ISSN 0439-755X
CN 11-1911/B

›› 2009, Vol. 41 ›› Issue (01): 44-52.

Prototype and Exemplar on Classification and Inference Learning

Liu Zhi-Ya;Mo Lei

1. Center for Studies of Psychological Application, South China Normal University, Guangzhou 510631, China
• Received:1900-01-01 Revised:1900-01-01 Published:2009-01-30 Online:2009-01-30
• Contact: Mo Lei

Abstract: In this paper, we compare classification learning with another mean of learning categories: inference learning. In inference task, participants predict the value of a missing feature of an item given its category label and other feature values. In the classification task, participants predict the category label of an item given its feature values. Yamauchi and Markman (1998, 2000, 2002) showed that these two types of learning do not result in the learning of equivalent knowledge. Categories defined by a family resemblance structure were more easily learned by inference learning than by classification learning, whereas categories defined by a nonlinearly separable structure were more easily learned by classification learning than by inference learning. Chin-Parker and Ross (2002, 2004) found that Classification learners were highly sensitive to diagnostic features but not sensitive to nondiagnostic, but prototypical, features. Inference learners were less sensitive to the diagnostic features than were classification learners and were also sensitive to the nondiagnostic, prototypical, features. The “5-4”category structure from Medin and Schaffer (1978) has played an important role in the recent dominance of exemplar/instance-based category representation over prototype/central-tendency based representation. Yamauchi and Markman (1998) argued that the type of category representation subjects use, prototypes or exemplars, depends on how they learn the categories, but they did not strongly support this claim with formal models of the representations. This study contrasts the learning results of training subjects on the 5-4 category structure using either standard classification or feature inference. The research further systematically explores the two learning tasks that might lead to differential learning efficiency, strategy and outcome.
Chin-Parker & Ross (2002，2004) pointed out Yamauchi & Markman’s two precursory studies that 1998’s ignored irrelative transfer effects and 2002’s was less learning exemplars. We improved Yamauchi & Markman 2002’s experiment by designing a “student coming into social group” , using 5-4 category structure, learning-transfer task paradigm and feature category detecting method, and explored leaning results in different way at classification and inference learning. One hundred and forty-four volunteers who participated for partial credit at an introductory psychology course at South China University of Technology took part in the experiments (ninety-six for other two complementary experiments).
For participants given classification learning the study found that they classified/inferenced transfer stimuli A2 more accurately than they did the transfer stimuli A1. In contrast, participants given inference classified/inferenced transfer stimuli A1 more accurately than they did the transfer stimuli A2. The study also showed that inference learning was better than classification learning to catch the category prototype. At classify transfer phase, inference learners were more accurately than classification learner classified the two category prototypical stimulus. At inference transfer phase, inference learners were more likely to infer the absent feature as the prototypical feature than classification learner did. At single feature classifying, inference learners showed more accurately than classification learner did. There are more participants to reach the 89% learning criterion in the classification learning than in the inference learning. However, for the participants reached 89% learning criterion, study also found that those in the classification learning condition (M = 13.17) were not required significantly more blocks than those in the inference learning condition (M = 14.45).
The results are consistent with the hypothesis that inference learning induces prototype representation and classification learning induces exemplar representation. Classification learning was better than inference learning at prototype integrating. Classification learner was more quickly to use the single dimension strategy than inference learner did, but at high level strategy, inference learner kept same pace with classification learner and showed potential learning effective

CLC Number: