Visual statistical learning alters low-dimensional cortical architecture

Abstract

Our ability to implicitly extract environmental regularities over time and use this information to make predictions about the world is central to human cognition and behavior. While prior work has established the role of individual brain regions in this statistical learning (SL) process, it remains poorly understood how these regions coordinate their activity with the rest of the brain to guide SL. Here, using functional MRI, we measured human brain activity during visual SL, whereby individuals implicitly learned to associate pairs of images that were covertly embedded in a larger sequence of images. By projecting patterns of cortical and subcortical functional connectivity onto a low-dimensional manifold space, we found that SL was characterized by the expansion of several cortical areas along the whole-brain manifold. This included regions in higher-order visual cortex and the dorsal attention (DAN) and default-mode networks (DMN), and reflected their increased segregation from other brain networks. Consistent with these neural changes being associated with SL, we found that once the paired image structure was interrupted, these same brain areas immediately contracted back towards their pre-learning positions along the manifold. Notably, during this phase we also found that several new areas, belonging to the DAN and DMN, exhibited significant manifold contraction, with a subset of regions actually contracting beyond their baseline, pre-learning manifold locations. This suggests a putative mechanism in higher-order cortex by which learned associations can be represented over longer time horizons. Together, these results offer a new perspective on the distributed changes in whole-brain activity during visual SL.