号：TP389.1[自动化与计算机技术—计算机系统结构]