Constrained brain volume in an efficient coding model explains the fraction of excitatory and inhibitory neurons in sensory cortices

Arish Alreja, Ilya Nemenman, Christopher J. RozellbioRxiv 2020 · DOI

TL;DR

This paper proposes that the conserved ratio of excitatory to inhibitory neurons (E:I ratio) in mammalian sensory cortices is explained by an efficient coding model under a fixed brain volume constraint. They found that a specific E:I ratio optimizes stimulus representation, sparsity, and metabolic efficiency, and this optimal ratio correlates with species-specific neural activity sparsity.

What Problem Does This Paper Address?

The mammalian brain exhibits remarkable consistency in the E:I ratio (around 3:1 to 9:1) across different sensory areas and species, despite vast differences in the total number of neurons. This consistency is puzzling, as it's unclear why this specific balance of neuron types is maintained. While efficient coding models suggest that increasing both excitatory and inhibitory neurons improves stimulus representation, brain volume limitations prevent simultaneous increases. This paper aims to explain the conserved E:I ratio by investigating how it impacts model performance under a fixed volume constraint, exploring the trade-offs between representation fidelity and computational efficiency.

Methods

The authors employed a sparse coding model, a well-established framework for understanding sensory processing, particularly in the visual cortex. This model aims to represent stimuli using a minimal number of active neurons, balancing reconstruction accuracy (fidelity) with the sparsity of neural activity. The core innovation here is the introduction of a volume constraint, meaning the total number of neurons (excitatory + inhibitory) was fixed. They systematically varied the E:I ratio within this constraint and evaluated model performance using three key metrics:

  • Reconstruction Error: Measures how well the model can reconstruct the original stimulus.
  • Population Sparsity: Quantifies how sparsely the excitatory neurons represent the stimulus.
  • Metabolic Energy Consumption: Estimates the energy required for neural activity.

They also analyzed the structure and function of inhibitory connections and activity at different E:I ratios and sparsity levels.

Datasets

The models were trained and evaluated using a database of natural image patches. For experimental validation, they analyzed publicly available electrophysiology data from studies on mice, monkeys, and cats in response to natural images and movies.

Key Insights and Ideas

  • The "Sweet Spot" E:I Ratio: The study reveals that for a fixed brain volume, there's a specific E:I ratio that optimally balances computational accuracy and representation capacity. Deviating from this optimal ratio, either too many excitatory or too many inhibitory neurons, leads to performance degradation.
  • Sparsity Drives E:I Ratio: A crucial finding is that the optimal E:I ratio is directly correlated with the sparsity of neural activity. Species with sparser neural representations (higher sparsity) tend to have higher optimal E:I ratios. This provides a potential explanation for inter-species variations within the observed biological E:I range.
  • Inhibitory Role Evolution: As sparsity increases (and thus the optimal E:I ratio increases), the nature of inhibitory neurons changes. They become less diverse in their firing rates and their synaptic connections become more broadly tuned, rather than highly specific. This suggests a functional adaptation of inhibition to support sparser coding.
  • Metabolic Efficiency: The optimal E:I ratio also minimizes metabolic energy consumption, suggesting that evolutionary pressures for energy efficiency might play a role in shaping neural circuit structure.

Key Results

  • The model demonstrated that coding fidelity, population sparsity, and metabolic energy consumption were all simultaneously optimized at a specific E:I ratio, which fell within the biologically observed range (around 6.5:1 for a specific sparsity constraint).
  • Increasing the model's sparsity constraint (λ) consistently led to a higher optimal E:I ratio across all performance metrics.
  • Preliminary analysis of experimental data from mice, monkeys, and cats showed a correlation between species' observed E:I ratios and their population sparsity, aligning with the model's predictions. Specifically, mice (with higher E:I ratios) exhibited higher population sparsity than monkeys and cats.

Limitations and Future Work

  • Simplistic Volume Constraint: The model's volume constraint is a simplification, not explicitly accounting for differences in cell size or non-somatic elements like axons and dendrites.
  • Inhibitory Neuron Subtypes: The model does not differentiate between various subtypes of inhibitory interneurons, which could offer more nuanced insights into their specific roles.
  • Experimental Data Challenges: The authors acknowledge that direct empirical validation is challenging due to variations in experimental methodologies, anesthesia, and recording parameters across studies. Future work should focus on controlled comparative studies.
  • Broader Applicability: While the study focuses on visual cortex, exploring this framework in other sensory areas and under different experimental conditions could reveal if these principles apply more broadly.
  • Dynamic Inhibition: Future work could explore the role of thalamic input to inhibitory cells and the dynamic contributions of different inhibitory interneuron types.

Summary updated

Abstract

Abstract The number of neurons in mammalian cortex varies by multiple orders of magnitude across different species. In contrast, the ratio of excitatory to inhibitory neurons (E:I ratio) varies in a much smaller range, from 3:1 to 9:1 and remains roughly constant for different sensory areas within a species. Despite this structure being important for understanding the function of neural circuits, the reason for this consistency is not yet understood. While recent models of vision based on the efficient coding hypothesis show that increasing the number of both excitatory and inhibitory cells improves stimulus representation, the two cannot increase simultaneously due to constraints on brain volume. In this work, we implement an efficient coding model of vision under a volume (i.e., total number of neurons) constraint while varying the E:I ratio. We show that the performance of the model is optimal at biologically observed E:I ratios under several metrics. We argue that this happens due to trade-offs between the computational accuracy and the representation capacity for natural stimuli. Further, we make experimentally testable predictions that 1) the optimal E:I ratio should be higher for species with a higher sparsity in the neural activity and 2) the character of inhibitory synaptic distributions and firing rates should change depending on E:I ratio. Our findings, which are supported by our new preliminary analyses of publicly available data, provide the first quantitative and testable hypothesis based on optimal coding models for the distribution of neural types in the mammalian sensory cortices.
BibTeX
@misc{alreja2020constrained,
  author = {Arish Alreja and Ilya Nemenman and Christopher J. Rozell},
  title = {{Constrained brain volume in an efficient coding model explains the fraction of excitatory and inhibitory neurons in sensory cortices}},
  year = {2020},
  doi = {10.1101/2020.09.17.299040},
  howpublished = {bioRxiv},
}
Figure