Statistical learning in infancy: Transitional probabilities or perceptual primitives?

Johnson, S. 1 & Slone, L. 2

1 University of California, Los Angeles
2 Indiana University

Kirkham, Slemmer, and Johnson (2002) demonstrated visual statistical learning (VSL) in infancy by habituating infants to looming shape sequences, organized in pairs differing in transitional probabilities (TPs). Infants subsequently looked longer at identically structured vs. random sequences, implying that infants learned distributional input properties. However, Addyman and Mareschal (2013) found that 5-month-olds preferred random sequences without habituation, and that consecutive repetitions predicted looking behavior, taken as evidence that infants detected runs of pairs or triplets, if not TPs. We tested 8-month-olds in a VSL paradigm to better evaluate TP vs. repetition learning. Infants viewed 12 items (looming shapes) organized into four triplet units (N=38) or 9 items organized into three triplet units (N=42), followed by test items consisting of a familiar triplet alternating with a 'part-triplet' with lower mean TPs. We examined the part-triplet TPs and sequence repetitions during habituation for individual infants and found substantial variability both within and across conditions. Posthabituation preferences were correlated only with repetitions; TPs seemed not to support learning. Repetition thus may be a key source of information for segmentation in VSL tasks, perhaps as a 'perceptual primitive' that bootstraps learning from dedicated 'repetition detectors' (Endress, Nestor, & Mehler, 2009).