TL;DR
Contrary to the "efficient coding" theory, which predicts retinal ganglion cells should decorrelate natural scenes, this study shows that natural stimuli, particularly those with high spatial contrast during gaze shifts, actually drive correlated activity in these cells. This concerted population response, arising from nonlinear processing within the cells' receptive fields, appears to be crucial for detecting important visual features.
What Problem Does This Paper Address?
The prevailing theory of "efficient coding" in vision suggests that the retina's primary role is to remove redundancies in natural scenes, leading to decorrelated activity in its output neurons, the retinal ganglion cells. This theory has been a cornerstone for understanding retinal structure and function. However, this prediction has largely been tested with artificial stimuli, not the dynamic, gaze-driven natural visual input we experience daily. A key question remained: does the retina actually discard these natural scene correlations, especially during the rapid eye movements (gaze shifts) that are ubiquitous in natural vision? This paper investigates whether natural stimuli, specifically those encountered during gaze shifts, lead to the expected decorrelation or, surprisingly, to correlated activity in retinal ganglion cells.
Methods
The researchers employed a combination of advanced electrophysiology and computational modeling. They recorded from retinal ganglion cells in both marmoset and mouse retinas using multielectrode arrays. To mimic natural vision, they generated custom "natural movies" by combining natural images with recorded eye movement traces from head-fixed animals.
Key to their analysis was the development and application of a novel "subunit grid model." This model moves beyond traditional linear-nonlinear (LN) models by proposing that a ganglion cell's receptive field is composed of multiple smaller, nonlinearly interacting subunits. This framework allowed them to investigate how nonlinear processing within these subunits contributes to population-level responses and correlations. They also used standard LN models and a difference-of-Gaussians (DoG) model for comparison.
Datasets
The study utilized natural images from the Van Hateren database. For marmoset experiments, they used eye movement data recorded from head-fixed marmosets viewing natural images. For mouse experiments, they used horizontal gaze traces from freely moving mice. These datasets were combined to create species-specific natural movies.
Key Insights and Ideas
- Efficient Coding Violation: The study directly challenges the efficient coding prediction of decorrelation. Instead, they found that natural stimuli, especially those with high spatial contrast during gaze shifts, induce correlated activity.
- Nonlinear Processing is Key: The correlated responses are not explained by simple linear models. Instead, they arise from complex nonlinear interactions within the ganglion cells' receptive fields, specifically through the pooling of inputs from multiple subunits.
- Gaze Shifts Matter: The dynamic nature of natural vision, particularly the rapid shifts caused by eye movements, plays a critical role in driving these concerted responses.
- Feature Detection Hypothesis: The authors propose that this concerted, correlated activity might be a mechanism for robustly signaling the detection of important visual features (like edges or motion) to downstream brain areas, even if it means sacrificing some degree of efficient coding.
Key Results
- Natural movies, especially those with high spatial contrast during fixations, drove strong and correlated spiking responses in both marmoset and mouse retinal ganglion cells, both within and across different cell types.
- These correlations were significantly stronger than predicted by standard linear-nonlinear (LN) models, highlighting the importance of nonlinear processing.
- The novel subunit grid model, which captures nonlinear subunit interactions, provided a much better account of these correlated responses than LN models.
- The degree of nonlinear processing within a cell type correlated with its tendency to exhibit less stimulus decorrelation. For example, parasol cells, known for their strong nonlinearities, showed more correlated activity.
Limitations and Future Work
- Cell Type Specificity: While the study identified general trends, the degree of correlation and the influence of nonlinear processing varied across different retinal ganglion cell types. Future work could delve deeper into these cell-type-specific mechanisms.
- Model Complexity: While the subunit grid model offered significant improvements, fitting and interpreting these complex models remains a challenge.
- Feature Detection Link: The hypothesis that correlated activity aids in feature detection is compelling but requires further experimental validation. Future research could explore how these correlated signals are used by downstream brain areas for specific perceptual tasks.
- Generalizability: While the study included both marmoset and mouse, further investigation across a wider range of species and visual conditions could solidify these findings.
Summary updated