【摘要】 现有的忆阻神经网络存在学习速率慢、精度较低、电路复杂等问题，为了实现标准、高效的片上学习，设计了基于Ag/CeO₂/ITO忆阻器的片上学习神经网络。由一个晶体管和两个Ag/CeO₂/ITO忆阻器作为突触神经元，该突触结构具有更大的权值范围，Ag/CeO₂/ITO忆阻器的阈值特性简化了非破坏性读取的步骤和电路结构，神经网络通过四种电路单元实现无数模转换和无外部处理器的片上学习，每层神经网络可以在两个时钟周期内并行更新权值，该方法避免了由数据传输所造成的延迟、功耗和误差。最后，仿真验证设计的忆阻神经网络并应用于字符图像和鸢尾花识别，识别准确率均能达到95%以上，且忆阻器件的差异性对准确率的影响不大，证明其具有有效性和鲁棒性。更多还原

【Abstract】 The existing memristor neural network has the problems of slow learning rate, low accuracy and complex circuit. In order to achieve standard and efficient on-chip learning, an on-chip learning neural network based on Ag/CeO₂/ITO memristor is designed. A transistor and two Ag/CeO₂/ITO memristors are used as synaptic neurons, and the synaptic structure has a larger weight range. The threshold characteristics of Ag/CeO₂/ITO memristor simplify the steps and circuit structure of non-destructive reading. Each layer of the neural network can update the weights in parallel in two clock period. This method avoids the delay, power consumption and error caused by data transmission. Finally, the proposed memristor neural network is simulated and applied to the recognition of character images and iris flowers. The recognition accuracy can reach more than 95%, and the difference of memristor devices has little impact on the accuracy, which proves the effectiveness and robustness of the proposed memristor neural network.更多还原