The Role of Base Rates in Category Learning

Previous researchers have discovered perplexing inconsistencies in how human subjects appear to utilize knowledge of category base rates when making category judgments. In particular, Medin and Edelson (1988) found an “inverse base rate effect” in which subjects tended to select a rare category when tested with a combination of conflicting cues, and Gluck and Bower (1988) reported apparent “base rate neglect” in which subjects tended to select a rare category when tested with a single symptom whose objective diagnosticity was equal for all categories. In this article I suggest that two principles underlie those effects: First, base rate information is learned and consistently deployed during all training and testing cases. Second, the dominant effect of base rates is to cause frequent categories to be learned before rare categories, so that the common categories are encoded in terms of their typical features, and the rare categories are encoded by whichever features distinguish them relative to the already-learned, common categories. Four new experiments are reported which confirm those principles. Experiment 1 replicates and extends the basic inverse base rate effect and establishes a standard for subsequent experiments and modeling. Experiment 2 shows that pre-training on a subset of categories has comparable effects as giving those categories high base rates during training. Experiment 3 shows that apparent “base rate neglect” is simply an attenuated case of the “inverse base rate effect,” hence any model of one effect should account for the other. Experiment 4 uses probabilistic categories, and shows that apparent “base rate neglect” can be obtained even for cues that occur in less than 50% of the rare category’s exemplars, and that cues that are not associated with any category decrease the base-rate bias. The principles are formalized in a new connectionist model that modifies attention to stimulus features based on the network’s current knowledge. Quantitative fits to the empirical data are reported, and provide additional support for the principles.