Proceedings of the Annual Meeting of the Cognitive Science Society

Humans can learn and reason about abstract concepts quickly,flexibly, and often from very little data. Here, we study howpeople learn novel concepts within a binary grid domain, andfind that even this minimal task nonetheless necessitates theinference of highly structured parts as well as their compo-sitional relationships. Furthermore, by changing the presen-tation condition of the learning examples, we reveal differentapproaches involved in learning such visual concepts: giventhe same images, human generalizations differ between rapidand static presentation conditions. We investigate this differ-ence by developing several computational models that vary intheir use of structured primitives and composition. We find thatlearning in the rapid presentation condition is best described asinference in simple models, while learning in the static presen-tation condition is best described as inference in a more struc-tured space of graphics programs.