This week’s Cog Lunch features Steve Piantadosi, presenting work with Noah Goodman and Josh Tenenbaum.
Title: A formal model of number word acquisition
Time: Tues 9/30, 12PM
Location: 46-3310
We present a formal computational model for learning number word meanings from noisy, cross-situational evidence. The model learns to construct number word meanings in a developmentally-plausible “language of thought” and shows how rational probabilistic inference can explain several key developmental patterns. The model also illustrates how infinite, structured, recursive concepts may be learned from relatively little data in a fairly unrestricted hypothesis space.
The model also provides a formal framework for studying conceptual change. It implements bootstrapping, where each successive numerical concept is defined in terms of the previous one (Carey & Sarnecka 2006, Carey 2004). However, we provide a new formalization of bootstrapping that avoids a key problem posed by Rips, Asmuth & Bloomfield (2006, 2008)—that bootstrapping presupposes knowledge of number and therefore cannot explain how children learn numerical concepts. The model also shows how linguistic and conceptual structures may be crucially related in number word learning.