Technical Report #187

Learning to talk about the properties of objects: A network model of the development of dimensions

Linda B. Smith, Michael Gasser, Catherine Sandhofer

Abstract

~.... the undeniable fact being that anv number of sensorv sources. falline simultaneouslv on a mind WHICH HAS NOT EXPERIENCED THEM SEPARATELY, will fuse into a sinvle undivided obiect for that mind. William James, 1890, pg. 488, italics and capitalization as in the original).

A classic question in philosophy and psychology asks whether the object or its properties are more fundamental (see, e.g., Carnap, 1967; Locke, 1964; James, 1890; Boring, 1942). Clearly we perce*e both --- a dog is apprehended as an integral whole and as being big and brown and furry. But which perception is prior? This question is often interpreted in terms of the logical priority of parts and whole. And the contemporary concensus is that parts are logically and computationally prior; complex percepts and concepts are built from simpler primitives. Over a hundred years ago, however, William James concluded that whole objects are experientially prior, that constituent properties are a secondary product of perceptual learning. This chapter provides support for James' conclusion.

Our starting point is the protracted course of children's acquisition of dimensional language, a lengthy process that includes the acquisition of dimensional terms and the development of selective attention. We explain this developmental course by simulating it in a connectionist network. Our results suggest that dimensions are created, that they are the product of learning dimensional language.