Abstract
Psychophysical reverse correlation (PRC) is a versatile, data-driven method that can be used to study a wide array of sensory, perceptual, and cognitive tasks. In visual PRC experiments, subjects typically perform discriminations on a base image with white noise overlaid. After many trials, noise frames from correct and incorrect trials are combined systematically to generate classification and significance images (CIs, SIs) which highlight locations and features critical for performing a perceptual task. However, the large (5-10k) number of trials required to generate quality CIs/SIs greatly limits the practicality of PRC. We explored the possibility of improving PRC efficiency by optimizing the noise used in a series of simulations. In these experiments, a simulated observer detected the orientation of angled Gabors with a variety of different noise types overlaid including white noise, sinusoidal noise, and Gaussian blob noise. We also systematically varied trial number (1-10k) and the spatial frequency of the base image Gabor (2-18 Hz) to fully explore the optimization space. For each simulation, CI quality was assessed by its correlation with the base image, correlation with a reference CI (generated via 20,000 trials), and the number of significant pixels. PRC efficiency was assessed as the number of trials required for the resulting CIs to meet a given correlation value or number of significant pixels. Our results suggest that noise selection greatly impacts PRC efficiency, and that the optimal noise choice depends on specific features of the base image used. Highest efficiency seems to be attained when the feature profile of the noise most closely matches that of the base image. We believe that similar, tailored simulation methods could be used to improve PRC experiment efficiency by informing and optimizing the noise used prior to collecting human subject data.