Peking University's RNN model links fine-scale synaptic motifs to brain dynamics

A team from Peking University and Université Côte d'Azur has developed a novel recurrent neural network (RNN) framework that systematically links fine-scale synaptic motifs—specifically second-order correlated couplings—to macroscopic nonlinear population dynamics. Their work, published on arXiv (2606.27946), addresses a long-standing question: can microscale synaptic structures explain the heterogeneous activity observed across brain regions in ways that conventional circuit models cannot?

The researchers constructed random RNNs with multiple cell types, nonlinear non-negative neural responses, and controlled statistics for both marginal and pairwise synaptic correlations. They derived mean-field equations for P-population networks where pre- and postsynaptic population identities define motif strengths. Crucially, their framework requires only 2P latent dynamic variables: P for mean population activity and P for within-population variability. Both theoretical analysis and simulations reveal that chain motifs—sequential correlated couplings—create shared variability among synapses, allowing microscopic fluctuations to be integrated and influence mesoscopic dynamics.

Applying this approach to mouse primary visual cortex, the team reverse-engineered network connectivity that recapitulates heterogeneous population activity observed experimentally. The results bridge the gap between synaptic organization and nonlinear dynamics, providing a principled method to infer fine-scale connectivity from macroscopic recordings. This offers testable predictions about how specific motif structures support functional computations like sensory processing.