摘要
递归神经网络是一种非线性动力系统,被广泛应用在联想记忆.递归神经网络的激活函数是典型的分段线性函数,其硬件实现繁琐.忆阻器是非线性的动态纳米级器件,将其用于神经网络,可以实现低功耗、易扩展的忆阻激活函数电路.本文利用忆阻器磁通量和电荷之间的关系特性,提出了一种基于忆阻激活函数的递归神经网络,并将其应用到联想记忆中.首先,利用忆阻器设计出二值、三值忆阻激活函数,为递归神经网络集成电路的实现提供了有效的方法.其次,加入矩阵转移函数,提出了一种基于忆阻激活函数的递归神经网络模型.最后,实现了二值、三值静态图像的自联想记忆和异联想记忆,并且通过仿真实现了字母图像中的字母移动,因而实现动态图像的联想记忆.该研究能简化递归神经网络集成电路,并且为实现记忆视频运动提供新的有效途径.
Recurrent neural network is a nonlinear dynamic system, which is widely used in the aspect of associative memory. The activation function of recurrent neural network is a typical piecewise linear function,and its hardware implementation is tedious. The memristor is nonlinear dynamic nanoscale device. If used for the neural network, it can achieve low power consumption and easy extension of the memristive activation function circuit. Using the relationship features between flux and charge of memristor, this paper puts forward a recurrent neural network based on the memristive activation function and applies it to the associative memory. First of all,the two value and the three value memristive activation function is designed, which provides an effective method for the realization of the recurrent neural network integrated circuit. Secondly, a recurrent neural network model based on memristive activation function is proposed by joining the matrix transfer function. Finally, we realize the simulation of auto-associative memory and hetero-associative memory of the binary and three values static image, and then the moving of the letters in the alphabet image, which can be considered as associative memory of the dynamic image. The study can simplify the recurrent neural network integrated circuits, and provide a new effective way for the realization of the video memory sports.
出处
《中国科学：信息科学》
CSCD
北大核心
2017年第9期1226-1241,共16页
Scientia Sinica(Informationis)
基金
国家自然科学基金(批准号:61372139
61571372
61672436)
中央高校基本科研业务费专项资金(批准号:XDJK2016A001)资助项目
关键词
递归神经网络
忆阻器
激活函数
多值
联想记忆
recurrent neural network
memristor
activation function
multiple-valued
associative memory
