Not optimal, just noisy: the geometry of correlated variability leads to highly suboptimal sensory coding

Jesse A. Livezey, Pratik Sachdeva, Maximilian E. Dougherty, Mathew T. Summers, Kristofer E. BouchardbioRxiv 2022 · DOI

TL;DR

Neural populations exhibit "noise correlations" – shared variability in their firing patterns – which have often been thought to improve sensory coding. This paper introduces novel null models to rigorously assess if these correlations are optimal. The surprising finding is that, across multiple brain regions and recording modalities, the observed correlated variability leads to highly suboptimal sensory coding, suggesting that biological constraints prevent the brain from achieving optimal coding strategies.

What Problem Does This Paper Address?

The brain encodes sensory information using the coordinated activity of large populations of neurons. A key feature of this activity is "correlated variability" or "noise correlations" – the tendency for the firing rates of simultaneously recorded neurons to fluctuate together, even when presented with the same stimulus. While many studies have shown that these noise correlations can improve stimulus discrimination compared to a scenario with no correlations (often modeled by shuffling data), it's been unclear whether these observed correlations represent an optimal strategy for sensory coding. The challenge is that existing "null models" (like the shuffle null model) only compare observed correlations to a zero-correlation baseline, failing to explore the full range of biologically plausible correlation structures that might be optimal. This paper aims to fill this gap by developing and applying new null models to assess the true optimality of neural correlated variability.

Methods

The authors introduce two novel null models to assess the optimality of neural coding:

  • Uniform Correlation (UC) Null Model: This model preserves the individual neuron's response properties (mean and variance) but samples multivariate correlations uniformly from a distribution. This allows for a broader exploration of correlation structures than the shuffle null model.
  • Factor Analysis (FA) Null Model: This model is inspired by findings that neural variability can arise from both private (neuron-specific) and shared (population-wide) sources. It decomposes the covariance into fixed private variability and a low-rank shared variability component that can be rotated. This allows for biologically motivated exploration of how shared variability influences coding.

They then use these null models to calculate the Linear Fisher Information (LFI), a measure of how well a neural population can discriminate between stimuli. By comparing the LFI of experimentally observed neural activity to the LFI distributions generated by their null models, they can determine the "percentile" of the observed coding performance relative to optimal or near-chance performance.

Datasets

The study utilizes three diverse datasets:

  • Retina: Calcium imaging recordings from mouse retinal ganglion cells (RGCs) responding to drifting bars.
  • Primary Visual Cortex (V1): Single-unit spike counts from macaque monkey V1 responding to drifting gratings.
  • Primary Auditory Cortex (PAC): Micro-electrocorticography (µECOG) recordings from rat PAC responding to tone pips.

These datasets span different brain regions, animal models, and recording modalities, allowing for a broad assessment of the findings.

Key Insights and Ideas

  • The "shuffle null model" is insufficient for assessing optimality: Comparing observed correlations to a zero-correlation baseline doesn't tell us if the observed correlations are truly optimal, as other non-zero correlation structures might yield better coding.
  • Biological constraints are key: The paper strongly suggests that biological limitations prevent neural populations from achieving optimal correlated variability. This is a major departure from the idea that observed correlations are simply the best the brain can do.
  • Suboptimal coding is pervasive: Across all tested datasets and null models, the observed correlated variability leads to significantly suboptimal sensory coding.
  • Suboptimal coding worsens with population size: As more neurons are included in the population, the coding performance becomes even more suboptimal.
  • Biologically motivated subselection doesn't fix suboptimality: Subselecting neural units based on criteria like spatial proximity or stimulus tuning doesn't lead to optimal coding performance.
  • Optimal subpopulations are tiny: If optimal subpopulations exist, they are exponentially small as the dimensionality of the neural population increases, making them difficult to find through uniform sampling.

Key Results

  • Across all three datasets (retina, V1, PAC), the observed neural populations consistently showed suboptimal coding performance when compared to the uniform correlation and factor analysis null models.
  • The degree of suboptimality increased with the number of neurons included in the population.
  • For many subpopulations, achieving optimal coding would require biologically implausible neural response properties (e.g., extreme correlations, unusual Fano factors, or negative densities).
  • Biologically motivated subselection strategies (distance-based, tuning-based) did not yield optimal or near-chance coding performance.
  • The size of an optimal subpopulation, if it exists, decreases exponentially with the dimensionality of the neural population. For example, at higher dimensions, less than 2% of the population might be considered optimal.

Limitations and Future Work

  • Gaussian assumption: The null models and analyses are based on a multivariate Gaussian distribution. The authors acknowledge that real neural data can be non-Gaussian, and this might impact the results. Future work could explore null models for non-Gaussian distributions.
  • Decoding perspective: The study primarily assesses optimality from a decoding perspective (how well stimuli can be discriminated). The authors note that other normative perspectives exist (e.g., Bayesian models where variability is informative) and that correlated variability might serve other functions beyond pure decoding.
  • Model limitations: The factor analysis model, while more biologically plausible than the uniform correlation model, still has limitations. The authors suggest that further exploration of null distributions and parameterizations could be beneficial.
  • Future directions: The authors highlight the potential for using their framework to study how correlated variability is modulated by brain states (e.g., attention, behavior) and how emerging technologies will allow for recording from larger populations during naturalistic behaviors, enabling deeper investigation into the biological origins and functions of correlated variability.

Summary updated

Abstract

Abstract The brain represents the world through the activity of neural populations. Correlated variability across simultaneously recorded neurons (noise correlations) has been observed across cortical areas and experimental paradigms. Many studies have shown that correlated variability improves stimulus coding compared to a null model with no correlations. However, such results do not shed light on whether neural populations’ correlated variability achieves optimal coding. Here, we assess optimality of noise correlations in diverse datasets by developing two novel null models each with a unique biological interpretation: a uniform correlations null model and a factor analysis null model. We show that across datasets, the correlated variability in neural populations leads to highly suboptimal coding performance according to these null models. We demonstrate that biological constraints prevent many subsets of the neural populations from achieving optimality according to these null models, and that subselecting based on biological criteria leaves coding performance suboptimal. Finally, we show that the optimal subpopulation is exponentially small as a function of neural dimensionality. Together, these results show that the geometry of correlated variability leads to highly suboptimal sensory coding.
BibTeX
@misc{livezey2022not,
  author = {Jesse A. Livezey and Pratik Sachdeva and Maximilian E. Dougherty and Mathew T. Summers and Kristofer E. Bouchard},
  title = {{Not optimal, just noisy: the geometry of correlated variability leads to highly suboptimal sensory coding}},
  year = {2022},
  doi = {10.1101/2022.03.08.483488},
  howpublished = {bioRxiv},
}

Loading related content...

Figure