|
Statistical learning : historical, empirical, and computational perspectives| old_uid | 2311 |
|---|
| title | Statistical learning : historical, empirical, and computational perspectives |
|---|
| start_date | 2007/02/27 |
|---|
| schedule | 15h45 |
|---|
| online | no |
|---|
| summary | Since the mid-1990s, there has been a resurgence of interest in mechanisms of learning that could complement, and perhaps replace, innate mechanisms for language acquisition. The initial studies focused on word segmentation -- a low-level aspect of language learning that must be solved by infants before higher-level structures are acquired (e.g., syntax). These initial studies, of course, were not the first to study learning in infants, and many principles of learning had been documented in visual and auditory domains. The unique feature of statistical learning was the rapid presentation of stimulus materials, something not incorporated in previous studies of infant learning. What is now clear, after a decade of studies of statistical learning, is that it is both domain- and species-general. At issue is how such a powerful learning mechanism is constrained so that it extracts just the right structures in finite time without suffering from a computational explosion of irrelevant statistics. Many of these contraints will be described, along with a recent Bayesian computational model, that has the potential to account for the empirical facts about statistical learning. The challenge for the future is to extend these mechanisms to the far more important, and potentially language-specific, case of forming categories and rules. |
|---|
| responsibles | Zondervan |
|---|
| |
|