Uncovering statistical structure in large-scale neural activity with Restricted Boltzmann Machines

A team of researchers has successfully applied a foundational AI architecture—the Restricted Boltzmann Machine (RBM)—to decode the complex, coordinated activity of up to 2,000 neurons recorded simultaneously in the mouse brain. Using data from the Allen Institute's Visual Behavior Neuropixels dataset, the model moved beyond traditional statistical methods limited to pairwise interactions. By employing latent variables, the RBM captured higher-order statistical dependencies, accurately reproducing empirical correlations and the global distribution of population activity. This demonstrates that generative models from machine learning can serve as powerful tools for large-scale neuroscience.

The study's key finding is that the model's inferred parameters reveal an anatomically structured 'effective interaction network.' Neurons within visual cortical areas showed significantly stronger couplings, aligning with visually driven behavior, while connections between different brain regions were weaker. Remarkably, even when trained on temporally shuffled data, the model's simulations reproduced the global relaxation dynamics of neural activity. This work, published on arXiv, bridges statistical physics and modern AI, offering a scalable, predictive, and interpretable framework for understanding how microscopic neural interactions give rise to macroscopic brain-wide patterns.