Relationships between item and category learning: Evidence that abstraction is not automatic.

Abstract

Recent studies of concept formation using ill-defined categories suggest that abstraction of category-level information is more or less an automatic consequence of experience with exemplars. An experiment using ill-defined categories composed of photographs of women was designed to test this assumption. Experimental manipulations required subjects to learn either identification (exemplar level) and/ or classification (category level) responses to the photographs. Exemplar learning was generally found to proceed faster and with fewer errors than did category learning, and depending on the specific procedural manipulation, abstraction ranged from good to poor. Interactive-and independent-cue models were applied to the transfer data. The fits of the models indicated that when abstraction was good (as defined by transfer performance), the interactive-cue model best characterized the results; as degree of abstraction declined, the independent-cue model performed best. The results were discussed in terms of the relative contributions of categorylevel and exemplar-specific (idiosyncratic) information during classification learning.A considerable amount of recent research has been directed at the question of how people learn about and use ill-defined or "fuzzy" concepts and categories (see Mervis & Rosch, 1981; Smith & Medin, 1981, for reviews). A common theme running through virtually all these studies is the assumption, either implicit or explicit, that abstraction is more or less an automatic consequence of experience with exemplars of a category. It is well known in the literature on problem solving that transfer between two problems having the same underlying structure is often far from perfect (e.g., Reed, Ernst, & Banerji, 1974), and one might expect similar limitations in the learning of ill-defined concepts. Nonetheless, research on learning fuzzy concepts has focused on the what of abstraction, not the whether. Candidates for the what of abstraction include a prototype representing the central tendency of a category, simple or conjoint feature frequency, complex hypotheses or descriptions, best examples, or even the exemplars themselves (e.g.,

Prototype and Exemplar-Based Information in Natural Language Categories

“…To our best knowledge, this is the first study in which predictions of the prototype view and of the exemplar view are compared to predict category-based decisions for natural language concepts. Our results differ somewhat from the findings in the majority of papers that have investigated category learning, where it is assumed that all previously encountered exemplars of a category are stored and are activated whenever category-related decisions have to be made (Gluck & Bower, 1988;Medin, Altom, & Murphy, 1984;Medin, Dewey, & Murphy, 1983;Medin & Schaffer, 1978;Nosofsky, 1987Nosofsky, , 1988Nosofsky, , 1991. The data presented here suggest that activation of a limited number of frequently generated exemplars of the studied concepts suffices to predict typicality ratings, response times, exemplar-generation frequencies, and category-naming frequencies significantly and that elaborating the exemplar set can even decreases the predictive power.…”

SUSTAIN: A Network Model of Category Learning.

“…For example, Medin et al (1983) found that people are faster to associate unique names to photographs of nine female faces than they are to categorize the photographs into two categories. The logical structure of the two categories is shown in Table 3 (the logical structure of the categories is roughly equivalent to Shepard et al's Type IV problem).…”

“…Learning problems requiring item memorization should be more difficult than learning problems that promote abstraction and many models of categorization are constrained to predict that categorization will be, at worst, no harder than identification. Unlike Shepard et al (1961), Medin et al (1983) used distinctive stimuli (photographs of faces) and found that identification learning was actually more efficient than classification learning. As we shall we, SUSTAIN offers an explanation for this counter-intuitive finding.…”

language impairment and colour categories

“…ARTMAP predicts that critical feature patterns to which humans learn to pay attention are stored in memory. Under language/cultural supervision, these prototypes can be either specific ("exemplars"; Estes 1994;Medin & Smith, 1981;Medin et al 1983) or general ("prototypes"; Posner & Keele 1970;Smith & Minda 1998;Smith et al 1997). Typically, both specific and general information will be learned ("rule-plus-exceptions" ;Nosofsky 1984;Nosofsky et al 1992;Palmeri et al 1994).…”