Correlated Cues in Probabilistic Categorization 1 Running head: CORRELATED CUES IN PROBABILISTIC CATEGORIZATION Better Learning With More Error: Probabilistic Feedback Increases Sensitivity to Correlated Cues in Categorization

Despite the fact that categories are often composed of correlated features, the evidence that people detect and use these correlations during intentional category learning has been overwhelmingly negative to date. Nonetheless, on other categorization tasks such as feature prediction, people show evidence of correlational sensitivity. A conventional explanation holds that category learning tasks promote rule use which discards the correlated-feature information; whereas other types of category learning tasks promote exemplar storage which preserves correlated-feature information. Contrary to that common belief, we report two experiments that demonstrate that using probabilistic feedback in an intentional categorization task leads to sensitivity to correlations among non-diagnostic cues. Deterministic feedback eliminates correlational sensitivity by focusing attention on relevant cues. Computational modeling reveals that exemplar storage coupled with selective attention is necessary to explain this effect.