Deals with the development of ARTMAP-like neural networks to analyze feature space for classification purposes. The proposed tool provides information about the value of membership functions of the unknown input vector to each class of interest. The designed ARTMAP-like system is called MF-ARTMAP based on the fact that membership functions are calculated. The functions shape is predefined as Gaussian with adaptation of mean value and variance in each feature space dimension during the training procedure. The parallel version of this approach is designed and implemented too. The parallel MF ARTMAP have some advantages over regular MF ARTMAP. The usefulness of this approach is presented on the benchmark classification problems e.g. circle in the square and spiral and on real-world data from satellite images over Slovakia. Classification accuracy is calculated using the contingency tables approach on actual and predicted classes of interest.
展开▼
