ABSTRACT
In natural environments, the sensations arriving at two or more sensory modalities are often correlated. We derive an algorithm for a piecewise linear classifier which uses the relationship between patterns presented simultaneously to two or more networks as a supervisory signal. The algorithm is based on the idea of minimizing the disagreement error — the proportion of patterns disagreed upon — between two or more networks receiving correlated patterns. We test the algorithm on a ten class vowel classification problem and find that it performs better than a hybrid unsupervised/supervised algorithm and, with two iterations, almost as well as a related supervised algorithm (Kohonen’s LVQ2.1).
