Simulation Based Inference of a Simple Neural Network Structure

A team of researchers from IRMA (Institut de Recherche Mathématique Avancée) has developed a novel simulation-based inference method to uncover the structure of neural networks from sparse electrophysiological recordings. The work, led by Pierre Charitat, Ségolen Geffray, and Christophe Pouzat, addresses a fundamental challenge in neuroscience: while modern extracellular electrode arrays can record spikes from many neurons simultaneously, these recordings capture only a tiny fraction of the total neuronal population in a network. Traditional approaches using cross-correlation functions to infer connectivity are severely compromised by this under-sampling.

The researchers propose focusing on simpler, more robust spike train statistics—specifically the empirical spike frequency and interspike interval distribution. By estimating the sampling distributions of these statistics through simulations, they can infer the underlying network structure from just a few observed spike trains. Their method was tested on a toy model and showed significantly better performance than conventional sub-network reconstruction approaches, particularly for estimating the connection probability of the original network. This work, published on arXiv (2604.18599), opens new avenues for understanding neural circuit architecture from limited experimental data, with potential applications in both basic neuroscience and neuroprosthetics.