CENTER FOR RESEARCH IN
COGNITION & NEUROSCIENCES/ConnexConnex

Humans readily develop sensitivity to novel statistical relationships among previously unrelated elements in domains including vision, haptics, and language. A key implication in the domain of language, where infants, children and adults quickly develop sensitivity to a variety of word-like patterns embedded in novel phonological sequences, is the possibility that statistical learning provides a bootstrap mechanism that extracts coherent patterns as candidate forms for word learning. In this talk, I revisit some classic controversies in this domain. These include whether statistical learning is viable or useful for initial segmentation of child-directed speech, reported limitations of particular candidate computational mechanisms (such as Simple Recurrent Networks) for modeling human statistical learning, the nature of representations that emerge in models, and the implications of assuming language learning depends upon discrete representations of words.