Fast inference of interactions in assemblies of stochastic integrate-and-fire neurons from spike recordings

Rémi Monasson, Simona CoccoJournal of Computational Neuroscience 2011 · DOI · arXiv

Introduction

Neuroscientists can now record the electrical activity of hundreds of neurons simultaneously. However, transforming these "spike trains"—sequences of timestamps—into a map of functional interactions remains a significant computational challenge. Traditional methods, such as pairwise cross-correlation, are often too simplistic to distinguish between direct synaptic connections and shared external signals. Conversely, more sophisticated statistical models often ignore the underlying biophysics of the cell membrane.

Schematic of the inference problem: mapping observed spike trains to a network of interacting neurons.

This paper introduces two fast Bayesian procedures to map these hidden "conversations" using a Leaky Integrate-and-Fire (LIF) model. By calculating the most likely voltage trajectories of neurons, the authors achieve a computational speed that scales linearly with the number of spikes, enabling the analysis of massive datasets containing millions of spikes across dozens of neurons.

Methods

The authors frame the problem as a Bayesian inference task: given a set of spike times, what are the most likely synaptic couplings ($J_{ij}$) and external currents ($I_i$)?

The Optimal Path Approach

Instead of simulating every possible voltage fluctuation, the authors utilize a "Fixed Threshold" (FT) procedure. They calculate the optimal potential—the single most likely voltage trajectory that would lead to the observed spikes. This reduces a complex stochastic problem into a deterministic optimization.

Moving Threshold (MT) Procedure

In the presence of moderate noise, the FT method tends to underestimate interactions because it doesn't account for the "bias" that noise introduces; noise can push a neuron to fire earlier than expected. The MT procedure heuristically adjusts the firing threshold to account for these fluctuations, significantly improving accuracy in noisy regimes.

Visualization of the optimal voltage trajectories and thresholding mechanisms used in the inference procedures.

Computational Efficiency

A key insight is that the log-likelihood of the model is concave, allowing for the use of the Newton-Raphson method for optimization. The resulting algorithm scales as $O(S \cdot N^2)$, where $S$ is the total number of spikes and $N$ is the number of neurons. Because noise is assumed to be independent between neurons, the inference for a large network can be decoupled into $N$ independent problems, facilitating parallel computing.

Results

The algorithm was validated using both synthetic and biological datasets, including recordings from the salamander retina under spontaneous ("dark") and stimulated ("natural movie") conditions.

  • Speed: The algorithm is remarkably efficient. Inferring interactions for retina datasets (tens of neurons, hundreds of thousands of spikes) took approximately 40 seconds on a standard desktop core.
  • Accuracy: On synthetic data, the "Moving Threshold" procedure recovered positive couplings with high precision even at moderate noise levels, where the simpler FT method failed.
  • Biological Insights: When applied to the retina, the model captured the temporal structure of cross-correlograms better than previous "perfect integrator" models. It demonstrated that as membrane "leakiness" increases, the inferred interactions become more asymmetric and directional.

Discussion

A major advantage of this LIF-based approach is its ability to detect directional influence, distinguishing whether neuron A drives neuron B or vice versa. However, the study also highlighted the "stealth" of inhibition: strong negative interactions are fundamentally harder to infer than positive ones. An inhibitory pulse pushes the voltage away from the threshold; if the neuron was not already near firing, this change leaves little trace in the spike timing.

Future work may address current limitations, such as the assumption of instantaneous synapses and known membrane parameters. The authors suggest that future iterations could infer membrane conductance and capacitance alongside synaptic couplings to provide a more complete picture of neural network dynamics.

Summary updated

BibTeX
@article{monasson2011fast,
  author = {Rémi Monasson and Simona Cocco},
  title = {{Fast inference of interactions in assemblies of stochastic integrate-and-fire neurons from spike recordings}},
  journal = {Journal of Computational Neuroscience},
  year = {2011},
  doi = {10.1007/s10827-010-0306-8},
}
Figure