Identifying Connectivity Distributions from Neural Dynamics Using Flows

A team of researchers including Timothy Doyeon Kim, Ulises Pereira-Obilinovic, and Uygar Sümbül has published a novel AI framework that fundamentally shifts how we infer brain connectivity from neural recordings. The core problem in neuroscience is that multiple, vastly different wiring diagrams (connectivity structures) can produce identical patterns of neural activity, making standard inference methods degenerate and prone to recovering spurious, irrelevant connections. The team's solution moves beyond estimating a single connectivity matrix. Instead, their method uses principles of maximum entropy and a machine learning technique called continuous normalizing flows (CNFs) to learn the entire, maximally unbiased *distribution* of possible synaptic weights that could have generated the observed dynamics.

This approach, trained via flow matching, is uniquely capable of capturing complex, real-world connectivity patterns like the heavy-tailed distributions found in empirical data, which simpler models miss. The researchers rigorously validated their framework on synthetic neural networks engineered to produce specific computational motifs—like multistable attractors, limit cycles, and ring attractors—and demonstrated its practical utility by applying it to recordings from the rat frontal cortex during decision-making. The work represents a paradigm shift: from the flawed goal of recovering 'the' connectivity to identifying which connectivity features are computationally necessary versus which are mere artifacts of an underconstrained inference problem.