Abstract
Visual systems integrate signals across space. Spatial integration operates at virtually all steps of visual information processing from retinae to higher cortical visual areas. Spatial integration is useful computationally because it can improve the precision of estimates by averaging out noise when local spatial variability is present. Classic studies model spatial integration with a fixed linear weighting of local inputs, the optimal strategy when variation is solely due to noise. In natural scenes, local variability can be caused by noise or signal variability, so a more subtle strategy may be required. Here, we report that human spatial integration deviates substantially from the predictions of the classic model. Human observers judged the average luminance or the average stereoscopic depth of spatially variable test stimuli relative to a surrounding surface. In both tasks, the test stimulus consisted of nine horizontal bars (60×6 arcmin each) vertically flanking the fovea. The luminance or disparity across bars was either uniform or independently sampled from normal distributions. For each sample variance (calculated across all nine bars), we varied the sample mean to measure a psychometric function. We separately analyzed human performance with stimuli having zero, small, medium, and large sample variance. The classic model linearly weights luminance or disparity across the stimulus; the sample mean is the optimal decision variable. Thus, identical performance is predicted for all sample variances. However, human thresholds increase markedly with sample variance (3 and 5 times in two tasks, respectively). This result indicates that spatial integration depends on the local variability across each stimulus. The classic linear model for spatial integration should be modified to account for these results. The modification likely involves variability-dependent changes to the gain (e.g. normalization) or pattern of the integration weights. These modifications may optimize performance in natural scenes.
Acknowledgement: NIH R01-EY028571