What exactly is learned in visual statistical learning? Insights from Bayesian modeling

Siegelman, N. 1, 2 , Bogaerts, L. 2 , Armstrong, B. 3, 4 & Frost, R. 1, 2, 4

1 Haskins Laboratories
2 The Hebrew University of Jerusalem
3 University of Toronto
4 BCBL, Basque center of Cognition, Brain and Language

It is well documented that humans can extract patterns from continuous input through Statistical Learning (SL) mechanisms. The exact computations underlying this ability, however, remain unclear. One outstanding controversy is whether learners extract global clusters from the continuous input, or whether they are tuned to local co-occurrences of pairs of elements. Discussions of this theoretical issue have reached a stalemate given conflicting results. Here we adopt a novel framework to address this question, applying a generative latent-mixture Bayesian model to data tracking SL as it unfolds online using a self-paced learning paradigm. This framework not only speaks to whether SL proceeds through computations of global patterns versus local co-occurrences at the group-level, but also reveals the extent to which specific individuals employ these computations. Our results provide evidence for substantial inter-individual mixture: while some participants rely on global patterns, others learn through the assimilation of local co-occurrences. This suggests that the two types of computations co-exist across different individuals, thereby explaining the previous mixed results that examined only group-level performance. We discuss the implications of these findings for understanding the nature of SL and individual-differences in this ability.