LMS rules and the inverse base-rate effect: Comment on Gluck and Bower (1988).

Abstract

(}luck and Bower (1988) suggestmi that through the use of the Rescorla-Wagner learning rule, a connectionist network might be ~ble to model the inverse base-rate phenomenon found by Medin and Edelson (1988). I prove that a network of the type that they proposed does not capture this effect. However, one can also prove that with additional assumptions about the encoding of features, the Rescorla-Wagner learning rule can be made to model the inverse baserate effect. The importance of these assumptions and an outline of how they might be tested are then discussed. Gluck and Bower (1988) recently showed how a version of the Rescorla-Wagner learning rule (a variant of the least mean squared [LMS] error correction rule) could be used to predict human behavior in a category-learning task. At the end of their article, they theorized that this rule might be able to account for subjects' incorrect use of base-rate information found in a set of experiments by Medin and Edelson (1988). The inverse base-rate phenomenon (Medin & Edelson, 1988) is a surprising and counterintuitive effect shown by subjects in category-learning experiments in which the frequency of presentation of the categories is varied. Table 1 illustrates a simple case: Subjects are presented ~h exemplars for each of two categories. In the simple case, these categories have one common feature (A) and one distinctive feature (B or C). When one category is presented more frequently than the other, an interesting pattern or responses results. If subjects • are given the common feature and asked which category is more likely to be described by this feature, subjects respond with the more frequently presented category. If either of the distinctive features is presented during testing, the appropriate category is selected. However, ifboth of the distinctive features are presented together, the subject is more likely to select the less frequently presented category, responding in opposition to the base-rate information. Medin and Edelson (1988) discussed the possibility that the inverse base-rate effect stems from a competition between features. In a context in which the categories are diseases and the features are symptoms, they stated that