上海交通大学学报

上海交通大学学报（自然版） ›› 2014, Vol. 48 ›› Issue (07): 965-970.

• 自动化技术、计算机技术 • 上一篇    下一篇

跨越-侧抑制神经网络分析及其应用

杨刚1, 2,乔俊飞1
  

  1. (1. 北京工业大学 电子信息与控制工程学院, 北京 100124; 2. 华东交通大学 电气与电子工程学院,  南昌 330013)
     
  • 收稿日期:2013-07-02 出版日期:2014-07-28 发布日期:2014-07-28
  • 基金资助:

    国家自然科学基金重点项目资助(61034008)

Analysis of Span-Lateral Inhibition Neural Network and Its Application

YANG Gang1, 2,QIAO Junfei1
  

  1. (1. College of Electronic and Control Engineering, Beijing University of Technology, Beijing 100124, China; 2. School of Electrical and Electronic Engineering,
    East China Jiaotong University, Nanchang 330013, China)
  • Received:2013-07-02 Online:2014-07-28 Published:2014-07-28

摘要：

针对神经网络的结构设计, 根据仿生学原理提出一种跨越侧抑制神经网络(SLINN). 该多层网络结构引入了层内中间抑制神经元的侧向连接以及神经元在多个层内进行的信息传递. 通过分析网络的逼近能力证明网络的学习能力, 设计基于误差反向传播思想的梯度下降学习算法. 通过鲍鱼年龄预测回归问题的仿真实验表明, S-LINN在处理实际回归问题时, 不但能够保证较高训练精度,同时可获得更强的泛化能力.

 
 

关键词: 仿生学, 侧抑制, 神经网络, 通用逼近

Abstract:

A novel neural network model S-LINN was proposed based on the neurons connection and lateral inhibition mechanism in the cortex. Considering the multilayer and span connection of cortex neurons, the model was used to simulate the cerebral cortex structure. The approximation ability was verified throuth the universial approximate analysis. Meanwhile, a supervised algorithm based on the error backpropagation and gradient descent theory was developed to train the network parameters. Simulation results for the abalone age prediction demonstrate that the proposed model can achieve higher accuracy of approximation and generalization with a comparable compact network structure.
 

Key words: bionics, lateral inhibition, neural network, universal approximator

中图分类号: