Discovering flexible codes for prediction across timescales in the retina
Efficient coding theory postulates that a sensory system maximizes information between its response and the input, yet it is unclear if a different measure of optimality that takes into account output function might give a better fit to neural data. The sensory processing delays in many systems suggest that the maximization of predictive information is a reasonable objective function for driving fast, effective downstream behavior. We introduce a one-parameter family of optimal encoding distributions based on how far out in time a population of retinal ganglion cells is optimized to predict future stimuli. Analyzing the population response to a moving bar stimulus with rich temporal correlation structure identifies which particular optimal encoding best describes the neural activity. This allows for the discovery of how far out in time the retina is predicting, instead of simply testing for optimality at one timescale. As stimulus statistics change, so too does the time scale of prediction that best matches the population response. Focusing on this optimal timescale, the neural code can be evaluated in terms of classic efficient coding theory, revealing that the code also shows a peak in how these predictive bits are allocated in the population response repertoire. The stimulus has a fully controlled set of temporal statistics, but is still complex enough to show behaviors like starts and stops, constant motion, and motion reversals. Its tractable statistical structure allows for an information theoretic account of computations like motion anticipation and the retina reversal response in terms of the maximization of predictive information.