A Recurrent Neural Circuit Theory of Normalization and Inter-area Communication.

Reciprocal feedback connections are ubiquitous in the brain, but their function remains largely unknown. To investigate the possible role(s) of feedback, we developed an analytically tractable, hierarchical, recurrent neural circuit model with feedback connections. The model implements divisive normalization exactly across a hierarchy of cortical areas, via recurrent amplification. Our analysis shows that: 1) increasing the feedback gain from a higher to lower area in the hierarchy amplifies responses in both areas, with a more pronounced increase in higher cortical areas, consistent with experimental observations; 2) feedforward and feedback processing are associated with distinct frequency bands; 3) inter-areal communication is lower dimensional than within-area communication; 4) increasing feedback strength enhances inter-areal communication without altering the dimensionality. In addition, the model suggests a possible mechanism for feature-based selective attention, that dynamically routes information between cortical areas (i.e., changing functional connectivity) simply by changing the gain in one or another higher cortical area. The model thus provides an analytically tractable framework for exploring normalization, inter-areal communication, and functional connectivity.