MINIMAL NEURAL NETWORKS - DIFFERENTIATION OF CLASSIFICATION ENTROPY

Minimal neural networks (MNNs) differ from other neural networks in that they use localized processing. Expressions for the partial derivatives of the fuzzy entropy of classification have been obtained. The use of partial derivatives by minimal neural networks improves both efficiency and efficacy over the numerical computations. The MNNs described are FuRES-1 and FuRES-2 that build classification trees whose rules are composed of a neural unit and a neural layer, respectively. Training uses a directed simulated annealing approach that has robust characteristics. The stability of the algorithm with regard to initial conditions is evaluated for single unit and dual unit rules. FuRES-2 is compared to backpropagation neural networks using a mass spectral library search post-filter that identifies classes of saturated octanols and differentiates them from ethers.