Everything an organism can know about its visual environment is carried in a stream of photons that are absorbed stochastically by the two-dimensional array of photoreceptors. As a consequence, the initial representation of a visual scene in the photoreceptor array contains both deterministic and random components. This raises the question of how an organism decides whether particular spatiotemporal correlations within this initial representation of the scene are meaningful, or merely artifacts of randomness. Moreover, it must do this in real time without benefit of prior information about the particular scene. If it cannot do this reliably, the result will be ambiguity and confusion.

To address this question, Sean McCarthy and I defined the properties of an ideal physical device that makes a minimally uncertain estimate of the visual scene. By comparing the properties of neurons in the vertebrate retina with those of this device, we found that the retina generates a minimally uncertain representation of the visual scene and then creates a spatio-temporal map of that uncertainty. It is this map that is transmitted to the brain. We conclude that the retina functions as an ideal estimator and that the uncertainty in the estimate contains all of the information necessary for reliable perception.