基于脈沖神經(jīng)元膜電位增量的數(shù)據(jù)分布統(tǒng)計量及批歸一化

打開文本圖片集

關(guān)鍵詞：脈沖神經(jīng)網(wǎng)絡(luò)；批歸一化；脈沖時間依賴性；脈沖神經(jīng)網(wǎng)絡(luò)訓(xùn)練算法中圖分類號：TP183 文獻標(biāo)志碼：A 文章編號：1001-3695（2025）08-013-2341-07doi：10.19734/j.issn.1001-3695.2024.12.0529

Data distribution statistics and batch normalization based on membrane potential increments of spiking neurons

LiWeiqi1，ChenYunhual?，ChenPinghua1，Zhu Chunjia2 （1.SchoolofComputerSience&Tchologyuangdong UniersityfTchologyuangzhou5o，hina;2.ChinaSibldingR search&DesignCenter，NeijiangSichuan641199，China）

Abstract：SNN has garnered significant atention due to theirlow power consumptionand high-speedcomputation，stemming fromtheiravoidanceof multiplicationoperations.However，substantial challnges remainin theareas of training algorithms， hyperparameter tuning，andarchitecture designforSNN.Adressng the limitations of existing BNmethods inefectively handling temporal dependencies，thispaperanalyzed thepropagationof membrane potential incrementsacrosstimesteps.The proposed method computed the spatio-temporal accumulationof membrane potential increments step-by-stepas statistical measuresfornormalizingdataateachtimestep.Furthermore，itintroducedanexponentiallweighted moving average tocompute thespatio-temporalaccumulationofmembranepotentialincrements，foringaspatio-temporalatenuationcumulativebatch normalization（STBN）method.Experimentalresultsonthe CIFAR-10，CIFAR-10O，and CIFAR10-DVS datasets demonstrate that the proposed method significantly improves network clasificationaccuracyandreduces latency.Notably，onthe CIFAR100 dataset， the method achieves an accuracy of 76.30% using only two time steps， representing a 3.43% improvement over the previous best algorithm for similar models.

Key words：spiking neural network（SNN）；batch normalization（BN）；temporal dependencies inspiking；training algo rithmsfor spikingneural network

0 引言

由腦科學(xué)啟發(fā)設(shè)計出的脈沖神經(jīng)網(wǎng)絡(luò)（SNN）[1～3]作為第三代人工神經(jīng)網(wǎng)絡(luò)，有著更好的生物解釋性。（剩余17494字）