Published in Annual Meeting, Neuroscience, Nov 2016.
A converging view of peripheral vision holds that the brain represents statistical summaries of visual content in local portions of the visual field, resulting in a loss of information that can explain the phenomenon of "crowding". Here, we show that such statistical representation can either help or hinder visual discrimination performance depending on the observer's task. We created synthetic texture stimuli by matching a set of higher-order statistics measured from natural photographs (Portilla & Simoncelli, 2000). Observers were asked to discriminate these stimuli windowed by apertures of different sizes and at increasing eccentricities. When observers compared images with different statistics, performance increased with increasing patch diameter. This is expected, since the parameters of the model converge to different values as the patch size increases. Interestingly, when observers compared different images with identical statistics, performance decreased as patch diameter increased. Specifically, as the statistics converge to their target values with increasing patch diameter, subjects were no longer able to utilize the local cues that enable high performance with small patch sizes. We found these opposing effects of patch size regardless of whether subjects discriminated simultaneously presented stimuli across space or sequentially presented stimuli across temporal intervals.
These results are consistent with analogous effects observed for discrimination of auditory textures as a function of temporal window duration (McDermott & Simoncelli, 2013), and suggest a general processing strategy for sensory systems. Curiously, there was little effect of eccentricity on the pattern of discrimination performance, in contrast to the well known eccentricity-dependence of crowding. We show that both the opposing effects of stimulus size, and the relatively minor effect of eccentricity, are predicted by a model for the visual periphery in which higher-order statistics are computed within pooling regions that grow with eccentricity at the scale of V2 receptive fields (Freeman & Simoncelli, 2011). At increasing eccentricities, the decrease in the number of pooling regions responding to the stimulus is partially counteracted by the stability of statistical summaries averaged over larger regions, resulting in a relatively modest effect of eccentricity on performance. Thus, we extend previous observations on peripheral vision by demonstrating that a simple model computing local statistical summaries can capture a complex pattern of human discrimination performance across different image sizes and eccentricities.