CATEGORIZATION BY A CONNECTIONIST NETWORK

Three experiments tested a simple connectionist network approach to human categorization. The specific network considered consists of a layer of input nodes, each representing a feature of the exemplar to be categorized, connected in parallel to a layer of output nodes representing the categories. Learning to categorize exemplars consists of adjusting the weights in the network so as to increase the probability of making correct categorizations; weight changes are determined by the Rescorla-Wagner (1972) learning rule. The experiments used a simulated medical diagnosis procedure in which subjects have to decide which disease (the category) each of a series of patients is suffering from on the basis of the patients' symptoms (the features). After a series of trials, the subjects rated the extent to which particular symptoms were associated with particular diseases. In each of the experiments, it is shown that a process of selective learning occurs in this categorization task and that selection in turn depends on the relative predictiveness of the symptom for the disease. Such effects parallel results found in animal conditioning experiments and are readily reproduced by the connectionist network model. The results are also discussed in terms of a variety of traditional theories of categorization.