Neural Population Coding Is Optimized by Discrete Tuning Curves

Alexander P. Nikitin, Nigel G. Stocks, Robert P. Morse, Mark D. McDonnellPhysical 2009 · DOI · arXiv

Abstract

The sigmoidal tuning curve that maximizes the mutual information for a Poisson neuron, or population of Poisson neurons, is obtained. The optimal tuning curve is found to have a discrete structure that results in a quantization of the input signal. The number of quantization levels undergoes a hierarchy of phase transitions as the length of the coding window is varied. We postulate, using the mammalian auditory system as an example, that the presence of a subpopulation structure within a neural population is consistent with an optimal neural code.

Introduction

In neuroscience, the "tuning curve" describes how a neuron’s firing rate changes in response to a stimulus. Because neurons are inherently noisy—often modeled using Poisson statistics—researchers have long sought the optimal tuning curve that transmits the maximum information about a stimulus. While smooth, sigmoidal curves were often assumed to be ideal, these assumptions frequently relied on the unrealistic scenario of an infinite time window for spike counting. This paper addresses the biologically realistic case: determining the optimal tuning curve when a neuron has only a short, finite window of time to communicate.

Methods

The authors model the neuron as a communication channel where a continuous input $x$ is mapped to a mean firing rate $\nu$ via a tuning curve $g(x)$. This process generates a discrete number of spikes $k$ following a Poisson distribution. To find the optimal $g(x)$, the researchers:

  1. Leveraged Photonics Theory: Drew parallels between neural firing and "direct detection photon channels" in optical communications.
  2. Mathematical Proof: Demonstrated that mutual information is maximized when the distribution of firing rates is discrete.
  3. Numerical Optimization: Used stochastic gradient descent to find the exact threshold values ($\theta_i$) and firing levels ($\beta_i$) that maximize information for different maximum spike counts ($N$).

The Emergence of Quantization and Phase Transitions

A key finding is that "analog" neurons should behave like "digital" quantizers to optimize information transmission against Poisson noise. Rather than representing every minute change in input, the optimal strategy involves jumping between distinct, well-separated firing rates.

As the coding window $T$ (or the maximum possible spike count $N$) increases, the optimal number of discrete levels also increases. The system undergoes a series of "phase transitions" where, for example, a binary code splits into a ternary code, then quaternary, and so on.

Figure 1: Evolution of optimal thresholds $\theta_i$ and the resulting discrete tuning curves (insets) as the maximum spike count $N$ increases.

This suggests that a population of neurons can implement this optimal multi-step curve by organizing into subpopulations with different activation thresholds.

Key Results and Biological Alignment

The study reveals specific thresholds for coding strategies:

  • Low Signal Regime: For $N < 3$, the optimal strategy is simple binary coding.
  • The "Rule of 15": At $N=15$, the optimal code becomes quaternary (4 levels).
  • Information Scaling: Mutual information $I$ increases with $N$, showing how longer observation windows allow for more complex and informative coding schemes.
Figure 2: Optimal firing rates $\phi_i$ and mutual information $I$ as a function of the maximum spike count $N$.

These theoretical predictions align remarkably well with biological data from the mammalian auditory system. In the ear, each inner hair cell connects to approximately 15 nerve fibers, which cluster into 2 or 3 distinct subpopulations based on sensitivity thresholds. This matches the theoretical prediction for $N=15$, suggesting that evolution has optimized the auditory nerve for maximum information throughput.

Limitations and Future Work

While providing a powerful framework, the model has several limitations:

  • Sigmoidal Constraint: It assumes non-decreasing tuning curves, excluding bell-shaped responses.
  • Simplified Noise: The pure Poisson model ignores refractory periods and other non-Poisson characteristics that might smooth discrete steps in practice.
  • Metabolic Costs: The analysis focuses on information maximization without accounting for the energy cost of firing spikes.
  • Experimental Visibility: The authors suggest that these discrete steps might not be readily observed in labs because experimental conditions may not match the specific stimuli or time windows required to reveal this underlying structure.

Summary updated

Abstract

The sigmoidal tuning curve that maximizes the mutual information for a Poisson neuron, or population of Poisson neurons, is obtained. The optimal tuning curve is found to have a discrete structure that results in a quantization of the input signal. The number of quantization levels undergoes a hierarchy of phase transitions as the length of the coding window is varied. We postulate, using the mammalian auditory system as an example, that the presence of a subpopulation structure within a neural population is consistent with an optimal neural code.
BibTeX
@inproceedings{nikitin2009neural,
  author = {Alexander P. Nikitin and Nigel G. Stocks and Robert P. Morse and Mark D. McDonnell},
  title = {{Neural Population Coding Is Optimized by Discrete Tuning Curves}},
  booktitle = {Physical},
  year = {2009},
  doi = {10.1103/physrevlett.103.138101},
}
Figure