多种连接模型的忆阻神经网络学习  

Learning in memristive neural networks with various connection patterns

作　　者：李传东[1] 田园[1] 陈玲[1] 葛均辉[1] 

机构地区：[1]重庆大学计算机学院,重庆400044

出　　处：《重庆大学学报（自然科学版）》2014年第6期10-16,24,共8页Journal of Chongqing University

基　　金：国家自然科学基金资助项目(61374078)

摘　　要：忆阻器以其独特的非易失性、天然的记忆功能以及纳米级尺寸,在人工神经网络、信号处理和模式识别等方面展现了巨大的应用前景。采用了基于STDP学习规则的忆阻神经网络,运用了网络自适应突变以及网络拓扑结构变化的基因算法,其中包括隐藏层神经元个数,连接权重以及神经网络突触模型的变化。比较了基于HP线性忆阻器模型,非线性忆阻器模型以及阈值模型这3种不同忆阻器模型的忆阻神经网络,并提出了学习效果更好的混合型忆阻神经网络。Due to the useful properties of nonvolatile, memory and nanoscale, memristors have prospective promising applications in artificial networks, pattern recognition and signal processing. This paper exploits the learning rule of the memristive network based on spiking timing dependent plasticity （STDP） and uses the genetic algorithms with self-adaptation and variable topologies, which allows the number of hi^tden neurons,connection weights,and connectivity pattern to change self-adaptably. Three memristor models are respectively used as the synapse in the network,including HP linear memristor, non-linear memristor and threshold memristor. The comparison of the performance of the three memristive neural networks is ,presented,and the hybrid memristive networks＇ learning effects are analyzed.

关 键 词：忆阻器 突触可塑性 基因算法 拓扑变异 混合型忆阻神经网络 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程]