A simple and effective method for picking training samples in neural networks

Abstract

A simple and effective method for picking training samples in neural networks is proposed. The synchronous fluorescence spectra of 85 standards of Azorubin and New Red mixed with concentrations ranging from 5 μg ml⁻¹ to 20 μg ml⁻¹ were obtained by synchronous scanning the excitation and the emission monochromators maintained at an offset of 70 nm. The radial basis function neural networks (RBFNN) were used. The whole analytical properties domain was divided into nine small areas. A sample was placed into every small area. Numbers and distribution of the training samples were decided according to the accuracies of the samples placed. The method was completed in three steps in this work. Finally, the completed RBFNN was fully tested and the results were satisfactory with the root mean square error of 0.4745 and the total mean relative error (MRE) of 0.0338. The testing results show that the method proposed is simple and effective.