Feature expansion by a continuous restricted Boltzmann machine for near-infrared spectrometric calibration

Abstract

A modified algorithm for training a restricted Boltzmann machine (RBM) has been devised and demonstrated for improving the results for partial least squares (PLS) calibration of wheat and meat by near-infrared (NIR) spectroscopy. In all cases, the PLS calibrations improved by using the abstract features generated from the RBM so long as the nonlinear mapping increased the dimensionality. The evaluations were validated using bootstrapped Latin partitions (BLPs) with 5 bootstraps and 3-Latin partitions which proved useful because of the statistical learning and random initial conditions of the RBM networks. By using a noise decay parameter, initial large amounts of noise could be used and the benefits of simulated annealing achieved as the noise level is slowly decreased. This paper demonstrates for the first time that using abstract features and enlarging the spectral data can improve the calibration results and exemplifies the Copiosity Principle. Two NIR reference datasets were evaluated. The first set of wheat spectra was calibrated for protein concentration and the second set of meat spectra was calibrated for moisture, fat, and protein concentration. The RBM feature extraction improved the linearity of the models and reduced embedded noise. The RBM also can help eliminate some difficult spectral preprocessing stages such as variable alignment and feature selection. RBMs benefit from derivative preprocessing of the NIR spectra or other preprocessing that enhances the differences among the spectra.