Under Pressure: How Time-Limited Cognition Explains Statistical Learning by 8-Month Old Infants

Abstract

In a classic experiment, Saffran, Aslin, and Newport (1996) used a headturn preference procedure to show that infants can discriminate between familiar syllable sequences ("words”) and new syllable sequences ("non-words" and "part-words"). While several computational models have simulated aspects of their data and proposed that the learning of transitional probabilities could be mediated by neural-net or chunking mechanisms, none have simulated the absolute values of infants' listening times in the different experimental conditions. In this paper, we used CHREST, a model based on chunking, to simulate these listening times. The model simulated the fact that infants listened longer to novel words (non-words and part-words) than familiar words. While the times observed with the model were longer than those observed with infants, we make a novel finding with regard to phonological store trace decay. We also propose how to modify CHREST to produce data that fits closer to the human data.


Back to Table of Contents