Visual statistical learning alters low-dimensional cortical architecture

Abstract

Our brains are in a near constant state of generating predictions, extracting regularities from seemingly random sensory inputs to support later cognition and behavior - a process called statistical learning (SL). Yet, the activity patterns across cortex and subcortex that support this form of associative learning remain unresolved. Here we use human fMRI and a visual SL task to investigate changes in neural activity patterns as participants implicitly learn visual associations from a sequence. By projecting functional connectivity patterns onto a low-dimensional manifold, we reveal that learning is selectively supported by changes along a single neural dimension spanning visual-parietal and perirhinal cortex (PRC). During learning, visual cortex expanded along this dimension, segregating from other networks, while dorsal attention network (DAN) regions contracted, integrating with higher-order transmodal cortex. When we later violated the learned associations, PRC and entorhinal cortex, which initially showed no evidence of learning-related effects, now contracted along this dimension, integrating with the default mode and DAN, while decreasing covariance with visual cortex. Whereas previous studies have linked SL to either broad cortical or medial temporal lobe changes, our findings suggest an integrative view, whereby cortical regions reorganize during association formation, while medial temporal lobe regions respond to their violation.