Prototype abstraction and the rejection of extraneous patterns

The abstraction of prototypical information from ill-defined categories was investigated for categories defined by three and six instances. The major question of interest was whether manipulation of category size would facilitate prototype abstraction when only two categories were learned and a default strategy of classification could not be used. To maximize learning difficulty, categories were defined by high-level (7.7-bit) distortions, and transfer was to new patterns at a high or extreme (8.3-bit) level of distortion. Nearly half the transfer patterns belonged to neither of the learned categories, and subjects were allowed to classify transfer patterns into a “junk” category. All subjects were tested immediately and after a delay of 1 week. The results indicated that category size maintained its potent influence on generalization performance. Although both categories deteriorated across the delay, the category defined by six instances continued to allow new patterns to be classified more accurately.