Reading a population code: a multi-scale neural model for representing binocular disparity.

Although binocular neurons in the primary visual cortex are sensitive to retinal disparity, their activity does not constitute an unambiguous disparity signal. A multi-spatial-scale neural model for disparity computation is developed to examine how population activity might be interpreted to overcome ambiguities at the single neuron level. The model incorporates a front end that encodes disparity by a family of complex cell-like energy units and a second stage that reads the population activity. Disparity is recovered by matching the population response to a set of canonical templates, derived from the mean response to white noise stimuli at a range of disparities. Model predictions are qualitatively consistent with a variety of psychophysical results in the literature, including the effects of spatial frequency on stereoacuity and bias in perceived depths, and the effect of standing disparity on increment thresholds. Model predictions are also consistent with data on qualitative appearance of complex stimuli, including depth averaging, transparency, and corrugation. The model also accounts for the non-linear interaction of disparities in compound grating stimuli. These results show that a template-match approach reduces ambiguities in individual and pooled neuronal responses, and allows for a broader range of percepts, consistent with psychophysics, than other models. Thus, the pattern of neural population activity across spatial scales is a better candidate for the neural correlate of depth perception than the activity of single neurons or the pooled activity of multiple neurons.