A Computational Account of the Development of the Generalization of Shape Information

Leonidas A. A. Doumas*, John E. Hummel

*Corresponding author for this work

Research output: Contribution to journalArticlepeer-review

Abstract

Abecassis, Sera, Yonas, and Schwade (2001) showed that young children represent shapes more metrically, and perhaps more holistically, than do older children and adults. How does a child transition from representing objects and events as undifferentiated wholes to representing them explicitly in terms of their attributes? According to RBC (Recognition-by-Components theory; Biederman, 1987), objects are represented as collections of categorical geometric parts ("geons") in particular categorical spatial relations. We propose that the transition from holistic to more categorical visual shape processing is a function of the development of geon-like representations via a process of progressive intersection discovery. We present an account of this transition in terms of DORA (Doumas, Hummel, & Sandhofer, 2008), a model of the discovery of relational concepts. We demonstrate that DORA can learn representations of single geons by comparing objects composed of multiple geons. In addition, as DORA is learning it follows the same performance trajectory as children, originally generalizing shape more metrically/holistically and eventually generalizing categorically.

Original languageEnglish
Pages (from-to)698-712
Number of pages15
JournalCognitive Science
Volume34
Issue number4
DOIs
Publication statusPublished - May 2010

Keywords

  • Shape bias
  • Learning geons
  • Relation learning
  • Development
  • Computational modeling
  • DISTRIBUTED REPRESENTATIONS
  • RECOGNITION
  • DIMENSIONS
  • CORTEX

Cite this