Abstract
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 but remain largely limited to pairwise interactions and modest system sizes. Here, we use Restricted Boltzmann Machines (RBMs) to model the activity of $\sim1500$-$2000$ simultaneously recorded neurons from the Allen Institute Visual Behavior Neuropixels dataset, spanning multiple cortical and subcortical regions of the mouse brain. RBMs extend the maximum-entropy framework through latent variables, enabling the capture of higher-order dependencies while allowing explicit extraction of effective interaction networks. Recent advances in efficient Markov Chain sampling and training enable accurate learning of these models at this scale. 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. The inferred parameters provide direct access to effective neuronal interactions, revealing coordination patterns in population activity. These couplings display clear anatomical structure: neurons within visual cortical areas show stronger interactions, consistent with visually driven behavior, while cross-area couplings are weaker. Despite being trained on temporally shuffled data, Markov Chain Monte Carlo simulations also reproduce the global relaxation dynamics of neural activity.
Introduction
In modern neuroscience, the ability to record the electrical activity of thousands of neurons simultaneously using tools like Neuropixels probes has created a massive statistical challenge. Understanding the collective "language" of these neurons requires models that can bridge the gap between simple, interpretable pairwise models and complex, "black-box" deep learning architectures.
Traditional approaches, such as Maximum Entropy (Ising) models, are highly interpretable but do not scale well; as the number of neurons ($N$) grows, the parameter space expands by $N^2$. Furthermore, Ising models typically ignore higher-order coordination. This study explores Restricted Boltzmann Machines (RBMs) as a "middle way"—a generative model that is scalable, expressive enough to capture higher-order interactions, and biologically interpretable.
Methodology
The study utilizes the Allen Institute Visual Behavior Neuropixels dataset, which includes recordings of 1,500–2,000 neurons across multiple cortical (e.g., visual cortex) and subcortical (e.g., thalamus, hippocampus) regions in mice performing a visual change-detection task. Neural activity was binned into 20ms windows and binarized.
The authors employ RBMs, a type of energy-based model with a bipartite graph structure.
The architecture consists of a visible layer representing the neurons and a hidden layer representing latent variables that coordinate neural activity. By mathematically "averaging out" the hidden units, the RBM is transformed into an effective Ising model, allowing for the calculation of 3-body and higher-order couplings. Training was conducted using Persistent Contrastive Divergence (PCD), and model complexity ($N_h$) was optimized using Parallel Trajectory Tempering to maximize test log-likelihood.
Results
The RBM demonstrated high statistical accuracy, successfully reproducing single-neuron firing rates, pairwise correlations, and three-body correlations.
Key findings include:
- Global Statistics: The model accurately matched the distribution of total population activity ($P(K)$), capturing the "heavy tails" where many neurons fire simultaneously.
- Anatomical Mapping: The inferred interaction network revealed that the strongest couplings were between neurons in the same functional brain areas (e.g., visual cortex), confirming that the model captures real biological structure.
- Emergent Dynamics: Despite being trained on static snapshots, the model's internal MCMC sampling dynamics reproduced the relaxation dynamics (the decay of activity patterns over time) seen in the empirical data.
- Parsimony: The RBM achieved these results using significantly fewer parameters than a fully connected Boltzmann Machine, representing a more efficient encoding of large-scale neural data.
Discussion and Future Work
While the RBM effectively captures statistical structure and global dynamics, it does not yet reproduce specific oscillatory patterns, such as brain waves, due to the temporal shuffling of training data. Future research may extend this framework to Recurrent RBMs to explicitly model time dependencies. Additionally, the latent variables identified by the model offer a potential pathway for decoding behavioral states or sensory stimuli directly from population activity.
Reference: Béreux, N., Catania, G., et al. (2026). Uncovering statistical structure in large-scale neural activity with Restricted Boltzmann Machines. arXiv:2603.11032v1.
Summary updated