A guide for the upper bound on the number of continuous-valued hidden nodes of a feed-forward network

Abstract

This study proposes and validates a construction concept for the realization of a real-valued single-hidden layer feed-forward neural network (SLFN) with continuous-valued hidden nodes for arbitrary mapping problems. The proposed construction concept says that for a specific application problem, the upper bound on the number of used hidden nodes depends on the characteristic of adopted SLFN and the observed properties of collected data samples. A positive validation result is obtained from the experiment of applying the construction concept to the m-bit parity problem learned by constructing two types of SLFN network solutions. © 2009 Springer Berlin Heidelberg.