Paper List

期刊: ArXiv Preprint
发布日期: 2026-03-11
Computational NeuroscienceStatistical Physics

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

Université Paris-Saclay, CNRS, INRIA, LISN, Gif-sur-Yvette, France | Universidad Complutense de Madrid, Spain | Princeton University, USA | City University of New York, USA

Nicolas Béreux, Giovanni Catania, Aurélien Decelle, Francesca Mignacco, Alfonso de Jesús Navas Gómez, Beatriz Seoane
Figure

30秒速读

IN SHORT: This paper addresses the core challenge of modeling large-scale neural population activity (1500-2000 neurons) with interpretable higher-order interactions, overcoming limitations of traditional pairwise maximum-entropy models.

核心创新

  • Methodology Demonstrates that Restricted Boltzmann Machines can be trained on thousands of simultaneously recorded neurons using efficient MCMC sampling, achieving accurate reproduction of both pairwise and higher-order correlations.
  • Methodology Provides a principled mapping from RBM parameters to explicit multi-body interaction spin models, enabling direct extraction of effective synaptic networks including higher-order couplings.
  • Biology Reveals anatomically structured effective interactions: stronger intra-area couplings within visual cortical regions and weaker, more diffuse cross-area couplings, correlating with functional engagement during visual tasks.

主要结论

  • RBMs accurately reproduce empirical statistics of neural recordings, matching pairwise correlations, higher-order correlations, and global population activity distributions with high fidelity.
  • The inferred effective couplings show clear anatomical organization: intra-visual cortical interactions are stronger and more coherent than cross-area couplings, consistent with functional specialization.
  • Despite being trained on static snapshots, RBM-generated samples via MCMC simulations accurately capture global neural relaxation dynamics, suggesting the model encodes temporal structure implicitly.
研究空白： Traditional maximum-entropy models are limited to pairwise interactions and scale poorly to modern large-scale recordings (thousands of neurons), while deep generative models lack interpretability for extracting explicit interaction networks.

摘要: Large-scale electrophysiological recordings now allow simultaneous monitoring of thousands of neurons across multiple brain regions, revealing structured variability in neural population activity. Understanding how these collective patterns emerge from microscopic neural interactions requires models that are scalable, predictive, and interpretable. Statistical physics provides principled frameworks to address this complexity, including maximum-entropy models that offer transparent descriptions of collective neural activity in small populations, but remain largely limited to pairwise interactions and modest system sizes. Here, we use Restricted Boltzmann Machines (RBMs) to model the activity of ∼1500–2000 simultaneously recorded neurons from the Allen Institute Visual Behavior Neuropixels dataset, spanning multiple cortical and subcortical regions of the mouse brain. RBMs are energy-based models that extend the maximum-entropy framework through latent variables, enabling the capture of higher-order dependencies while allowing explicit extraction of effective synaptic networks, including interactions beyond pairwise. Recent advances in efficient Markov Chain sampling and training enable accurate learning of these models at this scale. We show that RBMs reproduce the complex statistics of neural recordings with high accuracy. Generated samples match empirical pairwise and higher-order correlations, as well as global statistics such as the distribution of population activity. Beyond accurate data reconstruction, the inferred parameters provide direct access to effective interactions between neurons, revealing dominant coordination patterns in population activity. These couplings exhibit clear anatomical structure: neurons within visual cortical areas form coherent blocks of stronger interactions, consistent with shared engagement during visually driven behavior, whereas cross-area couplings are weaker and more diffuse. Furthermore, despite not being trained to reproduce temporal dependencies, Markov Chain Monte Carlo simulations of the model accurately reproduce the global neural relaxation dynamics. These results establish RBMs as scalable tools to extract interpretable statistical structure from large-scale neural recordings, linking collective neural activity to the organization of brain regions and task-related behavior.