Learning and Applying Category Knowledge in Unsupervised Domains

Theattribute-listing experiments revealed patterns of unsupervised learning characterized by initial discovery of categories based on featural contrasts with previous default expectations, gradual learning of defaults within a new category, and an overall bias to report informative features of the instances.

Abstract

This chapter discusses that the attribute-listing experiments revealed patterns of unsupervised learning characterized by initial discovery of categories based on featural contrasts with previous default expectations, gradual learning of defaults within a new category, and an overall bias to report informative features of the instances. The task also reveals the distinct stages in learning categories in hierarchical domains. Norms about a prior source category transferred to new derivative categories when they were applicable. The attribute-listing task shows promise as a useful paradigm within which unsupervised learning is studied in complex, hierarchically organized domains, and in which detailed theories is tested and refined. The similarity experiments provided independent confirmation of the proposed attentional biases toward surprising features. The similarity-from-memory procedure appears promising for investigating factors determining perceived informativeness. Such tasks provide greater insights into comparative judgments themselves, by suggesting constraints regarding the representation of stimuli in such tasks. The memory studies provide further support for the model's attentional assumptions, and explicate their impact on encoding processes, organization of category and instance information in memory, and later retrieval of facts about instances. The chapter also attempts to sketch a general model of learning that describes the abstraction of observations into general concepts and then the utilization of these concepts to encode later instances. Since top-down and bottom-up components are closely intertwined in human learning, this approach has some advantages compared to strictly bottom-up or top-down approaches.