基于注意力的联想忆阻脉冲神经网络及其无监督图像分类应用  

An attention-based associative memristive spiking neural network and its application in unsupervised image classification

作　　者：邓泽坤 王春华 蔺海荣 邓全利 Yichuang SUN Zekun DENG;Chunhua WANG;Hairong LIN;Quanli DENG;Yichuang SUN(College of Computer Science and Electronic Engineering,Hunan University,Changsha 410082,China;School of Electronic Information,Central South University,Changsha 410083,China;School of Engineering and Computer Science,University of Hertfordshire,Hatfield AL109AB,UK)

机构地区：[1]湖南大学信息科学与工程学院,长沙410082 [2]中南大学电子信息学院,长沙410083 [3]School of Engineering and Computer Science,University of Hertfordshire,Hatfield AL109AB,UK

出　　处：《中国科学：信息科学》2024年第11期2554-2571,共18页Scientia Sinica(Informationis)

基　　金：国家自然科学基金(批准号:62271197,62201204);广东省基础与应用基础研究基金(批准号:2024A1515011910);中国博士后科学基金(批准号:2022M71104)资助项目。

摘　　要：无监督学习不需要对训练数据进行人工标注,在硬件系统的图像分类应用中具有重要价值.现有忆阻脉冲神经网络(memristive spiking neural network,MSNN)的无监督学习主要集中于前后神经元之间的突触调节过程,这导致学习仅限于局部神经活动而忽略了神经反应之间的内部联系.联想记忆是大脑实现记忆的重要方式,其通过无监督方式将不同刺激关联起来以建立互联的网络记忆.同时,人类视觉系统利用注意机制从海量信息中选择重要信息,以有效减少输入神经元的数量和神经网络的规模.本文提出了一种基于注意力的联想忆阻脉冲神经网络(attention-based associative memristive spiking neural network,AAMSNN)的全电路设计,并将其应用于无监督图像分类应用.其中,注意力编码模块和注意力选择模块启发于人脑的注意力机制,用于搜索并选择重要特征信息,减少AAMSNN的输入神经元数量.联想忆阻脉冲神经网络由巴甫洛夫联想忆阻交叉阵列构成,通过调节联想记忆权重实现无监督图像分类.与其他MSNN相比,AAMSNN具有更小的MSNN规模和更少的忆阻器数量,并实现了更优的无监督图像分类准确率.Unsupervised learning does not require manual annotation of training data,showing significant value in image classification applications of hardware systems.Unsupervised learning in the existing memristive spiking neural networks(MSNNs)mainly focuses on the synaptic adjustment process between anteroposterior neurons,limiting learning to local neural activity and neglecting internal connections between neural responses.Associative memory is a crucial way for the brain to achieve memory,which associates different stimuli through unsupervised learning to establish interconnected network memories.Meanwhile,the human visual system utilizes attention mechanisms to select important information from massive data,effectively reducing the number of input neurons and the scale of neural networks.This paper proposes a fully circuit-designed attention-based associative memristive spiking neural network(AAMSNN)and applies it to unsupervised image classification.Inspired by the attention mechanism,attention encoding circuits and attention selection circuits are designed to search for and select important information,reducing the number of input neurons of AAMSNN.The associative memristive spiking neural network module consists of Pavlovian associative memristive cross arrays,and achieves unsupervised image classification by adjusting the associative memory weight.The AAMSNN has smaller MSNN size and fewer memristors than other MSNNs,and achieves superior unsupervised image classification accuracy.

关 键 词：忆阻脉冲神经网络 注意力 联想记忆 无监督学习 图像分类 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术] TP183[自动化与计算机技术—计算机科学与技术]