Abstract
Sensory neurons have been hypothesized to efficiently encode signals from the natural environment subject to resource constraints. The predictions of this efficient coding hypothesis regarding the spatial filtering properties of the visual system have been found consistent with human perception, but they have not been compared directly with neural responses. Here, we analyze the information that retinal ganglion cells transmit to the brain about the spatial information in natural images subject to three resource constraints: the number of retinal ganglion cells, their total response variances, and their total synaptic strengths. We derive a model that optimizes the transmitted information and compare it directly with measurements of complete functional connectivity between cone photoreceptors and the four major types of ganglion cells in the primate retina, obtained at single-cell resolution. We find that the ganglion cell population exhibited 80% efficiency in transmitting spatial information relative to the model. Both the retina and the model exhibited high redundancy (~30%) among ganglion cells of the same cell type. A novel and unique prediction of efficient coding, the relationships between projection patterns of individual cones to all ganglion cells, was consistent with the observed projection patterns in the retina. These results indicate a high level of efficiency with near-optimal redundancy in visual signaling by the retina.
Introduction
The Efficient Coding Hypothesis posits that sensory systems have evolved to maximize the transmission of information about the natural environment while minimizing biological costs such as energy consumption and physical space. While this theory has successfully explained many aspects of human visual perception, testing it directly against the underlying neural circuitry has remained a significant challenge. Such a test requires both a rigorous mathematical framework for "optimal" coding under realistic constraints and high-resolution data on the functional connectivity of sensory neurons.
In this study, researchers bridged this gap by analyzing the spatial information transmitted by the primate retina. By comparing the functional connectivity between individual cone photoreceptors and retinal ganglion cells (RGCs) to a theoretically optimized model, they sought to determine just how "efficient" the biological retina truly is.
Methods and Modeling
The researchers developed a linear-Gaussian model to represent the transformation of signals from the cone mosaic to the RGC output. The model is defined by the equation $r = W(s + \nu) + \delta$, where $s$ is the signal from the cone mosaic, $W$ is the connectivity matrix (synaptic weights), and $\nu$ and $\delta$ represent input and output noise, respectively.
To find the theoretically optimal connectivity, the authors used Karush–Kuhn–Tucker (KKT) conditions to maximize the Mutual Information between the input image and the RGC response. Crucially, the optimization was subject to three biological constraints:
- Cell Count: The number of RGCs was fixed to match biological observations.
- Response Variance: This served as a proxy for the metabolic cost associated with neural firing.
- Synaptic Strength: The total magnitude of the weights in $W$ was constrained, representing the structural cost of maintaining synapses.
The theoretical model was compared against physiological data obtained from macaque monkey retinas. Using high-density multielectrode arrays, the researchers mapped the functional connectivity between hundreds of cones and the four major RGC types: ON-Parasol, OFF-Parasol, ON-Midget, and OFF-Midget cells.
Results
80% Efficiency and the Role of Redundancy
The analysis revealed that the biological retina is remarkably efficient, preserving approximately 60% of the information available in the cone lattice. Compared to the theoretical maximum of 75% for an optimized system under the same constraints, the retina achieves roughly 80% relative efficiency.
A significant finding was the presence of "redundancy." Traditional views of efficient coding often emphasize "redundancy reduction," suggesting neurons should be as independent as possible. However, this study demonstrates that in the presence of noise, an optimal system should maintain a degree of redundancy (~30%) to ensure robust signal transmission. The biological retina's redundancy levels closely matched those predicted by the optimal model.
Spatial Structure and Projective Fields
Despite the model having no built-in requirements for spatial locality or circularity, the "best-fit" optimal weights resulted in receptive fields (RFs) that mirrored biological RGCs: they were circularly symmetric with a classic center-surround organization.
The study also introduced the concept of the Projective Field (PF)—the pattern of divergence from a single cone to the entire RGC population. While many different RF configurations can be equally efficient, the PF provides a more unique signature of the underlying coding strategy. The researchers found that the PFs in the biological retina were highly consistent with the predictions of the efficient coding model.
Discussion and Future Directions
The high level of efficiency observed suggests that the retina's "wiring" is finely tuned to the statistics of natural images. Furthermore, the authors demonstrated that this efficient state could potentially be reached through simple, biologically plausible developmental rules based on local activity, rather than requiring a global blueprint.
However, several limitations remain. This study focused exclusively on spatial information in achromatic images. The retina is a complex spatiotemporal processor that also handles color (e.g., the S-cone pathway) and motion. Future research will need to incorporate these dimensions, as well as the non-linearities inherent in RGC responses, to fully characterize the limits of retinal efficiency. Expanding this framework to include the dozens of other specialized RGC types will be a key step toward a complete understanding of the first stages of visual processing.
Summary updated