Published in Computational and Systems Neuroscience (CoSyNe), Mar 2018.
A converging view of peripheral vision holds that the brain represents statistical summaries of image content in local regions of the visual field. The resulting loss of information can explain the phenomenon of "crowding", in which recognition of peripheral objects is impaired by surrounding distractors (Balas et.al., 2009; Freeman & Simoncelli, 2011). Here, we show that such statistical representation can either help or hinder visual discrimination, depending on the observer's task. We created samples of synthetic texture constrained to match a set of higher-order statistics obtained from natural photographs, and measured the ability of observers to discriminate these stimuli within apertures of different sizes. Observers performed two tasks: category discrimination between images with different statistics, and sample discrimination between different images with matching statistics. For both tasks, performance of an ideal observer should improve as stimulus size (and thus stimulus information) increases. In contrast, humans became better at category discrimination but worse at sample discrimination. This occurred regardless of whether stimuli were presented simultaneously at different spatial locations, or in separate temporal intervals at the same location. We found that these opposing effects are predicted by a decision model operating on noisy higher-order statistics that are computed within localized regions that grow with eccentricity at roughly the scale of V2 receptive fields. This model predicts discriminability should scale in a lawful way with eccentricity, similar to the scaling of perceptual crowding, even though the contents of our stimuli are entirely task-relevant, and do not include distractors or clutter. We experimentally verified this prediction, further solidifying the relationship to crowding. These results are consistent with analogous effects in discrimination of auditory textures as a function of temporal window duration (McDermott et.al. 2013), suggesting they may be a general consequence of cascaded sensory transformations.