The construction and approximation of neural networks operators with Gaussian activation function

Zhixiang Chen; Feilong Cao

Vol. 18 No. 1 (2013)

Regular Articles

The construction and approximation of neural networks operators with Gaussian activation function

Zhixiang Chen
Feilong Cao

Submission of an original manuscript to the Mathematical Communications will be taken to mean that it represents original work not previously published, that it is not being considered elsewhere for publication; and, if accepted for publication, it will be published in print and online form and it will not be published elsewhere in the same form, for commercial purposes, in any language, without the consent of the publisher. The journal takes the stance that the publication of scholarly research is meant to disseminate knowledge and, in a not-for-profit regime, benefits neither publisher nor author financially. It sees itself as having an obligation to its authors and to society to make content available online now that the technology allows for such a possibility.

Abstract

This paper studies the construction and approximation of neural network operators with a centered bell-shaped Gaussian activation function. Using a univariate Gaussian function a class of Cardaliaguet-Euvrard type network operators is constructed to approximate the continuous function, and the Jackson type theorems of the approximation and some discussions about the convergence are given. Furthermore, to approximate the multivariate function, a class of bivariate Cardaliaguet-Euvrard type network operators is
introduced, and the corresponding estimates of the approximation rate are deduced.

PDF