Title:Emergent rate-based dynamics in duplicate-free populations of spiking neurons

Abstract:Can Spiking Neural Networks (SNNs) approximate the dynamics of Recurrent Neural Networks (RNNs)? Arguments in classical mean-field theory based on laws of large numbers provide a positive answer when each neuron in the network has many "duplicates", i.e. other neurons with almost perfectly correlated inputs. Using a disordered network model that guarantees the absence of duplicates, we show that duplicate-free SNNs can converge to RNNs, thanks to the concentration of measure phenomenon. This result reveals a general mechanism underlying the emergence of rate-based dynamics in large SNNs.