Correlated Variability in the Brain Leads to Highly Suboptimal Sensory Coding

Academic Background

The brain perceives the world through the activity of neural populations, but it remains unclear whether the computational goal of sensory coding is to support the discrimination of sensory stimuli or to generate an internal model of the sensory world. Experimentally, correlated variability (noise correlations) across neural populations is commonly observed, and many studies have shown that correlated variability can improve the discriminative capacity of sensory coding compared to a null model with no correlations. However, these studies have not explored whether correlated variability is optimal for discriminative sensory coding. If the computational goal of sensory coding is discriminative, then correlated variability should be optimized to support this goal. This paper evaluates the optimality of correlated variability for discriminative sensory coding in neural populations by developing two novel null models and finds that correlated variability exhibits highly suboptimal discriminative sensory coding across multiple datasets.

Source of the Paper

This paper was co-authored by Jesse A. Livezey, Pratik S. Sachdeva, Maximilian E. Dougherty, Mathew T. Summers, and Kristofer E. Bouchard. The authors are affiliated with multiple research institutions, including Lawrence Berkeley National Laboratory, University of California, Berkeley, and University of California, San Francisco. The paper was first published on November 6, 2024, in the Journal of Neurophysiology, with the DOI 10.1152/jn.00313.2024.

Research Process

1. Data Collection and Preprocessing

The study used three distinct neural datasets, covering different animal models, brain regions, and recording modalities: - Retina Dataset: Calcium imaging was used to record the responses of mouse retinal ganglion cells (RGCs) to drifting bar stimuli. - Primary Visual Cortex (V1) Dataset: Single-unit electrophysiology was used to record the responses of macaque V1 neurons to drifting grating stimuli. - Primary Auditory Cortex (PAC) Dataset: Micro-electrocorticography (μECoG) was used to record the responses of rat PAC to tone stimuli.

Each dataset underwent preprocessing, including baseline correction and normalization, to ensure data comparability and analytical accuracy.

2. Development of Null Models

To evaluate the optimality of correlated variability for discriminative sensory coding, the study developed two novel null models: - Uniform Correlation (UC) Null Model: Maintains the private variance of each neuron while allowing random variations in correlations between neurons. - Factor Analysis (FA) Null Model: Decomposes the observed covariance into private variance and shared variability, generating null models by rotating the shared variability.

These two null models are based on different biological assumptions. The UC model assumes that correlations between neurons can be adjusted by changing the strength of connections between neurons, while the FA model assumes that shared variability originates from unobserved neural activity.

3. Calculation of Linear Fisher Information (LFI)

The study used Linear Fisher Information (LFI) as a metric to measure the discriminative sensory coding capacity of neural populations. LFI quantifies the amount of information in neural population activity for discriminating between different stimuli. For each dataset, the study calculated the observed LFI and compared it to the LFI distributions generated by the null models to assess whether the observed correlated variability was optimal.

4. Results Analysis

Through extensive analysis of subpopulations and stimulus pairings, the study found that the observed correlated variability exhibited highly suboptimal discriminative sensory coding across multiple datasets. Specifically: - UC Null Model: The observed LFI was lower than the LFI distribution generated by the UC null model in most cases, indicating that correlated variability was suboptimal for discriminative sensory coding. - FA Null Model: The observed LFI was also generally lower than the LFI distribution generated by the FA null model, further confirming the suboptimality of correlated variability.

Additionally, the study found that suboptimality became more pronounced as the size of the neural population increased, suggesting that larger neural populations performed worse in discriminative sensory coding.

Key Findings

1. Suboptimality of Correlated Variability

The results indicate that the observed correlated variability exhibited highly suboptimal discriminative sensory coding across multiple datasets. Specifically: - UC Null Model: The observed LFI was lower than the LFI distribution generated by the UC null model in most cases, indicating that correlated variability was suboptimal for discriminative sensory coding. - FA Null Model: The observed LFI was also generally lower than the LFI distribution generated by the FA null model, further confirming the suboptimality of correlated variability.

2. Impact of Biological Constraints

The study also found that biological constraints limit the ability of neural populations to achieve optimal correlated variability. Specifically: - UC Null Model: Optimal correlation matrices often had absolute pairwise correlations close to 1, which were never observed in the experimental data. - FA Null Model: Optimal covariance matrices often had single-unit distributions different from the experimental data, indicating that optimal covariances were biologically infeasible.

3. Exponential Shrinkage of Optimal Subpopulations

The study further found that optimal subpopulations shrank exponentially as the size of the neural population increased. Specifically: - UC Null Model: In the retina and V1 datasets, almost no subpopulations achieved optimality. - FA Null Model: In the PAC dataset, the proportion of optimal subpopulations decreased rapidly as the neural population size increased.

Conclusion

By developing two novel null models, this study evaluated the optimality of correlated variability for discriminative sensory coding in neural populations and found that correlated variability exhibited highly suboptimal discriminative sensory coding across multiple datasets. The results indicate that biological constraints limit the ability of neural populations to achieve optimal correlated variability, and optimal subpopulations shrink exponentially as the neural population size increases. These findings reveal the suboptimality of correlated variability in sensory coding and provide new insights into the fundamental principles of neural computation.

Research Highlights

• Novel Null Models: The study developed two novel null models (UC and FA) to evaluate the optimality of correlated variability for discriminative sensory coding.
• Validation Across Multiple Datasets: The study validated the suboptimality of correlated variability across multiple datasets, covering different animal models, brain regions, and recording modalities.
• Impact of Biological Constraints: The study revealed the limitations imposed by biological constraints on neural populations achieving optimal correlated variability.
• Exponential Shrinkage of Optimal Subpopulations: The study discovered that optimal subpopulations shrink exponentially as the neural population size increases.

Research Significance

This study provides new perspectives on the role of correlated variability in sensory coding within neural populations, revealing its suboptimality and the biological constraints involved. These findings are not only significant for the field of neuroscience but also provide a theoretical foundation for developing more effective neural coding models in the future.