计算机工程 ›› 2020, Vol. 46 ›› Issue (3): 53-59.doi: 10.19678/j.issn.1000-3428.0054208

• 人工智能与模式识别 • 上一篇    下一篇

基于脉冲神经网络的迁移学习算法与软件框架

尚瑛杰, 董丽亚, 何虎   

  1. 清华大学 微电子与纳电子学系, 北京 100084
  • 收稿日期:2019-03-13 修回日期:2019-04-15 发布日期:2020-03-14
  • 作者简介:尚瑛杰(1992-),男,硕士,主研方向为脉冲神经网络;董丽亚,硕士;何虎,副教授。
  • 基金项目:
    国家自然科学基金(91846303)。

Transfer Learning Algorithm and Software Framework Based on Spiking Neuron Network

SHANG Yingjie, DONG Liya, HE Hu   

  1. Department of Microelectronics and Nanoelectronics, Tsinghua University, Beijing 100084, China
  • Received:2019-03-13 Revised:2019-04-15 Published:2020-03-14

摘要: 使用脉冲序列进行数据处理的脉冲神经网络具有优异的低功耗特性,但由于学习算法不成熟,多层网络训练存在收敛困难的问题。利用反向传播网络具有学习算法成熟和训练速度快的特点,设计一种迁移学习算法。基于反向传播网络完成训练过程,并通过脉冲编码规则和自适应的权值映射关系,将训练结果迁移至脉冲神经网络。实验结果表明,在多层脉冲神经网络中,迁移学习算法能够有效解决训练过程中收敛困难的问题,在MNIST数据集和CIFAR-10数据集上的识别准确率分别达到98.56%和56.00%,且具有微瓦级别的低功耗特性。

关键词: 脉冲神经网络, 迁移学习, 反向传播, 多层网络, MNIST数据集, CIFAR-10数据集, 低功耗

Abstract: Spiking Neuron Network(SNN) uses spike sequence for data processing,so it has the excellent characteristic of low power consumption.However,due to the immaturity of learning algorithm,the multilayer network training has difficulty in convergence.Utilizing the mature learning algorithm and fast training speed of the back propagation network,this paper proposes a transfer learning algorithm.The algorithm completes the training process based on the back propagation network and transfers the training results to the spiking neuron networks through the spike coding rules and the adaptive weight mapping relationship.Experimental results show that the transfer learning algorithm can effectively solve the convergence problem in the training process of multilayer spiking neuron networks.The recognition accuracy on the MNIST dataset and CIFAR-10 dataset can be up to 98.56% and 56.00% respectively,with low power consumption at the microwatt level.

Key words: Spiking Neuron Network(SNN), transfer learning, back propagation, multilayer network, MNIST dataset, CIFAR-10 dataset, low power consumption

中图分类号: