Rockwood Memorial Lecture


Jeffrey L. Elman
Department of Cognitive Science, Kavli Institute for Brain and Mind, University of California, San Diego

11:00AM, Monday, April 21, 2005

Title: "Generalizing beyond our experience: Lessons from neural networks"

Over the past two decades, connectionist models of learning have provided impressive demonstrations of how much information is present in the environment. These results have been surprising to some, particularly in the domain of language, where it has been claimed that the input available to children is often insufficient to account for children's eventual knowledge (the so-called Poverty of the Stimulus problem.

In fact, considerably controversy remains regarding the extent to which experience provides a sufficient basis for language acquisition and linguistic generalization. Important questions have yet to be answered. Are there limits to statistically-based learning, and if so, what are they? Do language users literally record their experience in some numerical form? Recent empirical studies have established that the input available to young children is in fact massive, but it also occupies a very limited range of the total linguistic possibilities. How do we explain cases where generalization appears to go beyond experience if it is limited in this way?

These questions will be the focus of my talk. I shall describe several examples of simulations in which generalization appears to go beyond the input. The analysis of how generalization occurs in these cases suggests that the processes of learning and generalization in child language acquisition may be richer than initially supposed.

Host: Terry Sejnowski