Validating an unsupervised weightless perceptron
Iuri Wickert, Felipe M. G. França · 2004
The paper presents a comparison between two unsupervised neural network models: (i) the well-known fuzzy ART, and (ii) AUTOWISARD, a new unsupervised version of the classic WISARD weightless neural network model. It is shown that AUTOWISARD is simple, fast and stable, whilst keeping compatibility with the original WISARD architecture. Experimental test results over binary patterns benchmarks have shown that, although both unsupervised learning models are remarkably simple, AUTOWISARD consistently exhibits better classification skills than fuzzy ART. It is also shown that such superiority happens thanks to AU-TOWISARD's richer internal representation of the trained patterns and the training methods employed by the algorithm, such as the learning window and partial training strategies.