Non‐linear Trainable Classifier in IRd Available to Purchase

Describes a new kind of non‐linear trainable classifier, successfully tested in computer‐vision pattern recognition. Class regions are not described, as usually, through analytical means but as a reunion of standard sets. Defines the notion of E‐separability for the class regions in the feature space IRd considered as a metric space with a distance related to the Euclidean distance. Studies and proves the convergence of the decision regions to the class regions in this metric space. For a given E (is a member of) provides a stopping rule for the training phase. Then describes the working phase, showing how classification actually takes place. Finally, presents significant results.