Adaptation of retinal discriminability to natural scenes

Xuehao Ding, Dongsoo Lee, Joshua B. Melander, Surya Ganguli, Stephen A. BaccusbioRxiv 2024 · DOI

TL;DR

This paper presents a novel computational model of the retina that captures both neural sensitivity and variability to analyze how visual systems discriminate natural scenes. They found that the retina dynamically adapts its receptive fields to enhance the discrimination of local motion, and surprisingly, that correlated neural noise actually reduces information transmission, likely as a byproduct of shared circuitry.

What Problem Does This Paper Address?

How do sensory systems, specifically the retina, become so good at processing the complex and ever-changing visual world we experience? Two key questions have long puzzled neuroscientists: 1) Are visual systems optimized for the statistics of natural scenes, and 2) Does the correlated activity (noise) between neurons play a beneficial role in this optimization? While theoretical approaches and experimental measurements have offered clues, the lack of sophisticated computational models that can handle the natural complexity of stimuli and neural responses has made it difficult to definitively answer these questions. This paper aims to bridge that gap by developing a detailed model of the retinal neural code to directly analyze how the retina discriminates natural scenes.

Methods

The authors developed a stochastic encoding model of the retina that goes beyond simpler models by incorporating both neural sensitivity (how average responses change with stimuli) and neural variability (including correlated noise between neurons).

Key techniques include:

  • Deep Convolutional Neural Network (CNN): A three-layer CNN, adapted from previous work, was trained on natural images to capture the visual sensitivity of the retina. This model was designed to be mechanistically interpretable.
  • Generative Adversarial Network (GAN): A GAN was used to parameterize the manifold of natural images, allowing the researchers to generate natural stimulus changes and define the "natural tangent space" (NTS).
  • Information Geometry: This mathematical framework was employed to analyze the Fisher information matrix, which quantifies the discriminability of stimuli based on neural responses. This allowed them to calculate the "most discriminable input" (MDI) and "most discriminative response" (MDR).
  • Stochastic Modeling: The model explicitly incorporated Gaussian noise in early layers (simulating photoreceptor and bipolar cell noise) and independent binomial noise for ganglion cell spiking to capture neural variability.

Datasets

  • Natural Image Database: The study utilized a natural image database (Ref. 18) for training the CNN and generating stimuli. This is a standard dataset for visual neuroscience research.

Key Insights and Ideas

  • Retina as a Motion Detector: The research strongly suggests that a primary function of the retina's dynamic adaptation to natural scenes is the enhanced discrimination of local motion. This goes beyond simply encoding static image statistics.
  • Noise Correlations are Detrimental, Not Beneficial: Contrary to some hypotheses, the study found that correlated neural noise reduces information transmission in the retina for natural scenes. This suggests that these correlations might be an emergent property of shared neural circuitry rather than an adaptive mechanism for optimization.
  • Dynamic Receptive Fields are Key: The adaptation of the retina to natural scenes is driven by the dynamic changes in its receptive fields, not by the presence of correlated noise.
  • Nonlinearities Enhance Motion Discrimination: The complex nonlinearities within the retina are crucial for enhancing the discriminability of local motion while suppressing global motion.

Key Results

  • The retinal neural code was found to be optimized to discriminate changes within the natural image manifold. This was demonstrated by showing that the most discriminable stimulus changes (MDIs) projected onto the natural tangent space changed little.
  • Discriminability of natural stimulus changes was significantly higher than for random changes in a space of the same dimensionality.
  • The authors found that nonlinearities in the retina enhance discriminability for local motion but suppress it for global motion. This was visualized through the analysis of enhanced and suppressed eigenvectors of the Fisher Information Difference matrix.
  • Removing noise correlations in their model significantly increased discriminability compared to the model with correlated noise, for all population sizes tested.

Limitations and Future Work

  • Infinitesimal Stimulus Changes: The current analysis focuses on infinitesimal stimulus changes. The authors note that predicting discrimination across larger stimulus changes will require analyzing geodesics under the information metric.
  • Generalizability to Other Systems: While the approach is presented as generalizable, further work is needed to confirm similar insights in other sensory systems.
  • Role of Noise Correlations in Specific Conditions: The paper acknowledges that while noise correlations limit information in their ensemble analysis, they might still increase information under specific stimulus conditions or for certain cell types. Future work could explore these niche roles.
  • Behavioral Link: While the study focuses on retinal processing, connecting these findings directly to specific visual behaviors would be a valuable next step.

Summary updated

Abstract

Sensory systems discriminate between stimuli to direct behavioral choices, a process governed by two distinct properties - neural sensitivity to specific stimuli, and neural variability that importantly includes correlations between neurons. Two questions that have received extensive investigation and debate are whether visual systems are optimized for natural scenes, and whether correlated neural variability contributes to this optimization. However, the lack of sufficient computational models has made these questions inaccessible in the context of the normal function of the visual system, which is to discriminate between natural stimuli. Here we take a direct approach to analyze discriminability under natural scenes for a population of salamander retinal ganglion cells using a model of the retinal neural code that captures both sensitivity and variability. Using methods of information geometry and generative machine learning, we analyzed the manifolds of natural stimuli and neural responses, finding that discriminability in the ganglion cell population adapts to enhance information transmission about natural scenes, in particular about localized motion. Contrary to previous proposals, correlated noise reduces information transmission and arises simply as a natural consequence of the shared circuitry that generates changing spatiotemporal visual sensitivity. These results address a long-standing debate as to the role of retinal correlations in the encoding of natural stimuli and reveal how the highly nonlinear receptive fields of the retina adapt dynamically to increase information transmission under natural scenes by performing the important ethological function of local motion discrimination.
BibTeX
@misc{ding2024adaptation,
  author = {Xuehao Ding and Dongsoo Lee and Joshua B. Melander and Surya Ganguli and Stephen A. Baccus},
  title = {{Adaptation of retinal discriminability to natural scenes}},
  year = {2024},
  doi = {10.1101/2024.09.26.615305},
  howpublished = {bioRxiv},
}
Figure