Inference of a mesoscopic population model from population spike trains

Alexandre René, André Longtin, Jakob H. Macke

Abstract

To understand how rich dynamics emerge in neural populations, we require models exhibiting a wide range of activity patterns while remaining interpretable in terms of connectivity and single-neuron dynamics. However, it has been challenging to fit such mechanistic spiking networks at the single neuron scale to empirical population data. To close this gap, we propose to fit such data at a meso scale, using a mechanistic but low-dimensional and hence statistically tractable model. The mesoscopic representation is obtained by approximating a population of neurons as multiple homogeneous 'pools' of neurons, and modelling the dynamics of the aggregate population activity within each pool. We derive the likelihood of both single-neuron and connectivity parameters given this activity, which can then be used to either optimize parameters by gradient ascent on the log-likelihood, or to perform Bayesian inference using Markov Chain Monte Carlo (MCMC) sampling. We illustrate this approach using a model of generalized integrate-and-fire neurons for which mesoscopic dynamics have been previously derived, and show that both single-neuron and connectivity parameters can be recovered from simulated data. In particular, our inference method extracts posterior correlations between model parameters, which define parameter subsets able to reproduce the data. We compute the Bayesian posterior for combinations of parameters using MCMC sampling and investigate how the approximations inherent to a mesoscopic population model impact the accuracy of the inferred single-neuron parameters.

Introduction

In neuroscience, a persistent trade-off exists between model interpretability and statistical tractability. Microscopic mechanistic models simulate individual neurons with biological richness (e.g., synaptic weights, adaptation) but are notoriously difficult to fit to experimental data due to their massive dimensionality. Conversely, phenomenological models like RNNs or state-space models excel at data fitting but offer little insight into the underlying biological machinery. This paper addresses this gap by focusing on the mesoscopic scale, asking whether a detailed spiking model can be simplified enough for statistical tractability without losing the biological meaning of its parameters.

Methods and Framework

The authors utilize the "mesoGIF" model, which approximates large populations of neurons as multiple homogeneous "pools." The core of the approach is a closed-form likelihood function for aggregate population activity, formulated as a Markovian process to ensure mathematical tractability.

Conceptual framework for mesoscopic model inference.

The framework involves several key technical components:

  • Likelihood Derivation: Enabling efficient parameter estimation by deriving a closed-form likelihood for aggregate activity.
  • Optimization: Using the adam optimizer for Maximum A Posteriori (MAP) estimation to find best-fit parameters.
  • Bayesian Inference: Employing Hamiltonian Monte Carlo (HMC) sampling via PyMC3 to quantify parameter uncertainty and identify correlations between biological variables.

Validation was performed using simulated datasets from microscopic Generalized Integrate-and-Fire (GIF) neurons across various architectures, including 2-population (Excitatory-Inhibitory) models and 4-population cortical column models. Input protocols included sinusoidal noise, Ornstein-Uhlenbeck processes, and sharp impulse ramps.

Key Results

The method demonstrates robust parameter recovery and generalization. In a 2-population model, 14 distinct parameters (connectivity, time constants, noise) were successfully recovered. The approach scaled to a 4-population model with 36 free parameters while maintaining high performance.

Parameter recovery and activity matching in the 2-population model.

Key findings include:

  • Effective Parameters: The inference method finds "effective" parameters that accurately reproduce true population dynamics, even when simple parameter averaging fails due to population heterogeneity.
  • Parameter Co-dependence: Bayesian posteriors revealed high correlations between parameters. For instance, an increase in excitatory connectivity can be statistically canceled by an increase in inhibitory connectivity, explaining how diverse parameter sets can yield similar brain activity.
  • Data Efficiency: High accuracy was achieved with as little as 5-10 seconds of population data (approximately 20,000 spikes).
  • Generalization: Models fit on simple stimuli generalized well to complex, unseen inputs, provided the underlying biological assumptions (like the "quasi-renewal" property) held true.

Discussion and Limitations

While powerful, the mesoscopic approximation relies on the "quasi-renewal" property—the assumption that neurons quickly "forget" their history. This can lead to inaccuracies during high-frequency bursting or very abrupt stimulus transitions. Additionally, while more efficient than microscopic fitting, Bayesian sampling remains computationally intensive, particularly for larger models (the 4-population model required over 200 hours of computation). Future work will focus on handling latent (unobserved) populations and applying the framework to real-world experimental data, such as Neuropixels recordings.

Summary updated

Abstract

To understand how rich dynamics emerge in neural populations, we require models exhibiting a wide range of activity patterns while remaining interpretable in terms of connectivity and single-neuron dynamics. However, it has been challenging to fit such mechanistic spiking networks at the single neuron scale to empirical population data. To close this gap, we propose to fit such data at a meso scale, using a mechanistic but low-dimensional and hence statistically tractable model. The mesoscopic representation is obtained by approximating a population of neurons as multiple homogeneous `pools' of neurons, and modelling the dynamics of the aggregate population activity within each pool. We derive the likelihood of both single-neuron and connectivity parameters given this activity, which can then be used to either optimize parameters by gradient ascent on the log-likelihood, or to perform Bayesian inference using Markov Chain Monte Carlo (MCMC) sampling. We illustrate this approach using a model of generalized integrate-and-fire neurons for which mesoscopic dynamics have been previously derived, and show that both single-neuron and connectivity parameters can be recovered from simulated data. In particular, our inference method extracts posterior correlations between model parameters, which define parameter subsets able to reproduce the data. We compute the Bayesian posterior for combinations of parameters using MCMC sampling and investigate how the approximations inherent to a mesoscopic population model impact the accuracy of the inferred single-neuron parameters.
BibTeX
@article{rene2020inference,
  author = {Alexandre René and André Longtin and Jakob H. Macke},
  title = {{Inference of a mesoscopic population model from population spike trains}},
  journal = {Neural Computation},
  year = {2020},
  url = {https://openreview.net/forum?id=LM0kEc-uutT},
}
Figure