| 基于神经元晶体管和忆阻器的 Hopfield 神经网络及其在联想记忆中的应用 |
| |
| 引用本文： | 朱航涛,王丽丹,段书凯,杨婷.基于神经元晶体管和忆阻器的 Hopfield 神经网络及其在联想记忆中的应用[J].西南农业大学学报,2018,40(2):157-166. |
| |
| 作者姓名： | 朱航涛 王丽丹 段书凯 杨婷 |
| |
| 作者单位： | 西南大学 电子信息工程学院/非线性电路与智能信息处理重庆市重点实验室，重庆 400715 |
| |
| 基金项目： | 国家自然科学基金项目(61571372)；国家自然科学基金面上项目(61672436)；重庆市基础科学与前沿技术研究专项重点项目(cstc2017jcyjBX0050)；教育部中央高校基本科研业务费专项创新团队项目(XDJK2014A009)；教育部中央高校基本科研业务费专项创新团队项目(XDJK2016A001) |
| |
| 摘 要： | 随着人工智能的高速发展,越来越多的神经元模型相继被提出,现有的神经元电路主要由普通晶体管、运算放大器等高功耗器件构成,存在结构复杂、集成度不高、兼容性差、功耗高、阈值调节难度高的缺点.针对以上不足,首次提出了一种全新的神经元结构,该结构仅由神经元晶体管、忆阻器和普通电阻构成,相比传统神经元电路,不包含复杂的差分运算电路以及电流与电压信号的转换电路,电路结构简单,同时具有良好的电路兼容性,可用于大规模集成.该结构利用神经元晶体管的加权求和特性以及阈值可控功能来模拟神经元信息传导过程,同时利用阈值忆阻器的阈值特性和阻值连续变化能力来设定和更新突触权值,使得该新型神经元结构不仅能实现传统神经元电路功能的同时,还具有能耗低、阈值动态可控、权值可编程的优点,不仅极大地简化了网络结构,还能加强网络性能.其次,还提出了基于这种新型神经元结构的忆阻离散Hopfield神经网络,该忆阻神经网络有助于促进人工神经形态系统的硬件实现,使神经网络系统能耗降低,集成度极大地提高,将这种网络运用在联想记忆和彩色数字图像恢复中,进一步说明了基于全新神经元结构的忆阻离散hopfield神经网络的实用性以及有效性.
|
| 关 键 词： | 神经元晶体管 忆阻器 Hopfield神经网络 联想记忆 |
A Hopfield Neural Network Based on Neuron Transistor and Memristor and Its Application in Associative Memory |
| |
| ZHU Hang-tao,WANG Li-dan,DUAN Shu-kai,YANG Ting.A Hopfield Neural Network Based on Neuron Transistor and Memristor and Its Application in Associative Memory[J].Journal of Southwest Agricultural University,2018,40(2):157-166. |
| |
| Authors: | ZHU Hang-tao WANG Li-dan DUAN Shu-kai YANG Ting |
| |
| Abstract: | With the development of artificial intelligence, more and more models of neurons have been proposed. The existing neuronal circuits are made of some high power consumption devices such as transistors and operational amplifiers, which have the disadvantages of complex structure, low integration, poor compatibility, high power consumption and difficulty in threshold adjustment. Given this situation, the authors of this paper put forward, for the first time, a novel neuron structure that consists only of a neuMOS, memristors and resistances. Compared to the traditional circuits of neurons, this circuit has simpler structure. It has no differential operational circuits and switching circuits, but has good compatibility, low power consumption and VLSI application. This structure utilizes the weighted summation and threshold-controllable features of neo-MOS to simulate the synaptic transmission between neurons, and the threshold characteristics and continuous change of memristance to set and update synaptic weight. The new neuron structure can achieve the function of traditional neuronal circuits, and has the advantages of low power consumption, threshold control and weight programmability, thus greatly simplifying the network structure and enhancing network performance. In addition, this paper proposes a novel discrete Hopfield neural network based on the neuron structure called Mem-Hopfield. That promotes neural morphological system of neural networks hardware implementation with lower consumption and higher integration. The effectiveness and practicability of the new optimized neural circuit construction is proved further by applying the Mem-Hopfield neural network in associative memory and color digital image recovery. |
| |
| Keywords: | |
| 本文献已被 CNKI 等数据库收录！ |
| 点击此处可从《西南农业大学学报》浏览原始摘要信息 |
| 点击此处可从《西南农业大学学报》下载免费的PDF全文 |
|
