Self-controlled bio-inspired extreme learning machines for scalable regression and classification: a comprehensive analysis with some recommendations

摘要

The main goal of the current investigation is to answer some open questions regarding the applicability of bio-inspired techniques for simultaneous evolving and training of extreme learning machines (ELMs). Over the past decade, ELMs have successfully been applied to a wide range of regression and classification problems. However, in most of the reports, classical learning systems have been used for training ELMs which can result in cumbersome mathematical formulations and unstable solutions when handling large databases. Moreover, there is no guarantee for converging to global optimum solutions when traditional methods are used for the regularized learning of ELMs, especially when the variables in the database are correlated. To cope with such flaws, some researchers have used bio-inspired computation (BIC) for the topology evolving and parameter tuning of ELMs. However, the research stream still experiences its infancy. In an attempt to take a significant stride towards fulfilling the existing gap, the authors conduct a comprehensive analysis by adopting different BIC and classical methods for training ELMs. Based on the simulation results, some recommendations are given to guide researchers on using appropriate BICs for ELM training for small-scale, medium-scale, large-scale and very large-scale regression and classification problems.