计算机工程与应用 ›› 2016, Vol. 52 ›› Issue (8): 7-12.

• 理论与研发 • 上一篇    下一篇

逐层可加的急速学习机

郑雪辉,王士同,许小龙   

  1. 江南大学 数字媒体学院,江苏 无锡 214122
  • 出版日期:2016-04-15 发布日期:2016-04-19

Extreme learning machine with incremental hidden layers

ZHENG Xuehui, WANG Shitong, XU Xiaolong   

  1. School of Digital Media, Jiangnan University, Wuxi, Jiangsu 214122, China
  • Online:2016-04-15 Published:2016-04-19

摘要: 已有的急速学习机(Extreme Learning Machine)的学习精度受隐节点数目的影响很大。无论是已提出的单隐层急速学习机还是多隐层神经网络,都是先确定隐藏层数,再通过增加每一层的神经元个数来提高精度。但当训练集规模很大时,往往需要引入很多的隐节点,导致违逆矩阵计算复杂度大,从而不利于学习效率的提高。提出逐层可加的急速学习机MHL-ELM(Extreme Learning Machine with Incremental Hidden Layers),其思想是首先对当前隐藏层神经元(数目不大且不寻优,因而复杂度小)的权值进行随机赋值,用ELM思想求出逼近误差;若误差达不到要求,再增加一个隐含层。然后运用ELM的思想对当前隐含层优化。逐渐增加隐含层,直至满足误差精度为止。除此以外,MHL-ELM的算法复杂度为[l=1MO(N3l)]。实验使用10个UCI,keel真实数据集,通过与BP,OP-ELM等传统方法进行比较,表明MHL-ELM学习方法具有更好的泛化性,在学习精度和学习速度方面都有很大的提升。

null

关键词: 急速学习机, 多隐层神经网络, 逐层优化

Abstract: The number of hidden nodes has strong influence on the accuracy of ELM(Extreme Learning Machine). More hidden nodes are needed as the increase of the size of training data set. Either ELM or multi-hidden layer neural network determines the number of hidden layer in advance and then increases the number of nodes in every layer to achieve a smaller RMSE. Thus the computational complexity of the involved matrix becomes bigger, leading to worse learning efficiency. In this paper extreme learning machine with incremental hidden layers is proposed, in which the weights of hidden nodes can be assigned randomly to the current hidden layer nodes(small number, don’t optimize and small complexity) and then the corresponding RMSE is obtained like ELM. MHL-ELM increases a hidden layer unless the RMSE of the network reaches the request. Besides, the complexity of MHL-ELM is[l=1MO(N3l)]. The experiment shows MHL-ELM has better generalization performance, smaller RMSE and shorter learning time, compared with some traditional algorithms like BP or OP-ELM based on ten UCI, keel data sets and real data sets.

Key words: Extreme Learning Machine(ELM), multi-hidden layer neural network, layer-wise pre-training