Abstract for Martin Cooperson

Pattern-recognition is a critical aspect of machine cognition. A popular framework, instance based classifiers, is built around the conjecture that similar objects are likely to represent the same class. For this reason, a large number of preclassified examples are examined to find those that are most similar to the object at hand. This can entail prohibitive computational costs during the classification phase. Not surprisingly, a lot of research has gone into how to reduce the size of this store. The solution posed by this work is to find three or more groups of preclassified examples in such a manner as to maximize the likelihood that each group represents a somewhat different aspect of the pattern to be recognized. Systematic experiments show that very high classification accuracy can then be achieved with very small numbers of stored examples.