借鉴大脑结构和信息处理基本原理，基于神经形态电路构建类脑计算芯片，是解决摩尔定律失效和冯·诺依曼架构“存储墙”问题的最佳方案之一。二端忆阻器具有高密度、低功耗、易扩展和丰富的阻变特性，十分适合作为神经形态器件构建类脑计算芯片。作为一个系统性工程，类脑计算芯片的构建需要神经形态器件、网络和算法方面的协同研究，故而本论文针对忆阻器在突触行为模拟、硬件网络构建和软件网络算法方面存在的问题，通过实验测量和模型仿真，对Ta2O5神经形态器件及网络展开系统性研究，主要内容和成果包括：（1）针对突触短期可塑性行为模拟问题，发展了高仿生度扩散型Ta2O5器件，有利于神经形态电路的构建。电压激励下，器件外部电流响应与突触短期可塑性定量模型相一致；器件仿真表明，阻变时器件内部粒子运动也与生物突触相似。（2）针对突触长短期可塑性行为模拟问题，发展了新型自掺杂结构Ta2O5器件，可用于构建新型神经形态电路。电压激励下，器件具有可控易失和非易失阻变行为，能实现突触的多种长短期可塑性行为模拟；实验和仿真表明，器件双导电细丝粒子源引发的阻变行为具有较高的生物仿真度。（3）针对神经形态器件网络构建问题，基于Ta2O5扩散型器件发展了选通器和神经元。通过电场诱导扩散型器件实现了双向选通行为，开关比 > 10^7，高阻态漏电流 < 10^-12 A（±0.25 V），有潜力作为选通器缓解网络漏电流问题；基于扩散型器件构建了简单化、小型化神经元，电路仿真表明，其能在多脉冲和单脉冲策略下完成可控泄漏-整合-发放行为，为构建大规模脉冲神经网络提供基础。（4）针对神经形态器件网络算法问题，发展了随机稀疏动量算法和长短期可塑性转化算法。基于非易失Ta2O5器件的阻变行为建立网络仿真器，研究了器件阻变和权重编程中非理想因素对网络性能的影响，针对非理想网络提出了随机稀疏动量算法，网络仿真表明，该算法可使多层感知器和卷积神经网络的图像分类精度分别从26.12 %和65.98 %提高到90.07 %和92.38 %，编程脉冲数分别降低90 %和40 %，训练速度为原来的3倍，且硬件友好性较好；基于自掺杂Ta2O5器件构建了带有长短期可塑性转化层的感知器网络和训练算法，网络仿真表明，其对带100 %强度噪声的图像分类精度高于单独感知器对50 %噪声的分类精度。上述研究成果在器件和网络、硬件和算法层面，为构建基于神经形态器件的类脑计算芯片提供了一定的技术基础。

One of the best ways to address the failure of Moore’s law and the “memory wall” issue of von Neumann architecture is developing brain-inspired computing chips, which is constructed by the neuromorphic circuits, and inspired from the structures and information processing principles of the brain. The advantages of two terminal neuromorphic devices, such as high density, low power, scalable architecture, and rich dynamics of the resistive switching, make them suitable to construct brain-inspired computing chips. As a systematic engineering, the construction of brain-inspired computing chips needs cooperative research in neuromorphic devices, networks, and algorithms. Therefore, on the issues of memristors in synaptic behavior emulations, hardware network constructions, and software network algorithms, this dissertation systematically investigates Ta2O5 neuromorphic devices and networks through experimental measurements and model simulations. The main contents and results of this dissertation are as follows:(1)For the emulation of synaptic short-term plasticity (STP) behaviors, a high bio-fidelity Ta2O5 diffusive device is developed, which is beneficial to the construction of neuromorphic circuits. Under voltage stimuli, the current response of the device agrees with the quantitative model of biological synapses in STP. Simulation results show that particle dynamics inside the device during the resistive switching are similar to that in biological synapse during STP. (2)For the emulation of synaptic short- and long-term plasticity (STP and LTP) behaviors, a Ta2O5 self-doping structure device is developed, which can construct novel neuromorphic circuits. Under voltage stimuli, the device has controllable volatile and non-volatile resistive switching behaviors. Several synaptic STP and LTP behaviors are emulated by the device. Experimental and simulation results show that particle dynamics of double conductive filament sources have relatively high bio-fidelity to that in bio-synaptic plasticity.(3)For the construction of hardware networks, selectors and neurons are developed based on Ta2O5 diffusive devices. Bidirectional threshold switching behaviors are induced by the electric field in a Ta2O5 diffusive device, the on/off ratio and the off current of the device at ±0.25 V are 10^7 and 10^-12 A, respectively, which indicate that the device can act as selectors to suppress the sneak current issue of networks. Moreover, a small and simple neuron circuit is constructed based on Ta2O5 diffusive devices. Simulation results show that, under a multi- or single-pulse scheme, controllable leaky-integrate-and-fire behaviors are performed by the neuron, which indicates that the neuron is a promising element in large scale spiking neural networks.(4)For the learning algorithms of device networks, a stochastic sparse learning algorithm with momentum adaptation (SSM) and a STP to LTP transition (S2L) training algorithm are developed. A network simulator is built based on non-volatile Ta2O5 devices. The effect of non-ideal properties in device resistive switching and weight programming to the network performance are studied. A SSM algorithm is proposed to train non-ideal networks efficiently. With this algorithm, simulation results show that the classification accuracy of multilayer perceptron (MLP) and convolutional neural network (CNN) based on non-ideal devices improves from 26.12% to 90.07% and from 65.98% to 92.38%, respectively. The total numbers of weight programming pulses decrease 90% and 40% in MLP and CNN, respectively. The convergence rates are both 3× compared with the condition without SSM. Meanwhile, the learning scheme has good hardware friendliness. Moreover, a perceptron network with S2L layer based on Ta2O5 self-doping devices is proposed. Under a S2L algorithm, simulation results show that the network can solve recognition tasks of the figure with random noise efficiently. The network accuracy to the figure with 100 % amplitude random noise is higher than single perceptron to the figure with 50 % amplitude random noisy. The aforementioned results provide a certain foundation for the construction of neuromorphic device based brain-inspired computing chips in devices, networks, hardware, and software aspects.