Hierarchical and non-hierarchical network flows generate complementary representational dynamics in human visual cortex

Hierarchy is considered a fundamental organizing principle of visual cortex, but its functional implications remain debated given the presence of direct (non-hierarchical) connections. Building on recent advances in measuring direct region-to-region functional connectivity in the human brain, and in using that connectivity (rather than, e.g., visual classification training) to construct deep neural network models, we tested the hypothesis that hierarchical and direct connectivity pathways make distinct contributions to the generation of visual functionality. Detailed measurement of visual functionality, connectivity, and their interaction was achieved using 7T MRI and empirical neural network (ENN) models parameterized by empirical connectivity estimates. The classic V1 to V4 hierarchy was recovered in terms of (i) network distance from V1 along the human brains direct region-to-region resting-state functional connectome and (ii) on-task representational transformation distance (visual representation dissimilarity) from V1. In silico ENN lesion experiments revealed that hierarchical pathways (V1{leftrightarrow}V2{leftrightarrow}V3{leftrightarrow}V4) reduce the dimensionality of neural representations relative to more rapid and high-dimensional representational contributions from direct pathways (e.g., V1{leftrightarrow}V4). These findings reveal distinct but complementary roles of hierarchical and direct pathways in generating cortical functionality. Significance StatementHierarchy is a foundational organizing principle of cortex, yet its functional consequences remain unclear because of direct, non-hierarchical connections. The visual system, often portrayed as the clearest example of cortical hierarchy, provides a testbed for dissociating hierarchical and non-hierarchical contributions. Using high-resolution 7T MRI with recent advances in measuring direct region-to-region functional connectivity, we mapped the classic V1 to V4 hierarchy in the human brain. Using empirical neural network (ENN) models parameterized by these empirical connections, we determined that hierarchical pathways (V1{leftrightarrow}V2{leftrightarrow}V3{leftrightarrow}V4) reduce representational dimensionality relative to more rapid, high-dimensional contributions from direct pathways (e.g., V1{leftrightarrow}V4). These results reveal complementary hierarchical and direct contributions and establish ENN modeling as a general approach for determining pathway-specific functions throughout the brain.