Linking visual perception and V2 physiology by crowdsourcing psychophysics

C M Ziemba, J Freeman, E P Simoncelli and J A Movshon

Published in Annual Meeting, Neuroscience, Oct 2012.
The cortical areas of the primate ventral visual pathway encode increasingly complex stimulus attributes. In previous work, we showed that cells in the second visual area (V2) respond more strongly to "naturalistic" images synthesized using a model that captures the statistical structure of natural textures, compared to ßpectral noise"images matched for orientation and spatial frequency content. For reasons that remain obscure, naturalistic images synthesized from some textures evoked stronger differential responses and were more perceptually distinct than others. These findings were corroborated using human fMRI. But these relationships between perception and physiology, like many in sensory neuroscience, were based on a small ensemble of stimuli and a small number of human observers. Here, we use internet-based psychophysical experiments to greatly extend and validate these physiological-behavioral relationships.

We developed a three-choice oddity task requiring minimal instruction and suitable for presentation in a web browser, and obtained data from hundreds of participants on hundreds of different images using the "Mechanical Turk"service provided by amazon.com. Stimuli were derived from images of natural texture from commercial and personal photographs. For each image, we used our model to interpolate between the statistics of the natural texture and the statistics of a matched spectral noise image, yielding an axis of "naturalness". We measured discrimination performance along that axis, collecting a small number of trials from each of many subjects. We developed an expectation-maximization procedure for inferring the reliability of each individual subject and iteratively reweighted their contributions to estimate a single psychometric function.

On a subset of images, we confirmed that perceptual sensitivities obtained from internet observers were highly correlated with, but slightly lower than, sensitivities measured in a controlled laboratory setting. There was substantial variability in sensitivity across images, extending both above and below that measured in our original laboratory-based experiment. We selected images at the upper and lower ends of that distribution and used them in an fMRI experiment to confirm the previously identified relationship between physiology and perception. Our results validate the internet as a reliable tool for mid-level visual psychophysics, and highlight its unique capabilities for generalizing relationships between behavior and neural representation.

  • Listing of all publications