As if I needed more reasons to hope for admittance to UCSD. Here's a well written piece by Dr. Jeff Elman titled Connectionist Models of cognitive development: where next?. It appeared in TRENDS in Cognitive Science in March, and is a review of important and recent literature surrounding connectionist (those using neural networks) models of cognitive development. There is an emphasis on those models which deal with language acquisition.

I'm always telling people that computational models are good for directing research in the wet sciences, but I don't have a lot of concrete examples. This paper is full of them.

In one particularly cool bit, Elman discusses research which is attempting to discover how we learn to differentiate spoken words. Contrary to the way it seems to a practiced native speaker, there are no clear breaks in spoken language. Some computational models have apparently converged on one way in which this discrimination could take place:

... word boundaries are locations where the conditional probability of the next sound, given what has preceded it, is low. A network (or child) that attempts to anticipate what it (she) will hear next will tend to do worse at the onsets of words, and better as more of a word is heard. Error maxima thus constitute likely word boundaries.

Post a comment