Auto-adjustable method for Gaussian width optimization on RBF neural network. Application to face authentication on a mono-chip system
Abstract
This paper describes an automatic method for optimizing a radial basis function (RBF) neural network parameter during training stage. The neural network is used as a classifier to realize a human face authentication system. The aim of this project is to obtain a low cost system on chip (SoC) to replace password identification on mobile devices. The system is designed in order to respect the AAA methodology for algorithm selection and implantation on hardware platform. Several classifier parameters need to be adjusted to obtain better performances. One of these critical parameters is the Gaussian width. Contrary to other applications, there is a unique class (corresponding to the trained person); standard methods for width optimization therefore do not work. We suggest a new method to compute an optimized width.