Fuzzy Hypercube Artificial Neural Network Classifier

摘要

An Artificial Neural Network classifier, based on fuzzy Min-Max, Adaptive Resonant Theory (ART), and fuzzy ART is described. The outputs of the classifier are fuzzy hypercubes representing functional categories of its input functions. A hypothesis and test paradigm compares input data and existing hypercube categories, and results in either network resonance or dissonance, depending on the test outcome. A hypothesis is formed by two match functions: Degree of Inclusion, and Degree of Perfect Match. An overall hypothesis is chosen with the best Degree of Match. Tests are then performed to verify the hypothesis. The vigilance test measures the top down match between the hypothesized category and the input. The overall hypervolume test ensures that any category adjustments keep the total category hypercube volume within bounds. The fuzzy hypercube classifier was tested using two standard sets: Iris Flower and Wisconsin Diagnostic Breast Cancer. The network produced 88% and 76% correct classification, respectively. A speaker recognition system using a fuzzy hypercube classifier was also tested using the Switchboard and Greenflag databases. Test results are discussed.