Michael Gasser and 
Eliana Colunga
Computer Science Department
Linguistics Department
Cognitive Science Program
Indiana University
Bloomington, IN 47405, USA
{gasser,
ecolunga}@cs.indiana.edu
In this report we argue that the study of the acquisition of word meaning requires taking seriously non-linguistic cognition, in particular human vision and the pre-linguistic development of concepts. We consider the implications of this claim for the acquisition of spatial relations, and we present Playpen, an evolving neural network architecture for modeling the development of spatial language and spatial cognition. Playpen includes modules for high-level vision, the lexicon, and the conceptual space in which vision and lexicon come together, allowing for the mutual influence of all three. This report focuses on the basic building blocks of the network. Feature binding and object segregation are implemented through the use of phase angles, and the learning algorithm is a version of Contrastive Hebbian Learning [Movellan, 1990], adapted for units with phase angles. We argue that to represent and learn the meanings of relational terms, the network also requires units which represent micro-relations explicitly. In Playpen these take the form of relation units, hard-wired clusters of simpler units which become activated to the extent that they receive inputs from units representing distinct objects.