Lew-Williams, C. 1 , Reuter, T. 1 & Borovsky, A. 2
1 Princeton University
2 Florida State University
Learners process statistical co-occurrences incrementally, and in doing so, learn to predict upcoming input. In models of error-based learning, predictions are generated and, if incorrect, revised, which helps learners navigate the probabilistic structure of the environment. Here, using three converging investigations (N=126), we show that the ability to update (but not simply generate) predictions enables infants and young children to better represent the shifting statistics of the input. In Experiment 1, 1-year-old infants' ability to update visual predictions was tightly coupled with vocabulary growth. In Experiment 2, 3-year-olds' ability to redirect attention after a semantic prediction error was highly correlated with vocabulary. But are these individual differences in prediction updating meaningfully related to the learning of novel input? In Experiment 3, 3-year-olds' success in learning novel object-label pairings was contingent on rapid visual reorientation toward novel objects after a semantic prediction error. Across experiments, children showed skill in generating predictions, but the data point to an essential and complementary behavior: the ability to use unexpected input to update representations about the statistics of the environment. Together, the experiments indicate that young children's subsecond prediction abilities are sufficiently flexible to learn from the statistical variability and inherent novelty of incoming input.