一种基于忆阻激活函数的递归神经网络及其联想记忆

发布时间：2018-05-14 20:14

本文选题：递归神经网络 + 忆阻器　；参考：《中国科学:信息科学》2017年09期

【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.
[Abstract]:Recursive neural network is a nonlinear dynamic system, which is widely used in associative memory. The activation function of recurrent neural network is a typical piecewise linear function. The resistor is a nonlinear dynamic nano-scale device, which can be used in neural network to realize low power consumption and easy to extend the circuit of memory activation function. Based on the relationship between magnetic flux and charge, a recursive neural network based on memory activation function is proposed and applied to associative memory. Firstly, the binary and ternary mnemonic activation functions are designed by means of a resistive device, which provides an effective method for the realization of the integrated circuit of recurrent neural network. Secondly, a recursive neural network model based on memory activation function is proposed by adding matrix transfer function. Finally, the self-associative memory and the hetero-associative memory of the binary and ternary still images are realized, and the letter movement in the alphabetic image is realized by simulation, so the associative memory of the dynamic image is realized. This research can simplify the integrated circuit of recurrent neural network and provide a new effective way to realize the motion of memory video.
【作者单位】：非线性电路与智能信息处理重庆市重点实验室;西南大学电子信息工程学院;
【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
【分类号】：TP183

【相似文献】