Abstract
Visual processing in any natural task is limited by at least three sources of uncertainty: natural stimulus variability, early noise, and late noise. Natural stimulus variability is due to the task-relevant statistical structure of natural scenes, and is irreducible. Early (input) noise is due to factors prior to nonlinearities in the visual system (e.g. photon noise), and is also irreducible. These sources of uncertainty are 'external noise'. 'Internal' (or late) noise is due factors after nonlinearities in visual processing. Here, in three human observers, we use a suite of tools to predict the impact of each source of uncertainty on visual processing in the task of speed estimation from natural image movies. First, we measured the early noise in a target detection task using the equivalent input noise paradigm. Second, using an ideal observer for speed estimation in natural image movies, we determined the combined impact of natural stimulus variability and early noise on performance in a speed estimation experiment; every presented movie was unique. Third, we measured human performance with matched natural stimuli, and compared human and ideal performance. The ideal observer accounts for ~95% of the variance in the human responses across all conditions with a single free parameter (efficiency). Fourth, each human observer repeated the experiment with the exact same set of movies (double pass paradigm), and we determined the proportion of times that responses agreed. Response agreement is determined by the ratio of internal to external noise. Under the hypothesis that human inefficiency is due only to internal late noise, efficiency should tightly predict response agreement (zero additional free parameters) in the double pass experiment; maximum likelihood fits strongly confirm the prediction. Thus, an analysis of task-relevant natural stimulus variability, and strongly constrained noise models fully account for human performance in a speed estimation task.
Meeting abstract presented at VSS 2017