文献类型：期刊文献

英文题名：Incremental and Decremental Extreme Learning Machine Based on Generalized Inverse

作者：Jin, Bo[1];Jing, Zhongliang[2];Zhao, Haitao[3]

机构：[1]East China Normal Univ, Sch Comp Sci & Software Engn, Shanghai 200062, Peoples R China;[2]Shanghai Jiao Tong Univ, Sch Aeronaut & Astronaut, Shanghai 200240, Peoples R China;[3]East China Univ Sci & Technol, Sch Informat Sci & Engn, Shanghai 200237, Peoples R China

年份：2017

卷号：5

起止页码：20852

外文期刊名：IEEE ACCESS

收录：;EI(收录号：20174104263468);WOS:【SCI-EXPANDED(收录号:WOS:000413942100020)】；

基金：This work was supported in part by the Key Program of Shanghai Science and Technology Commission under Grant 15JC1401700, in part by the NSFC-Zhejiang Joint Fund for the Integration of Industrialization and Information under Grant U1609220 and Grant U1509219, and in part by the Municipality Projects of Shanghai Science and Technology Commission under Grant 15511104700 and Grant 16DZ1100600.

语种：英文

外文关键词：Extreme learning machine; online sequential ELM; incremental ELM; decremental ELM; generalized inverse

摘要：In online sequential applications, a machine learning model needs to have a self-updating ability to handle the situation, which the training set is changing. Conventional incremental extreme learning machine (ELM) and online sequential ELM are usually achieved in two approaches: directly updating the output weight and recursively computing the left pseudo inverse of the hidden layer output matrix. In this paper, we develop a novel solution for incremental and decremental ELM (DELM), via recursively updating and downdating the generalized inverse of the hidden layer output matrix. By preserving the global optimality and best generalization performance, our approach implements node incremental ELM (N-IELM) and sample incremental ELM (S-IELM) in a universal form, and overcomes the problem of self-starting and numerical instability in the conventional online sequential ELM. We also propose sample DELM (S-DELM), which is the first decremental version of ELM. The experiments on regression and classification problems with real-world data sets demonstrate the feasibility and effectiveness of the proposed algorithms with encouraging performances.