【作者】 刘军；

【导师】 王会海；

【作者基本信息】 中南大学 ， 电子信息（专业学位）， 2024， 硕士

【摘要】 随着人工智能与大数据的飞速发展,基于传统计算机架构的神经网络在计算效率上遇到了瓶颈,具有阻变特性和记忆特性的忆阻器成为模拟神经突触的热门器件。然而,目前的忆阻器模拟突触通常表现出有限状态,存在忆阻值的浪费。本文提出一种离散忆阻神经网络,实现忆阻突触状态连续调节,将有效提高忆阻突触的表达效率,降低忆阻神经网络的资源消耗。引入离散差分理论,构建了一种内部状态离散化的离散忆阻器,通过周期性激励验证其三指纹特性,证明该离散忆阻器符合忆阻器的广义定义。根据离散忆阻器数学模型设计实现了基于FPGA的仿真器电路,并在周期性双极信号的激励下成功在示波器上显示出8字型磁滞回线,证明该离散忆阻仿真器的正确性和可实现性。将离散忆阻器与BP神经网络进行耦合,以忆阻器替代神经网络突触,构建离散忆阻BP神经网络。基于所提出的学习算法设计了一个针对MNIST-10手写数字识别的三层神经网络。设计简化的离散忆阻BP神经网络的FPGA电路,该电路在神经网络中增加调节电路,使得神经网络误差反馈调节信号能作用在离散忆阻器上。分别用以下三个实验对所设计的离散忆阻BP神经网络进行验证:(1)MATLAB模型对MNIST-10手写数字集进行学习,识别率为97.16%。(2)FPGA模型在Model Sim上对简易数字完成原位训练,并识别所有目标数字序列。(3)将在MATLAB上训练的简化离散忆阻神经网络映射到FPGA上,实现了简易字符系统识别。更多还原

【Abstract】 With the rapid development of artificial intelligence and big data,neural networks based on traditional computer architecture have encountered bottlenecks in computing efficiency.Memristors with resistive switching characteristics and memory characteristics have become a popular device for simulating neural synapses.However,current memristor-simulated synapses usually exhibit finite states,resulting in a huge waste of memristor value.This article proposes a discrete memristive neural network to achieve continuous adjustment of memristive synaptic status,which will effectively improve the expression efficiency of memristive synapses and reduce the resource consumption of memristive neural networks.This thesis introduces the discrete difference theory to construct a discrete memristor with a discretized internal state.Its three-fingerprint characteristics are verified through periodic excitation,proving that the discrete memristor meets the broad definition of a memristor.An FPGA-based simulator circuit was designed and implemented based on the discrete memristor mathematical model,and an 8-shaped hysteresis loop was successfully displayed on the oscilloscope under the stimulation of periodic bipolar signals,proving the correctness of the discrete memristor simulator.feasibility and achievability.The discrete memristor is coupled to the BP neural network,and the neural network synapse is replaced by the memristor to construct a discrete memristor BP neural network.Based on the proposed learning algorithm,a three-layer neural network for MNIST-10 handwritten digit recognition is designed.Design a simplified FPGA circuit of discrete memristor neural network,which adds an adjustment circuit to the neural network so that the neural network error feedback adjustment signal can act on the discrete memristor.The following three experiments were used to verify the designed discrete memristive BP neural network:1.The MATLAB model learned the MNIST-10 handwritten digit set,and the recognition rate was 97.16%.2.The FPGA model completes in-situ training on simple numbers on Model Sim and recognizes all target number sequences.3.Map the simplified discrete memristive neural network trained on MATLAB to FPGA to realize simple character system recognition.更多还原