Files
Abstract
All information available to the visual system of an organism comes through the retina. This biological sensor is constantly inundated with high-dimensional, statistically complex inputs. As it can only encode a limited amount of information, the organisms must prioritize measuring behaviorally-relevant aspects of their input stimulus. To understand what constitutes a “good” measurement, it is necessary to understand what bits of information are most important for an organism’s survival. Rather than asserting a particular normative notion of good measurements, we examine what information is encoded in the population response of the retina and infer what these bits are best-representing. We construct a one-parameter family of optimal measurements and infer which one best describes retinal spiking activity. Across a range of driving statistics, there is strong agreement between the population response and one of the optimal measurement strategies from our one-parameter family. Interestingly, the particular value of the one-parameter family that best describes the population response changes with input statistics. Knowing what information matters is not enough to constrain the response properties of the retina. For the goal of making good measurements of the stimulus to constrain response properties, there must be a small number of ways to make “good” measurements. If there are many approaches that work similarly well, the response properties will be unconstrained. Using natural scene statistics, we explore how constrained an optimal retina population would be. Analyzing the sensitivity of the optimal population to perturbation, we find that the strength of agreement between the optimal population and the observed spiking response depends on sensitivity. Features of the optimal population that are highly sensitive to the perturbation show strong agreement with the observed population response. Agreement is weaker for less sensitive features. We find a small but highly important set of features strongly aligned with an optimal population. These features are preserved across different natural scenes.