Extreme Learning Machine with Harmony Search for High Dimensional Data Classifications

Extreme learning machine (ELM) with harmony search (HS) method for solving high dimensional data pattern classifications (abbreviated to HDHS-ELM) is studied in this paper. Firstly, HDHS mechanism which can enhance the global search ability to make features in feature space more separable is used to adjust the hidden nodes' parameters of neural classifier to improve the classification performance and reduce the hidden neurons' quantity. For balancing and reducing empirical and structural risks, the output weights are then computed through regularization technique. Compared with classifiers trained with another three algorithms related to ELM, the simulation results of classifier trained with HDHS- ELM illustrate better performance.

I. Introduction

Neural networks (NNs), as an important class of machine learning algorithms, have been widely researched on its applications [1]-[4]. It has been proved that single hidden layer feedforward neural networks (SLFNs) have universal approximation capability [5]. Moreover, theories of ELM [6] further show that SLFNs with randomly generated input weights and hidden biases can also approximate any complex functions. Due to such an useful property, classifiers trained with ELM have been developed for pattern classification in recent years [6], [23].