Coglunch - February 23, 2006

"Imitation, agency and language learning"
Michael Ramscar
Stanford University The topic of this talk is language learning. But some of the things I'll talk about may not appear to have much to do with language at all. I'm going to try and show how a simple, unsupervised learning model can capture the patterns of children's acquisition of the English plural system. The model makes some surprising predictions about how exemplar-based learning might proceed in children's memories, and I'll try and show how these predictions are borne out. A lot of the phenomena predicted by the unsupervised learning model may appear odd to those of you familiar with adult memory and learning. I'm hoping you will bear with me on this. It turns out that an interesting thing about plural learning is that it is something kids do well, and adults do poorly, and so I think that having a different model of learning for kids may be justified. In the second part of my talk, I'll try and show how the unsupervised learning model can be very easily modified so that it can capture adult patterns of data. A curious thing about this process is that, in doing so, one ends up re-inventing a famous model of executive processing from computational cognitive neuroscience which some of you may be familiar with. I think this may be no bad thing, and I'm going to talk about some recent findings from developmental neuroscience to try and persuade you of this. Time permitting, I'll talk about some convergent data from children in an entirely different domain that support the overall architecture I propose, and a potential patient model suggested by recent studies of autism.

Last modified: Fri Feb 10 11:17:44 PST 2006 by emma@csli.stanford.edu