基于忆阻桥突触的神经网络电路研究及应用

发布时间：2018-11-26 18:30

【摘要】：1971年,蔡少棠教授通过电路理论的完备性提出了忆阻器的概念,它具有一系列的优良特性,如纳米级尺寸、非线性特性、掉电后信息非易失性等。因而在信息存储、控制电路、非线性电路、人工神经网络等领域有着广泛的应用前景。随着信息化时代的不断发展,人们迫切需要更加智能化和微型化的信息处理系统,通过模拟大脑神经系统构造人工神经网络,提供了可行的解决方案,并且一直是科学研究的一个热门领域。忆阻器的记忆特性类似于大脑神经网络中的突触功能,利用忆阻器有望构建更具仿生智能的神经网络系统,从而加速信息化处理的能力。本文针对忆阻值漂移现象,基于误差原理分析和实验论证,证明了利用双极性脉冲能够有效减少忆阻值漂移造成的误差。同时,设计了能够产生大小相等、极性相反对称脉冲的双极性脉冲电路,并将其应用于神经突触和神经网络。进一步,通过分析神经元和神经突触的原理,讨论设计了更加灵活的神经网络电路实现。本文主要工作包含了以下内容:(1)介绍了忆阻器模拟突触的原理和可行性,随后,分析了忆阻串并联电路的特性,包括串联结构和并联结构。基于忆阻器简单组合电路,进一步分析了忆阻桥突触电路的原理特性,包括四个忆阻器和五个忆阻器构成的桥突触结构。最后,结合细胞神经网络和忆阻桥突触结构,介绍了忆阻桥神经网络的结构和特点。(2)分析了忆阻桥模拟神经突触的原理,并分别讨论了线性和非线性惠普忆阻器模型下,突触模拟过程中产生的忆阻器阻值漂移现象。由于忆阻值漂移现象,会导致一定程度的模拟误差,推导了忆阻值漂移产生的机理,提出了用双极性脉冲的对称性减少这种误差。基于此设计了一种产生对称脉冲信号电路,将其用于忆阻器突触,减少了突触模拟误差,并进行了数值分析和仿真比较,验证了所提出方法的有效性。(3)基于双极性脉冲发生器的忆阻突触结构,将其与细胞神经网络相结合,构造优化的忆阻突触神经网络。由于减少了忆阻值漂移造成的误差,它的突触权值模拟更为精确。在细胞神经网络中,利用模板算子和图像像素的二值进行卷积能够实现一些图像处理功能,这种模板算子通常是数值构成的矩阵形式,因此将突触权值对应细胞神经网络进行图像处理中的模板算子,能够实现图像处理能力。与传统神经网络处理能力相比,本文中优化的神经网络显示出了更加优越的图像处理效果,通过Matlab仿真论证了该神经网络的有效性。(4)基于BP神经算法,设计了新型忆阻桥神经元与神经突触电路,它更灵活地实现了突触权值的更新。最后,构建了更加灵活的神经网络电路结构,通过巴普洛夫联想记忆实验仿真论证了该神经网络能够实现联想记忆的能力。
[Abstract]:In 1971, Professor Cai Shaotang put forward the concept of resistive device through the completeness of circuit theory. It has a series of excellent characteristics, such as nanometer size, nonlinear characteristics, non-volatile information after power down, and so on. Therefore, it has a wide application prospect in the fields of information storage, control circuit, nonlinear circuit, artificial neural network and so on. With the continuous development of the information age, people urgently need more intelligent and miniature information processing system. Through simulating the neural system of the brain to construct artificial neural network, it provides a feasible solution. And has been a hot area of scientific research. The memory characteristics of the resistor are similar to the synaptic function in the neural network of the brain. It is expected that the memory device can be used to construct a more bionic intelligent neural network system so as to accelerate the ability of information processing. Aiming at the phenomenon of amnesia drift, based on the error principle analysis and experimental demonstration, it is proved that using bipolar pulse can effectively reduce the error caused by amnesia resistance drift. At the same time, a bipolar pulse circuit which can generate symmetrical pulses of equal size and opposite polarity is designed and applied to neural synapses and neural networks. Furthermore, by analyzing the principle of neuron and synapse, a more flexible neural network circuit is designed. The main contents of this thesis are as follows: (1) the principle and feasibility of analog synapse of amnesia are introduced. Then, the characteristics of series-parallel circuit are analyzed, including series structure and parallel structure. Based on the simple combinatorial circuit of the memory bridge, the principle characteristics of the bridge synaptic circuit are further analyzed, including the bridge synaptic structure composed of four memristors and five memristors. Finally, combined with cellular neural network and memory bridge synaptic structure, the structure and characteristics of amnesia bridge neural network are introduced. (2) the principle of amnesia simulated synapse is analyzed, and the linear and nonlinear models of Hewlett-Packard amnesia are discussed, respectively. The phenomenon of amnesia resistance drift during synaptic simulation. The phenomenon of amnesia drift will lead to a certain degree of simulation error. The mechanism of amnesia drift is deduced and the symmetry of bipolar pulse is proposed to reduce the error. Based on this, a symmetrical pulse signal circuit is designed, which is applied to the synapse of amnesia, which reduces the error of synapse simulation, and makes numerical analysis and simulation comparison. The effectiveness of the proposed method is verified. (3) based on the mnemonic synaptic structure of bipolar pulse generator, an optimized mnemonic synaptic neural network is constructed by combining it with cellular neural network. Because of reducing the error caused by amnesia resistance drift, its synaptic weight simulation is more accurate. In cellular neural networks, some image processing functions can be realized by convolution of template operator and binary value of image pixels. This kind of template operator is usually a matrix form of numerical value. Therefore, the image processing ability can be realized by using the template operator to correspond the synaptic weight to the cellular neural network. Compared with the traditional neural network processing ability, the optimized neural network in this paper shows more superior image processing effect. The effectiveness of the neural network is demonstrated by Matlab simulation. (4) based on the BP neural algorithm, A new type of synaptic circuit of amnesia bridge neuron and nerve is designed, which can update the synaptic weight more flexibly. Finally, a more flexible neural network circuit structure is constructed, and the ability of the neural network to realize associative memory is demonstrated by the experimental simulation of Pavlov's associative memory.
【学位授予单位】：西南大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP183;TN60

【参考文献】