Chapter 6 - Outlier-based models of the non-tuned neural network—concept

Abstract

The extreme learning machine (ELM) is a form of single-layer feed-forward neural network. It has recently garnered much attention from scholars and academics due to its ability for rapid model training and development as well as its respectable generalization potential. However, irrespective of these advantages, there remains a great difficulty in modeling real-world applications with the ELM model. Real data sets encountered in practice often contain a heteroscedastic nature, which can result in an unreliable ELM model with a poor predictive capacity. In order to address this shortfall in the application of ELM models, this chapter will seek to present several ELM-based methodologies which are capable of handling such variation in the data sets. To begin the chapter, a brief review of the ELM technique, as well as its advantages and disadvantages (the reader is directed to Chapter 4 for a more detailed discussion of advantages and disadvantages), is presented. Following this, the mathematical formulation of MATLAB programming is presented in complete detail for four ELM-based models, including (1) classical ELM, (2) regularized ELM, (3) weighted regularized ELM, (4) outlier-robust ELM.