Percept formation from neural populations in sensory decision-making tasks

Adrien Wohrer, Christian K. MachensarXiv 2013 · DOI · arXiv

Abstract

We study a standard linear readout model of perceptual integration from a population of sensory neurons. We show that the readout can be associated to a set of characteristic equations which summarize the joint trial-to-trial covariance structure of neural activities and animal percept. These characteristic equations implicitly determine the readout parameters that were used by the animal to create its percept. In particular, they implicitly constrain the temporal integration window $w$ and the typical number of neurons $K$ which give rise to the percept. Comparing neural and behavioral sensitivity alone cannot disentangle these two sources of perceptual integration, so the characteristic equations also involve a measure of choice signals, like those assessed by the classic experimental measure of choice probabilities. We then propose a statistical method of analysis which allows to recover the typical scales of integration $w$ and $K$ from finite numbers of recorded neurons and recording trials, and show the efficiency of this method on an artificial encoding network. We also study the statistical method theoretically, and relate its laws of convergence to the underlying structure of neural activity in the population, as described through its singular value decomposition. Altogether, our method provides the first thorough interpretation of feedforward percept formation from a population of sensory neurons. It can readily be applied to experimental recordings in classic sensory decision-making tasks, and hopefully provide new insights into the nature of perceptual integration.

Introduction

Neuroscientists have long sought to understand how the brain transforms noisy sensory signals into behavioral decisions. Sensory neurons exhibit significant trial-to-trial variability; the same stimulus can elicit different responses across repetitions. To mitigate this noise, the brain must integrate signals across large neural populations and over time. However, identifying the specific readout strategy—how many neurons ($K$) are used and the duration of the integration window ($w$)—is a fundamental challenge. A high behavioral accuracy could result from integrating many neurons briefly or a few neurons for a longer period. This paper introduces a statistical framework to "reverse-engineer" these parameters from limited experimental data.

The Standard Linear Readout Model

The authors model percept formation as a weighted sum of neural activities integrated over a temporal window $w$.

The standard linear readout model (a) and its geometric interpretation (b), showing how neural activity is integrated over time and across neurons.

In this model, the estimated stimulus $\hat{f}$ is derived from the spikes of $N$ neurons, each weighted by a factor $a_i$. The geometric interpretation views this process as projecting high-dimensional neural activity onto a specific readout axis. This projection effectively averages out independent noise while preserving the signal relevant to the task.

Statistical Framework and Characteristic Equations

The core of the proposed method lies in three "characteristic equations" that link neural tuning, noise covariance, and choice signals.

Overview of the neural and behavioral data components used to infer readout parameters, including tuning, noise covariance, and percept covariance signals.

As illustrated in Figure 1, the method utilizes several key data components:

  1. Tuning signals ($\beta_i(t)$): These describe how individual neurons respond to varying stimulus intensities.
  2. Noise covariance ($\gamma_{ij}(t,s)$): This captures the trial-to-trial correlations in noise between different neurons and across different time points.
  3. Percept Covariance (PCV) signals ($\pi_i^*(t)$): A time-resolved version of the classic "Choice Probability" measure, PCV describes how a neuron's activity correlates with the animal's decision error.

By combining these measures, the researchers can break the inherent symmetry between the number of neurons $K$ and the integration window $w$. While behavioral sensitivity alone is ambiguous, the PCV signals provide the necessary constraints to identify the global readout policy. This theoretical grounding, often analyzed via Singular Value Decomposition (SVD), explains why their statistical averages converge toward the truth as more data is collected, as the readout essentially selects specific "modes" of population activity.

Results from Artificial Encoding Networks

The framework was validated using a simulated network of 500 leaky integrate-and-fire (LIF) neurons with realistic cortical dynamics. With a "ground truth" readout set to $K=40$ neurons and $w=50ms$, the statistical method successfully recovered these parameters using only 100 recorded neurons and 150 trials—conditions typical for modern electrophysiology experiments.

A key finding was the robustness of the "Tuning-Multiplied PCV" measure. By multiplying a neuron's choice signal by its tuning, the authors created a population average that remains consistent regardless of which specific neurons are sampled. This stability allows for accurate extrapolation from a small recorded subset to the entire population.

Discussion and Limitations

While providing a powerful tool for interpreting neural data, the model relies on a linear readout assumption. While biologically plausible for many sensory tasks, the brain may employ non-linear integration in more complex decision-making scenarios. Additionally, the current model is strictly feedforward. If significant top-down feedback from decision areas influences sensory activity, the PCV curves may deviate from the model's predictions. Future work could extend this framework to account for non-stationary readout times, where an animal's internal decision deadline varies from trial to trial.

Summary updated

Abstract

We study a standard linear readout model of perceptual integration from a population of sensory neurons. We show that the readout can be associated to a set of characteristic equations which summarize the joint trial-to-trial covariance structure of neural activities and animal percept. These characteristic equations implicitly determine the readout parameters that were used by the animal to create its percept. In particular, they implicitly constrain the temporal integration window w and the typical number of neurons K which give rise to the percept. Comparing neural and behavioral sensitivity alone cannot disentangle these two sources of perceptual integration, so the characteristic equations also involve a measure of choice signals, like those assessed by the classic experimental measure of choice probabilities. We then propose a statistical method of analysis which allows to recover the typical scales of integration w and K from finite numbers of recorded neurons and recording trials, and show the efficiency of this method on an artificial encoding network. We also study the statistical method theoretically, and relate its laws of convergence to the underlying structure of neural activity in the population, as described through its singular value decomposition. Altogether, our method provides the first thorough interpretation of feedforward percept formation from a population of sensory neurons. It can readily be applied to experimental recordings in classic sensory decision-making tasks, and hopefully provide new insights into the nature of perceptual integration.
BibTeX
@misc{wohrer2013percept,
  author = {Adrien Wohrer and Christian K. Machens},
  title = {{Percept formation from neural populations in sensory decision-making tasks}},
  year = {2013},
  eprint = {1303.1939},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/1303.1939},
}
Figure