基于EFAST的在线极限学习机节点剪枝方法

A Online Extreme Learning Machine Node Pruning Method Based on EFAST

针对在线极限学习机(OS-ELM)的隐藏层网络结构优化问题,设计了一种能自适应调整网络结构的在线极限学习方法(FOS-ELM)。该方法首先利用扩展的傅里叶振幅敏感度测试(EFAST),对OS-ELM中的各个隐藏层节点敏感度进行分析,再通过移除低敏感的隐藏层节点,从而达到对OS-ELM的网络结构进行优化的目的。从实验结果中分析,相比标准的OS-ELM,CEOS-ELM和HOS-ELM,在保证泛化精度的条件下,通过本文训练方法所需的隐藏层节点数均少于这3种方法。

In order to optimize the hidden layer structure of the extreme learning machine（OS-ELM）, this paper designs an adaptive method to adjust the structure of the OS-ELM network（FOS-ELM）. We employ the extended fourier amplitude sensitivity test（EFAST） method to analyze the sensitivity of nodes in hidden layer. by removing the lower sensitivity nodes, we will achieve a simplified OS-ELM structure. And then a recursive least square algorithm is designed to adjusted the parameters in the output layer. In this way, the structure of OS-ELM can be optimized. In experiment, we compared our method to the standard OS-ELM, CEOS-ELM, and HOS-ELM, it proved that our method performed better than these state-of-art methods.