基于忆阻器的多模式识别CNN电路设计

Design of memristor based CNN circuit for multi-pattern recognition

现有的类脑计算电路存在的识别对象单一、识别率低等问题,限制了类脑芯片的发展。该文结合器件量子电导及存算一体架构,提出一种基于忆阻器的多模式识别CNN电路设计方案。该方案首先构建具有多阻态、高精度、可重构的忆阻器模型;然后采用CNN架构与权重量化方法,提高电路的训练速度与识别率;最后,在交叉阵列结构中进行电导映射,完成MNIST、Fashion-MNIST和EMNIST等多种模式测试。实验结果表明所设计忆阻器具有50个稳定阻态以及LTP突触可塑性;CNN硬件电路的多模式识别率均为90%以上,尤其是MNIST识别率高达99.08%,并验证了高斯噪声下电路的抗干扰能力。

The main challenges of brain-like computing circuits include single object recognition and low recognition rate,which limits the development of brain like chips.This paper proposes a memristive multi-pattern recognition CNN circuit based on the quantum conductance of device and memory-computing integrated architecture.Firstly,a model with multi-resistance states,high precision and reconfigurability is constructed.Then,the CNN architecture and weight quantization design method are adopted to improve the training speed and recognition rate of the circuit.Finally,conductance mapping is performed in the cross-array structure,and multiple mode tests such as MNIST,Fashion-MNIST,and EMNIST are completed.The experimental results show that the proposed memristor has 50 stable resistive states and LTP synaptic plasticity.The multi-pattern recognition rate of CNN hardware circuit is above 90%,especially the recognition rate of MNIST is as high as 99.08%,and the anti-interference ability under Gaussian noise is verified.