When do microcircuits produce beyond-pairwise correlations?

Andrea K. Barreiro, Julijana Gjorgjieva, Fred Rieke, Eric Shea-BrownFrontiers Comput. Neurosci 2014 · DOI

TL;DR

This paper investigates why neural microcircuits, particularly in the retina, often exhibit activity patterns that can be well-described by simple pairwise interactions, even though more complex, higher-order correlations are theoretically possible. They find that the shape of the input signal is a critical factor: bimodal inputs lead to significant deviations from pairwise models, while unimodal inputs, especially after filtering, result in activity that is largely captured by pairwise statistics.

What Problem Does This Paper Address?

Understanding how information is encoded in the brain relies heavily on deciphering the collective activity of large neural populations. A major challenge is the sheer complexity of these interactions. While theoretically, the number of possible correlations grows exponentially with the number of neurons, empirical studies have shown that for some neural systems, like retinal ganglion cells (RGCs), pairwise interactions can explain a surprisingly large portion of the observed activity. This raises a fundamental question: what circuit features and input statistics allow for this simplification, and when do higher-order interactions become essential? The authors aim to systematically explore the conditions under which pairwise models succeed or fail, moving beyond simply observing the phenomenon to understanding its underlying mechanisms.

Methods

The authors employ a combination of computational modeling and statistical analysis. They use:

  • Mechanistic Circuit Models: They build detailed models of RGCs, incorporating biologically realistic features like integrate-and-fire dynamics, synaptic conductances, and receptive field properties. These models are then simulated under various input conditions.
  • Pairwise Maximum Entropy (PME) Models: This is a standard statistical framework used to predict the probability distribution of neural activity based solely on pairwise correlations. The authors compare the output of their mechanistic models to the predictions of PME models.
  • Kullback-Leibler (KL) Divergence (DKL): This is a key metric used to quantify the difference between the true (simulated) probability distribution of neural activity and the PME model's prediction. A low DKL value indicates that the PME model is a good fit.
  • Simplified Thresholding Model: To gain analytical insight, they also use a simpler feedforward circuit model where neurons spike if their summed input crosses a threshold. This allows for more systematic exploration of parameter space and theoretical analysis.

Datasets

The study does not use external datasets for training or evaluation. Instead, they generate simulated neural activity based on:

  • Experimental Data for RGC Models: They use measured properties of primate ON parasol ganglion cells, including their light-dependent synaptic inputs and voltage dynamics, to parameterize their mechanistic RGC model.
  • Synthetically Generated Stimuli: They create various input stimuli, including Gaussian, binary, Cauchy, and heavy-tailed distributions, with controlled correlations, variances, and refresh rates.

Key Insights and Ideas

  • Input Signal Shape is Paramount: The most striking finding is that the shape of the common input signal is a primary determinant of whether higher-order correlations emerge. Bimodal inputs lead to significant deviations from pairwise models, while unimodal inputs generally do not.
  • Filtering Suppresses Bimodality: Neural circuits, like the RGCs studied, often have filtering properties (e.g., biphasic filters) that can transform bimodal input stimuli into unimodal conductance inputs, thereby suppressing higher-order correlations.
  • Recurrent Coupling Has a Non-Monotonic Effect: While recurrent connections can increase higher-order interactions, this effect is not monotonic. It tends to be maximized at intermediate coupling strengths, and can even decrease at very strong coupling.
  • Pairwise Models are Surprisingly Robust: For many biologically plausible scenarios, especially those involving unimodal inputs and typical RGC filtering, pairwise models provide an excellent approximation of neural activity.

Key Results

  • Low DKL for Unimodal Inputs: In simulations with unimodal common inputs (Gaussian, skewed), the KL divergence between the simulated activity and the PME model was consistently low (e.g., DKL(P, P) < 0.0038), indicating good fits.
  • Higher DKL for Bimodal Inputs: Bimodal common inputs led to significantly larger KL divergences (e.g., up to 0.091 in the simplified model), indicating substantial deviations from pairwise predictions.
  • RGC Filtering Reduces DKL: When simulating RGCs with their characteristic filtering properties, even bimodal stimuli often resulted in unimodal conductance inputs and low DKL values.
  • Recurrent Coupling Impact: For RGCs, experimentally measured gap junction coupling had minimal impact on DKL. However, in the simplified model, intermediate coupling strengths led to increased DKL, with the effect being more pronounced at higher population sizes.
  • Scaling with Population Size: In the simplified model, the deviation from pairwise models (measured by DKL/N) increased moderately with population size for unimodal inputs, but this effect saturated for bimodal inputs.

Limitations and Future Work

  • Focus on Small, Symmetric Circuits: Much of the analysis was performed on small, often permutation-symmetric circuits. The authors acknowledge that results might differ for larger, more heterogeneous networks with more complex connectivity.
  • Specific RGC Type: The RGC model was based on ON parasol cells. Future work could explore other RGC types and stimulus conditions where filtering might be less effective at suppressing bimodality.
  • Adaptation of Filters: The study did not consider the adaptation of stimulus filters in response to higher-order stimulus characteristics. This is an avenue for future research.
  • Information Encoding: The paper focuses on statistical correlations. Future work could investigate how these higher-order correlations impact information encoding and decoding in neural populations.
  • More Complex Network Architectures: The study primarily explored feedforward and simple recurrent connections. Investigating lateral and feedback coupling, as well as multi-layer networks, could reveal further mechanisms for generating higher-order interactions.

Summary updated

BibTeX
@article{barreiro2014when,
  author = {Andrea K. Barreiro and Julijana Gjorgjieva and Fred Rieke and Eric Shea-Brown},
  title = {{When do microcircuits produce beyond-pairwise correlations?}},
  journal = {Frontiers Comput. Neurosci.},
  year = {2014},
  doi = {10.3389/fncom.2014.00010},
}
Figure