UNSUPERVISED LEARNING MODES FOR THE WISARD WEIGHTLESS NEURAL NETWORK

Iuri Wickert

November/2001

Advisor: Felipe Maia Galvão França

Department: Engenharia de Sistemas e Computação

In this thesis we present the AUTOWISARD neural model, a new unsupervised learning algorithm for the well-known WISARD weightless neural model. AUTOWISARD is able to classify binary patterns in a single learning epoch, reaching a stable state. The model implements several class generation and learning control methods, as the learning window, the partial learning and the learning control function. A comparison was made with AUTOWISARD and Fuzzy ART, in an optical character recognition application, using labelled handwritten digits images. To compare such different neural paradigms and networks, the results were normalized and converted into a graphical representation which yields the visualization and the comparison of the average performance for both networks relatively to the quality of classification (quantity of classes which recognises multiple digits and saturation level of the classes). The results showed that the classifications generated by AUTOWISARD are consistently superior in quality to the Fuzzy ART's. This work also presents a revision of related models: WISARD, Fuzzy ART and WIS-ART, conclusions and future works, presenting the hierarchical AUTOWISARD model.