Representation of visual uncertainty through neural gain variability

Olivier J. Hénaff, Zoe M. Boundy-Singer, Kristof Meding, Corey M. Ziemba, Robbe L. T. GorisNature Communications 2020 · DOI

TL;DR

This paper proposes and validates a novel model where neural circuits represent visual uncertainty not just through average firing rates, but through the variability in their response gain. They demonstrate that this "gain variability" is specifically tuned to stimulus uncertainty and can be decoded by downstream circuits, offering a new perspective on how the brain handles unreliable sensory information.

What Problem Does This Paper Address?

Perception is inherently uncertain. Our senses receive noisy and incomplete information about the world. To make accurate decisions, our brains need to not only process sensory features but also estimate the reliability of that information. While it's understood that neural circuits must represent this uncertainty, how they do so remains a significant question. A prominent hypothesis suggests that neural response variability might encode uncertainty, but the specific aspect of this variability and its underlying mechanisms have been debated. This paper tackles this by proposing and testing a model where variability in neural response gain specifically reflects stimulus uncertainty.

Methods

The authors propose a model where a neuron's firing rate is a product of a deterministic response mean (encoding stimulus features) and a stochastic response gain (encoding uncertainty). They formalize this using a modulated Poisson process. To test this, they recorded from macaque V1 and V2 neurons while manipulating stimulus uncertainty in two ways: by altering stimulus contrast and spread.

Key methodological contributions include:

  • Modulated Poisson Model with Uncertainty Receptive Field: Extending the standard modulated Poisson model by introducing an "uncertainty receptive field" that specifically maps stimulus uncertainty to the variance of the response gain.
  • Stochastic Normalization Model: Developing a probabilistic formulation of divisive normalization. This model suggests that noise in the normalization signal, when processed through a standard gain control mechanism, naturally leads to stimulus-dependent gain variability.
  • Analysis of Gain Dynamics: Investigating whether gain fluctuations are fast or slow by fitting different temporal models to variance-to-mean relationships across varying counting window durations.
  • Population Decoding: Simulating V1 populations and developing a method to decode both stimulus features (orientation, contrast, spread) and their associated uncertainty (gain variability) from population activity.

Datasets

The study utilized existing datasets from macaque V1 and V2 neurons.

  • Orientation Experiment: Data collected from V1 neurons responding to stimuli with manipulated orientation uncertainty (contrast and spread).
  • Texture Experiment: Data from V1 and V2 neurons responding to naturalistic textures and unstructured noise stimuli, where noise was used to induce textural uncertainty.

No new benchmarks were introduced; rather, existing neural recording data was re-analyzed through the lens of the proposed model.

Key Insights and Ideas

  • Gain Variability as the Uncertainty Signal: The core idea is that the variability in a neuron's response gain, rather than just the average firing rate or overall response variance, is the key signal for representing stimulus uncertainty.
  • Uncertainty Receptive Field: The concept of a dedicated "uncertainty receptive field" that is tuned to features that induce uncertainty, analogous to classical receptive fields for stimulus features.
  • Normalization as the Mechanism: The finding that a stochastic version of divisive normalization, a well-established neural computation, can naturally generate stimulus-dependent gain variability, providing a plausible biological mechanism.
  • Slow Dynamics of Gain: Gain variability exhibits slow temporal dynamics, meaning individual neurons cannot rapidly communicate uncertainty on a trial-by-trial basis. This implies that uncertainty information is conveyed through population-level activity.
  • Population Decoding of Uncertainty: Despite slow individual neuron dynamics, downstream circuits can effectively decode both stimulus features and their uncertainty from the joint activity of a neural population, even with limited integration times and correlated gain fluctuations.

Key Results

  • Gain Variability Tracks Uncertainty: In both V1 and V2 neurons, gain variability systematically increased with stimulus uncertainty (manipulated by contrast and spread for orientation, and by texture vs. noise for texture). This relationship was quantitative and held across different sources of uncertainty.
  • Selectivity for Uncertainty: The selectivity of gain variability for stimulus uncertainty was largely invariant to the source of that uncertainty (e.g., contrast vs. spread), suggesting a unified mechanism.
  • V2 Specialization: V2 neurons, which are selective for texture features, showed gain variability tuned to textural uncertainty, while V1 neurons, lacking this selectivity, did not. This supports the idea of hierarchical emergence of uncertainty representation.
  • Normalization Model Fit: The stochastic normalization model quantitatively predicted the observed gain variability across different stimulus conditions.
  • Slow Dynamics Confirmed: Gain fluctuations were found to have slow dynamics, consistent with predictions from the normalization model and suggesting population-level encoding.
  • Successful Population Decoding: Simulated V1 populations could reliably decode stimulus orientation, contrast, and spread, and importantly, the decoded gain variability strongly correlated with the actual stimulus uncertainty and predicted orientation decoding error.

Limitations and Future Work

  • Mechanistic Origin: While the study proposes normalization as a plausible mechanism, the precise biophysical underpinnings of gain variability (e.g., neuromodulation, membrane potential fluctuations) remain to be fully resolved, requiring intracellular recordings.
  • Behavioral Link: The study primarily focuses on neural representations. While behavioral effects of uncertainty are discussed, directly linking neural gain variability to specific behavioral choices in real-time experiments would strengthen the findings.
  • Generalizability: While tested across orientation and texture, further investigation into other sensory modalities and stimulus features would be valuable to confirm the generality of the uncertainty receptive field.
  • Heuristic Decoding: The paper proposes a neurally plausible heuristic for estimating gain variability at the population level, but further experimental validation of this specific decoding mechanism in behaving animals would be beneficial.
  • Interneuronal Correlations: While the study explored the impact of shared gain fluctuations on decoding, understanding the precise role and origin of these correlations in uncertainty representation warrants further investigation.

Summary updated

Abstract

Uncertainty is intrinsic to perception. Neural circuits which process sensory information must therefore also represent the reliability of this information. How they do so is a topic of debate. We propose a model of visual cortex in which average neural response strength encodes stimulus features, while cross-neuron variability in response gain encodes the uncertainty of these features. To test this model, we studied spiking activity of neurons in macaque V1 and V2 elicited by repeated presentations of stimuli whose uncertainty was manipulated in distinct ways. We show that gain variability of individual neurons is tuned to stimulus uncertainty, that this tuning is specific to the features encoded by these neurons and largely invariant to the source of uncertainty. We demonstrate that this behavior naturally arises from known gain-control mechanisms, and illustrate how downstream circuits can jointly decode stimulus features and their uncertainty from sensory population activity.
BibTeX
@inproceedings{henaff2020representation,
  author = {Olivier J. Hénaff and Zoe M. Boundy-Singer and Kristof Meding and Corey M. Ziemba and Robbe L. T. Goris},
  title = {{Representation of visual uncertainty through neural gain variability}},
  booktitle = {Nature Communications},
  year = {2020},
  doi = {10.1038/s41467-020-15533-0},
}
Figure