【作者】 张洋；

【导师】 王小平；

【作者基本信息】 华中科技大学 ， 控制科学与工程， 2017， 博士

【摘要】 人工神经网络理论发展至今,已经日趋成熟,其硬件上的实现有望构成非冯·诺伊曼结构的计算机系统。由于晶体管工艺的局限性,传统计算机系统中基于晶体管的集成电路体积已经缩小到了极限,很难再继续满足摩尔定律。用新型材料的电路设备取代晶体管已成为了研究热点。忆阻作为最有希望取代晶体管的新型设备之一,具有许多优良的性质,比如非易失性、高集成度、低功耗、良好的可扩展性等,被用于研发新型的存储器和神经计算系统。忆阻单层神经网络在芯片上的实现有望彻底改变计算机处理信息的方式,构建非冯·诺依曼计算机系统。而忆阻多层神经网络的设计仍然是一个难题,其应用前景十分广泛,可用于逻辑运算、图像处理、模式识别等领域。本论文首先根据某一特定材料制成的忆阻设备的测试结果,即忆阻真实设备两端施加正弦信号时电压电流变化的Ⅰ-Ⅴ曲线以及忆阻阻值在正负不同电压下随时间的变化曲线,以行为级建模为主,通过参数拟合的方式逼近忆阻真实设备所表现出来的特性。更进一步地,通过对相关模型参数的测试与优化,增加模型的兼容性使之可以模拟更多不同材料特性的忆阻设备。对单忆阻神经突触进行研究,用一个忆阻通过与固定的电阻连接到反相器,实现具有正负权值的神经突触。利用忆阻能通过控制所施加的脉冲数量来连续调整忆阻阻值的特性,突触权值由一个忆阻神经突触来表示并保存,在对突触权值调整过程中考虑到每一步的误差因素,构建对噪声容忍度高的忆阻神经网络电路。将对能用电路实现的忆阻神经网络算法进行研究,其中可以采用的算法有反向传播(BP)算法、赢者通吃算法、随机调整算法等。根据不同的算法可以搭建不同的忆阻神经网络电路。根据具体电路设计基于忆阻的神经网络学习训练算法,使忆阻神经网络可以在硬件上实现相应的逻辑运算和模式识别功能。本文从四个方面对基于忆阻的多层神经网络的设计和应用进行了较深入的研究,主要的创新工作及研究成果如下:(1)根据最新的忆阻真实设备的测试数据,建立一种新型的更加符合忆阻本质特性的忆阻数学模型和电路模型,该模型可以模拟多种不同材料的真实忆阻器设备;(2)充分利用忆阻的记忆特性和内在阻值随施加电压变化而改变的特性,设计新型的忆阻神经元突触,该突触电路体积更小、功耗更低、不受漏电流影响、能持续多次正确调整权值;(3)利用忆阻的非易失性、纳米级尺寸、以及功耗低等特性,设计一种新型的具有容错性好、稳定性强等优点的忆阻多层神经网络电路;(4)根据具体突触电路,设计新型的基于忆阻多层神经网络的算法,为忆阻多层神经网络的应用打下基础。将稳定的能够容忍噪声的忆阻多层神经网络应用到逻辑运算和模式识别功能中,在考虑误差因素的忆阻神经网络中实现学习训练功能,正确的完成相应的逻辑运算和模式识别功能,且设计的忆阻神经网络学习训练速度更快、误差更小、噪声容忍性更好。本文的研究成果可以为硬件实现更大规模的忆阻多层神经网络电路奠定基础,实现更加复杂更加具有实践意义的功能,比如图像识别中的人脸识别功能。将机器学习和深度学习最近的算法理论应用到忆阻神经网络的硬件中去,实现计算机视觉、语音识别、自然语言处理等其他领域的相关功能。更多还原

【Abstract】 Artificial neural network has been developed for many years.Up to now,it is increasingly mature.The hardware implementation of the artificial neural network is expected to realize the structure of non-Von neumann computer system.As the limit of the traditional transistor technology in Von Neumann computer system,the area of the transistor-based integrated circuits has been reduced to its minimum size and has become increasing difficult to meet the Moore’s law.More and more efforts have been invested in the research of new electronic devices to replace the transistor.One of the most promising candidates is the memristor,which has several advantages such as non-volatility,high density,low power,and good scalability.Memristor are mostly utilized in developing new memory and neuromorphic computing system.The realization of memristive on-chip singer layer neural networks makes it possible to change the way that the computers proceed the data and build the non-Von Neumann computer system.However,the design of memristive multilayer neural network is still a problem,which can be used in logic operation,image processing,and pattern recognition.In this dissertation,a new memristor model is build based on the experimental data of the memristor made of a certain material,which can match the I-V curves of both sinusoidal and repetitive sweeping inputs and the changes of the memristances by voltage pulses.Using behavior-level modeling,parameters are fit to match the characteristics of the real devices.Moreover,the parameters of the model are verified and optimized.The proposed model can therefore simulate memristors made of different materials.A single memristor-based synaptic array is presented,where both plus-polarity and minus-polarity connection matrices are realized by a single crossbar array and a simple constant-term circuit.The memristance of memristor can be adjusted continuously by the applied voltage pulse numbers,the synaptic weight can therefore be stored by one memristive synapse.Considering the error factors in each step of the weight adjustment,the memristive neural network is built with high noisy tolerance.In addition,the neural network algorithm is realized on chip,and several algorithms can be considered,such as back propagation(BP)algorithm,winner-take-all algorithm,and random adjustment algorithm.Different memristive neural networks are build based on different algorithms.The applications of logic operation and pattern recognition can be realized on chip by applying the learning algorithms in the memristive neural networks.In this dissertation,four aspects of design and applications of memristor-based multilayer neural network are deeply studied.The primary innovation and research results are as follows:(1)A new memristive model is build based on the experimental data of the recent memristive device,which can match the intrinsic characteristics of the memrisoter.The proposed model can also simulate memristors made of different materials;(1)A new memristive neural synapse is designed by making full use of the memory characteristic and the adjustable memristance of the memristor,which has the advantages of smaller area,lower power,no sneak currents,and continuous correct weight adjustments;(3)Memristive multilayer neural network with good error tolerance and stability is designed by making use of the memristive characteristics of non-volatility,nano-scale size,and low power.(4)The multilayer neural network algorithm is designed based on the synaptic circuit,which is the fundamental of the applications of the memristive multilayer neural networks.The logic operation and pattern recognition is realized in the memristive multilayer neural network with high error tolerance.The learning and training are proceeded correctly,considering the error factors,which has the advantages of faster training speed,smaller learning error,and better noisy tolerance.The research results of this dissertation can make foundation for hardware implementation of larger multilayer neural network circuits,achieving more complex and practical functions,such as the facial recognition in image recognitions.Machine learning and deep learning algorithm theory can be applied to the memristor-based neural network hardware,realizing functions in the field of computer vision,speech recognition,natural language processing,and other related areas.更多还原