忆阻神经网络联想学习电路设计与分析

发布时间：2018-10-05 13:23

【摘要】：现如今,人工神经网络一直是众多研究者关注的重大课题,但其发展大多基于神经信息演算的原理和法则,在冯诺依曼机上构造具有生物神经系统结构特征与功能的神经计算理论模型。但人工神经网络发展的最终目标在于,在神经细胞的水平上,研究并理解生物神经系统的智能原理以及智能行为,从而以生物神经系统的方式构造具有智能行为的智能机器,这与冯诺依曼体系计算机具有本质的区别,可定义为第五代智能计算机。在硬件水平上实现人工神经网络,最重要的问题是解决神经元电路与突触电路集成度与连接度。随着近年来在新型元器件的研究领域的发展,纳米级元件的出现为此带来了新的突破口。比传统CMOS更小尺寸的纳米级硅薄膜晶体管(Nc-Si TFT)的出现保证原有性能的基础上提升了反应速度,降低了功耗。突触电路作为神经元之间连接的重要部件,不仅担负着神经元间信息传递的功能,同时其自身应具有很高的可塑造性来以实现神经系统的记忆与学习。忆阻器的出现为此提供了基础,忆阻器是具有动态特性的电阻,阻值可依赖于激励电压来变化,同时具有低能耗、纳米级、长久记忆等特性,使其具有类似于生物神经突触连接强度与可塑造的特性,可用作为突触电路的基础部件存储突触连接强度。本文首先介绍了课题研究的背景,对忆阻器的发展以及应用进行了详细介绍,并对HP型忆阻器进行了SPICE环境下的仿真与性能分析;然后对人工神经网络的发展以及对人工神经网络发展有重要影响的联想学习算法进行了介绍;最后详细给出了纳米级TFT SPICE模型的过程,并进行了电气环境下的性能仿真。在此基础上为实现忆阻器突触电路的学习功能,建立了“整合—激发”型神经元SPICE仿真电路,分析了电路中每个部件的工作特性,完善了原始Mead神经元电路结构,并对电路的脉冲信号产生过程进行了SPICE仿真。同时结合TFT及忆阻器的特性提出了新型可实现HEBB学习的神经元突触电路结构,使突触电路更符合真实生物神经突触特征,同时提高了突触电路的控制灵活性与可扩展性。在应用此设计的基础上,实现了两个神经元所构成神经网络之间的基于单脉冲HEBB学习与平均激发率的学习规则。并进一步基于多个神经元的神经网络完成了Pavlov实验,证明了此神经系统结构设计在联想学习方面的可用性。同时,为证实此电路可提供与传统逻辑门电路相同的比较功能,进行了自学习基础上输入信号重合检测功能的仿真实验,实现传统电路基本功能是神经计算型智能计算机发展的重要基础。
[Abstract]:Nowadays, artificial neural network (Ann) has always been a major concern of many researchers, but its development is mostly based on the principles and rules of neural information calculus. A neural computational model with the structural characteristics and functions of the biological nervous system was constructed on the von Neumann computer. But the ultimate goal of the development of artificial neural networks is to study and understand the intelligent principles and behaviors of the biological nervous system at the nerve cell level, so as to construct intelligent machines with intelligent behavior in the way of the biological nervous system. This is essentially different from the Von Neumann system computer, which can be defined as the fifth generation intelligent computer. The most important problem in the implementation of artificial neural networks at the hardware level is to solve the integration and connectivity of neuronal circuits and synaptic circuits. With the development of new components in recent years, the appearance of nanometer components has brought a new breakthrough. The appearance of nanocrystalline silicon thin film transistor (Nc-Si TFT) which is smaller than the traditional CMOS can improve the reaction speed and reduce the power consumption. As an important part of neuron connection, synaptic circuit not only bears the function of transmission of information between neurons, but also has the ability to be molded to realize the memory and learning of the nervous system. The appearance of the resistor provides the basis for this. The resistor is a resistor with dynamic characteristics, the resistance value can depend on the excitation voltage to change, at the same time, it has the characteristics of low energy consumption, nanometer level, long memory, etc. It can be used as the basic part of the synaptic circuit to store the synaptic connection strength. In this paper, the background of the research is introduced, the development and application of the resistor are introduced in detail, and the simulation and performance analysis of the HP type resistor under the SPICE environment is carried out. Then, the development of artificial neural network and the associative learning algorithm which has important influence on the development of artificial neural network are introduced. Finally, the process of nano-scale TFT SPICE model is given in detail, and the performance simulation in electrical environment is carried out. On this basis, in order to realize the learning function of the memory device synaptic circuit, a "integration-excitation" type neuron SPICE simulation circuit is established. The working characteristics of each component in the circuit are analyzed, and the original Mead neuron circuit structure is improved. The pulse signal generation process of the circuit is simulated by SPICE. At the same time, combining the characteristics of TFT and amnesia, a new type of synaptic circuit structure for HEBB learning is proposed, which makes the synaptic circuit more consistent with the real biological synaptic characteristics, and improves the control flexibility and extensibility of the synaptic circuit. On the basis of this design, the learning rules based on monopulse HEBB learning and average excitation rate between two neural networks are realized. Furthermore, the Pavlov experiment based on neural network with multiple neurons is carried out, which proves the usability of the neural structure design in associative learning. At the same time, in order to prove that the circuit can provide the same function as the traditional logic gate circuit, the simulation experiment is carried out to detect the coincidence of input signal on the basis of self-learning. Realizing the basic functions of traditional circuits is an important basis for the development of neural computing intelligent computer.
【学位授予单位】：重庆大学
【学位级别】：硕士
【学位授予年份】：2015
【分类号】：TN702;TP183

【相似文献】