Over the past couple of decades, statistical learning—the discovery of patterned regularities by way of distributional properties in the input—has become a key explanatory construct in the cognitive sciences. However, the promise of a powerful domain-general learning mechanism underlying linguistic and other aspects of cognitive behavior is hampered by both methodological and theoretical shortcomings. In this talk, I will argue that statistical learning is not supported by a unitary mechanism but instead involves separate neural networks working in different modalities, each relying on a set of domain-general computational principles. This provides a possible explanation for why some studies have observed similar sensitivity to statistical patterns across different domains, while others have revealed substantial modality effects on statistical learning. To further elucidate the intricate relationships between statistical learning and other aspects of cognition, cognitive scientists need to adopt an individual differences approach that compare sensitivity to statistical patterns across different domains. I conclude by highlighting the importance of statistically-based chunking as a key component of language and cognition, linking statistical learning more closely to basic work on learning and memory.

Morten H. Christiansen, Professor, Aarhus University, Cornell University, and Haskins Laboratories