Abstract
Conventional neurophysiological and psychophysical studies focus on characterizing responses to restricted sets of artificial stimuli presented hundreds of times. These studies are well designed for quantifying internal variability (e.g. neural noise), but they reveal little about how stimulus variation affects responses in natural viewing. Natural stimuli have variable and uncertain spatiotemporal structure. Here, we examine the role of natural stimulus variability on neural representations of retinal image speed by recording MT neurons in fixating macaques to a large set of naturalistic image movies. Recently, we developed an ideal observer for speed estimation with these naturalistic movies (Burge and Geisler, in prep.). For a given speed, the ideal observer over- and under-estimates speed due to variation in the spatial features of each movie. Spike counts and speed-tuning curves were obtained for individual natural movies, and across the set of movies. (Similar measures were obtained for traditional sine-wave stimuli at different spatial frequencies). At a given speed, spike count variances for individual movies accounted for ~85% of the total variance across all movies, indicating significant image-specific variability beyond conventional internal neural noise. These results indicate that speed estimation in natural vision is limited both by internal noise and by retinal image variation due to differences in the projected images of objects in the environment. Analyzing responses to natural stimuli, and quantifying the relative contribution of internal and external variability, is required for understanding the visual computations that occur in critical sensory-perceptual tasks under natural conditions.