Connectionism and the learning of probabilistic concepts

Three experiments examine the claim of Gluck and Bower (1986, 1988a, 1988b) that the learning of medical concepts can be simulated by a connectionist network in which the symptoms are the input and the diagnosis is the output. The first experiment replicates the main finding of Gluck and Bower. In this experiment, subjects were required to estimate the probability of each of two diseases, given a particular target symptom. In fact these two probabilities were identical, but because one illness was more common than the other, the target symptom was a better predictor of the rare disease than of the common disease. Contrary to a normative probability judgement account, subjects were biased in that they judged the probability of the rare disease given the target symptom to be greater than the probability of the common disease given the target symptom. Gluck and Bower argued that such a result was predicted by a connectionist network using the Rescorla-Wagner learning rule, but it is argued that Gluck and Bower's network simulation was not appropriate for the experiment they had performed. In fact, it appeared that the connectionist network failed to predict the bias in the subjects' probability estimates.

However, this conclusion rests on an assumption that Gluck and Bower implicitly made. They arranged for P(rare disease/target symptom) and P(com-mon disease/target symptom) to be identical across all trials on which the target symptom occurred, both on its own and with other symptoms present. Gluck and Bower assumed that the subjects were estimating these probabilities. But the results of the second experiment showed instead that the subjects were estimating the probability of each disease given only the target symptom. In the final experiment the design was changed so that this problem might be circumvented. In this experiment, again, the subjects were biased in their probability judgements exactly as the connectionist network predicted. Thus, finally, evidence was found which was compatible with the network model but not with a normative account, but this was true only if the network did not include a layer of hidden units.