[PS-3.11]Rapid Generalization from Statistical Learning

Yuan, L. & Smith, L.

Indiana University

Much human learning emerges by incrementing and statistically aggregating experiences (SL) given massive input. But in some domains human learners also show rapid generalization from quite limited data. Most SL research has focused on linguistic and visual domains. Here we examined SL in the case of learning how multi-digit numbers are symbolically represented to understand the data structures that learners use to extract latent rules from a relatively few experienced items and to broadly generalize that learning. Multi-digit numbers provide an excellent domain since reading and mapping number names to written numerals such as 1283 require applying general principles not the learning of specific instances. In the experiment, we provided 3- to 5-year-olds with a small set of name-number associations (three-hundred twenty-six -> 326) over 3 days of training and then tested them with novel names and written forms; they showed rapid learning and near perfect generalization. The manipulation of two training sets and a control set provide further insights into the latent structures in data sets that may support rapid learning and broad generalization from very few instances.