OPTIMAL MINIMAL NEURAL INTERPRETATION OF SPECTRA

A new method ''optimal minimal neural-network interpretation of spectra'' (OMNIS) based on principal component analysis and artificial neural networks is presented. OMNIS is useful whenever spectra are measured for the purpose of classification or quantitative determination. The spectra can be visible light, near-infrared (NIR) light, sound, or any other large amount of correlated data. OMNIS is unique in several respects: It employs principal component analysis as a preprocessor to a neural network. The neural network contains direct connections from input to output ensuring that OMNIS is a true generalization of PCR (principal component regression). The neural network size is optimized so that the resulting solution contains the minimum of connections necessary to interpret the data. Cross validation is used systematically to optimize the network. OMNIS is based on recent insights in neural network research showing that deliberate search for the minimum network compatible with the data is a unique way of obtaining the optimal generalization ability. As a result OMNIS gives the best cross validation. In comparison, with PCR and PLS (partial least squares) on two NIR calibration data sets, OMNIS is demonstrated to reduce the standard error of prediction by 50 % to 75 %.