Predicting the partition of behavioral variability in speed perception with naturalistic stimuli

A core goal of visual neuroscience is to predict human perceptual performance from natural signals. Performance in any natural task can be impacted by at least three sources of uncertainty: stimulus variability, internal noise, and sub-optimal computations. Determining the relative importance of these factors has been a focus of interest for decades, but most successes have been achieved with simple tasks and simple stimuli. Drawing quantitative links directly from natural signals to perceptual performance has proven a substantial challenge. Here, we develop an image-computable (pixels in, estimates out) Bayesian ideal observer that makes optimal use of the statistics relating image movies to speed. The optimal computations bear striking resemblance to descriptive models proposed to account for neural activity in area MT. We develop a model based on the ideal, stimulate it with naturalistic signals, predict the behavioral signatures of each performance-limiting factor, and test the predictions in an interlocking series of speed discrimination experiments. The critical experiment collects human responses to repeated presentations of each unique image movie. The model, highly constrained by the earlier experiments, tightly predicts human response consistency without free parameters. This result implies that human observers use near-optimal computations to estimate speed, and that human performance is near-exclusively limited by natural stimulus variability and internal noise. The results demonstrate that human performance can be predicted from a task-specific statistical analysis of naturalistic stimuli, show that image-computable ideal observer analysis can be generalized from simple to natural stimuli, and encourage similar analyses in other domains.