High coherence among training exemplars promotes broad generalization of face families

How to best tailor learning experiences to promote conceptual knowledge that generalizes to new situations is a fundamental question across many domains. There have been conflicting findings on whether exposing a learner to highly variable exemplars during training improves generalization. In the present study, we used a face family category structure and manipulated the variability within the training set in terms of both the typicality (high vs. low training set coherence) and the number of unique examples (small vs. large training set size). Unlike some of prior work that trained participants to a performance criterion, the length of training was controlled across conditions. We found a clear benefit of high coherence training both in terms of category learning and broad generalization. Further, the category structure was designed such that participants could learn by picking up on shared features across category members and/or by memorizing individual category members and their unique features. The pattern of results from both categorization and an old/new recognition test indicated that high coherence training promoted the abstraction of shared features, whereas low coherence training resulted in difficulty both extracting commonalities among exemplars and memorizing unique exemplars. In contrast to the robust effects of training set coherence, there were minimal effects of training set size on generalization. Taken together, we show that training on typical examples can promote broad category knowledge by helping learners extract the central tendency of the category.