【作者】 赵亮；

【导师】 王小平；

【作者基本信息】 华中科技大学 ， 控制科学与工程， 2019， 硕士

【摘要】 脉冲神经网络被看作是第三代人工神经网络,也是科学家对生物神经网络更深层次的借鉴,具有实时、异步、高效的计算能力和更大的信息容量,有希望帮助人类达到更高的人工智能水平。忆阻器具有阻值非易失、可堆叠、可扩展、兼容CMOS等诸多优点,有望突破现有微电子技术的瓶颈,也为人工神经网络的电路实现提供了新的思路。本文将忆阻与脉冲神经网络相结合,探索了基于忆阻的脉冲神经网络在图像分类中的应用。针对现有忆阻数学模型无法同时准确表示不同实物忆阻特性这一问题,通过改进得到了一种比较通用的阈值型漂移速率自适应模型,给出了面向神经元和突触的忆阻参数。结合提出的忆阻脉冲神经元和类突触忆阻设计了STDP学习电路并扩展为交叉阵列结构,搭建了对应的赢者通吃电路。在尽可能体现电路特性的基础上,将设计的电路抽象为简单的处理模块用于在软件环境中设计忆阻脉冲神经网络,同时,详细地讨论了网络的其他设计细节和整个学习训练过程。将设计的脉冲神经网络用于图片分类任务,包括二分类和多分类,对分类结果进行分析并与已有的脉冲神经网络进行比较。本文的研究成果包括一种改进的阈值型漂移速率自适应忆阻模型、一种基于忆阻的脉冲神经元电路、一种基于忆阻的STDP学习电路和一种赢者通吃电路,并以此为基础在软件环境中设计了一种忆阻脉冲神经网络完成图像分类任务。本文对脉冲神经网络的纯电路实现进行了探索,在尽可能节约电路成本的基础上对网络的部分环节搭建了相应电路,软件环境中设计的忆阻脉冲神经网络和其他已有的脉冲神经网络相比具有一定优势。更多还原

【Abstract】 Spiking neural network(SNN),a reference to biological neural network in a deeper level,is regarded as the third generation of artificial neural network.SNNs have a real-time,asynchronous and efficient computing power and a greater information capacity.It is hopeful to help human reach a higher level of artificial intelligence.On the other hand,nanomemristor devices have many advantages,such as non-volatile resistance,high life length,stackable,extensible and CMOS-compatible.There is a possibility to break through the bottleneck of existing microelectronics technology with the memristor,which also bring a new thinking for the circuit implementation of artificial neural networks.In this thesis,the application of memristor-SNNs in image classification is explored based on the combination of memristor and SNN.A general model with adaptive drift rate and threshold is given through improving an existing model.This given model could more accurately represent the physical characteristics of some real memristors than others.The model parameters of the memristors used for neurons and synapses are adapted respectively.The STDP learning circuit is designed by combining the provided memristor-based leaky integrate and fire(MLIF)model and synaptic memristor and extended to a crossbar structure.The corresponding winner-take-all(WTA)circuit is constructed.On the basis of reflecting the circuit characteristics as much as possible,the designed circuit before is abstracted into a relative simple model for designing the memristor-SNN in software environment.At the same time,other design details and the learning process are discussed attentively.The designed memristor-SNN is applied to image classification tasks,including twoclassification and multi-classification.The results are analyzed and the comparison with other existing SNNs is given.The work of this thesis include an improved with-threshold drift rate adaptive memristor model,a MLIF circuit,a memristor-based STDP circuit and a WTA circuit.On this basis,a memristor-based SNN is designed for image classification in software environment.In this thesis,the pure hardware implementation of SNNs is explored.Corresponding circuits representing parts of SNN is constructed considering saving circuit cost as much as possible.The designed memristor-based SNN in software environment has certain advantages compared with others.更多还原