How do learners make sense of their intricately structured visual and auditory environments? One important learning mechanism available to both infants and adults is the ability to detect statistical structure within visual and auditory inputs. Despite the scope of this “statistical learning” (SL) ability and the large literature that now surrounds it, the processes underlying SL, and how those processes may differ across modalities and change across development, remain unclear.

The goal of this dissertation was to clarify these mechanistic questions. Paper 1 reviews previous work on visual and auditory SL across development, highlighting the debate concerning the domain-generality of SL. Paper 2 investigates the domain-generality of infants’ SL by employing a visual SL task comparable to previous auditory SL tasks. Findings from four experiments with 8-month-olds suggest that visual SL is constrained compared to auditory SL in infancy. Individual differences in visual SL performances were unrelated to visual short-term memory performances, but were related to overall cognitive ability. These findings suggest that visual SL ability is important for, and constrained by, early cognitive development.

Papers 3 and 4 examine the representations that result from visual SL. I evaluate the ability of two major classes of SL models – statistical models and chunking models - to correctly predict adults’ and infants’ performances on visual SL tasks. Across five experiments, adults and 8-month-old infants discriminated between high- and low-probability visual sequences, providing strong evidence of SL. Critically, adults represented pairs of items embedded within larger sequences, but did not represent so-called “illusory sequences.” These results support the competitive chunking model of Servan-Schreiber and Anderson (1990), and suggest that adults represented visual sequences in terms of hierarchical levels of chunks. In contrast, 8-month-olds did not show evidence of representing embedded or illusory sequences, suggesting that infants represented visual sequences in terms of only the highest level of chunks. Together, these studies not only suggest that the representations that result from visual SL are best captured by chunking models, but also that the type of chunking learners engage in may change between infancy and adulthood.