Two-Stage Extreme Learning Machine for SLFNs in Regression
Hieu Trung Huynh, Yonggwan Won · ITC-CSCC :International Technical Conference on Circuits Systems, Computers and Communications · 2007
Neural network approach has been massively used in regression problem. However, collected data for training often include outliers which affect the final results. In this paper, we propose a new approach for outlier elimination in regression based on the extreme learning machine (ELM) called two-stage ELM. Training process consists of two stages. In the first stage, the single-hidden layer feed forward neural network (SLFN) is trained based on the ELM algorithm using the whole of training set. The trained SLFN is used to verify training patterns. Patterns corresponding to the outputs exceeding a rejection threshold are removed from the training set. Finally, the remainder of training set is used to train the SLFN again based on the ELM algorithm. One interesting observation is that, our approach is simple to detect and eliminate outliers and it can be able to deliver lower error than that of the normal ELM with fast learning speed if there exist outliers in the training set.