汪晶, 王君鹏, 孙文昊, 陈松. 用于脉冲卷积神经网络的神经形态处理VLSI架构设计[J]. 微电子学与计算机, 2020, 37(12): 1-5.
引用本文: 汪晶, 王君鹏, 孙文昊, 陈松. 用于脉冲卷积神经网络的神经形态处理VLSI架构设计[J]. 微电子学与计算机, 2020, 37(12): 1-5.
WANG Jing, WANG Jun-peng, SUN Wen-hao, CHEN Song. A neuromorphic hardware design of a spiking convolutional neural network[J]. Microelectronics & Computer, 2020, 37(12): 1-5.
Citation: WANG Jing, WANG Jun-peng, SUN Wen-hao, CHEN Song. A neuromorphic hardware design of a spiking convolutional neural network[J]. Microelectronics & Computer, 2020, 37(12): 1-5.

用于脉冲卷积神经网络的神经形态处理VLSI架构设计

A neuromorphic hardware design of a spiking convolutional neural network

  • 摘要: 传统的卷积神经网络在训练和识别阶段通常都需要用高能耗的GPU,无法应用到需要小型低功耗设备的移动应用场景中.本文设计了一种用于识别手写体的数字脉冲卷积神经网络神经形态硬件VLSI架构,根据脉冲神经网络设计的神经形态硬件系统只有在有输入脉冲到来时硬件才会进行相应处理,从而能达到很低的能耗.在识别MNIST数据集时,卷积神经网络识别精度为99.0%,使用该神经形态硬件的识别精度能达到98.46%.相比于相同硬件结构的传统卷积神经网络,平均能耗大大降低.

     

    Abstract: In recent years, neural network technology has developed rapidly and reached a level comparable to human recognition in the field of image recognition. Traditional convolutional neural networks usually require GPU with high energy consumption in the training and recognition stage, so they cannot be applied to mobile applications requiring small and low-power devices. This paper presents a neuromorphic hardware architecture of spiking convolutional neural networks for recognizing handwriting numbers. Because the neuromorphic only operates when there is a spike input, it can achieve a low power consumption accordingly. When recognizing the MNIST data set, the accuracy of the traditional convolution neural network is 99.0%, whereas the accuracy of the neuromorphic hardware is 98.46%. Therefore, the neuromorphic hardware achieved a comparable recognition accuracy while lowering down the power consumption greatly compared to CNN with a similar hardware architecture.

     

/

返回文章
返回