How Goals Shape Category Acquisition: The Role of Contrasting Categories

摘要: How Goals Shape Category Acquisition: The Role of Contrasting Categories Tyler Davis (thdavis@mail.utexas.edu) Bradley C. Love (brad_love@mail.utexas.edu) Department Psychology, University Texas at Austin Austin, TX 78712 USA Abstract category ideal (Barsalou, 1985). For example, the goal derived foods to eat on a diet will have graded structure near 0 calories (e.g., celery). Ideals also been shown influence taxonomic categories when groups or other cultures that goals associated with are tested Atran, 1999). tree experts view trees as more typical extent they minimize weediness (Lynch, Coley, & Medin, 2000). majority explanations for how this idealization occurs suggest these effects depend upon real-world and cultural influences not present in artificial commonly used laboratory. Thus, findings often depicted being odds approaches categorization rely laboratory techniques such mathematical modeling. It is our perspective, however, methods continuous one another, can be mutually informative. Further, we’ll argue lower level learning mechanisms described by e.g., error-driven clustering models, provide least partial explanation affect structure, whilst making predictions may demonstrated organized around ideals rather than central tendencies. In real- world categories, items extreme perceived most typical. Here, we demonstrate similar learned Our experiment simulations low-level seek prediction error responsible aspects idealization. To error, centroids adjusted both increase similarity their members, well members contrasting categories. This dynamic distorts representations away from leading Keywords: learning, use, goals. Introduction Increasingly, research suggests learner’s within an environment place substantial constraints content (for review see Markman Ross, 2003). classification lead important differences between those inference, even though two tasks formally equivalent. These critical challenge formal many models do contain ways which exert any over (cf. Love, 2005). At same time, because challenge, offer opportunity cross- fertilization fields studying expertise/cultural psychology modeling/laboratory experimentation. following, discuss related studied using design novel model previous results cultural/expertise literatures. One finding cultural, expertise, goal-derived literatures determined ideals. Graded refers notion differ terms belong. Whereas category’s statistical tendency (Rosch, 1975), centered A Formal Model Clustering different perspective defined illustrate, food example above, would separate clusters representing non-diet initially mean number respective However, since distributions continuous, somewhat overlapping, sometimes highly activate cluster error. Error-driven compensate moving means calorie dimension further each other. After some trials, tend move apart actual means, producing marked shift outset produce strongest activation, end (i.e., ideal) will.