Paying Attention to Relevant Dimensions: a Localist Approach

Localist models of, for example, the classification of multidimensional stimuli, can run into problems if generalization is attempted when many of the stimulus dimensions are irrelevant to the classification task in hand. A procedure is suggested by which a localist model can learn prototype representations that focus on the relevant dimensions only. These permit good generalization which would be lacking in a simple exemplar-based model.