ADAPTIVE RESONANCE THEORY-BASED ARTIFICIAL NEURAL NETWORKS FOR TREATMENT OF OPEN-CATEGORY PROBLEMS IN CHEMICAL-PATTERN RECOGNITION - APPLICATION TO UV-VIS AND IR SPECTROSCOPY

A supervised classification problem is defined as open-categorical if the training classes are only partly known beforehand according to their number, shape and size. This is the case, for example, in applications of spectroscopic pattern recognition for automated materials sorting for recycling. Continuously modified or newly developed materials require a continuous self-adapting pattern recognition technique (plasticity). A second, but contrary demand is the robustness of the classifier against outlier pattern vectors (stability). This stability-plasticity dilemma can partly be solved using Grossberg's adaptive resonance theory based artificial neural networks (ARTs). The main difference between ARTs like artificial neural networks to other types of artificial neural networks is that besides the numerical size of the weights as fitting parameters, additionally the network structure itself (number of units, dimensions of the weights matrices) is also not fixed forming a final result of the training process. Basic properties of the ART-1 technique for its potential application to chemical pattern recognition are elucidated by a classification of simulated data, UV-Vis spectra of phenanthroline complexes and infrared reflectance spectra of optical glasses.