This paper introduces spatially embedded recurrent neural networks (seRNNs), a novel model that incorporates biophysical constraints into a fully artificial system. seRNNs learn basic inferences while existing within a 3D space, with communication constrained by a sparse connectome. The model converges on structural and functional features found in primate cerebral cortices, such as modular small-world networks and energetically efficient mixed-selective codes. These features emerge together, suggesting a strong interdependence driven by biological optimization processes. seRNNs serve as a bridge between structural and functional neuroscience research.

Jascha Achterberg, Danyal Akarca, D. J. Strouse, John Duncan, Duncan E. Astle