home · mobile · calendar · colloquia · 2003-2004 · 

Colloquium - Colunga

Statistical Bootstrapping: Creating Abstract Ideas from Specific Experiences
Department of Psychology, University of Colorado

At the heart of human cognition is the ability to form abstract ideas. But how do we do it? In this talk I will concentrate on the developmental origin of the ideas of Object and Substance as evidenced by the way children generalize novel nouns. Children generalize names for solid objects and for non-solid substances differently -- names for solid objects are extended by shape; names for non-solid substances are extended by material. Where does this knowledge come from?

I propose that the origin of this knowledge is in the statistics of the lexicon, perceptual properties and syntax. In a series of experiments with 2-year-olds and neural network simulations I will show 1) that a simple statistical learner, given the lexical and perceptual regularities available to children, will make an object-substance distinction like children's, 2) that networks' and children's expectations about the meanings of novel nouns are determined by the words they already know and the perceptual features of the object being named, and 3) that the knowledge of networks and children about object/substance is both abstract and rule-like on the one hand, and flexible and graded, finely tuned to the specifics of the child's environment on the other hand. Finally I will discuss the implications of this for the nature of knowledge, learning, and the learner.

Hosted by Michael Mozer.

Department of Computer Science
University of Colorado Boulder
Boulder, CO 80309-0430 USA
May 5, 2012 (14:13)