Abstract
Binocular disparity, an important depth cue, is initially represented by computing the cross-correlation between the left-eye and right-eye images (correlation-based representation). This process takes place in the primary visual cortex (V1). In later processing stages within and beyond V1, this primitive representation is refined into a match-based representation, which reflects the strength of feature matching between the left-eye and right-eye images and discards responses to falsely matched features. Although other mechanisms must also be involved, a nonlinearity is a well-supported mechanism to transform the correlation-based representation into the match-based representation. Some physiological studies find that V1 responses have this nonlinearity. Also, simulations with the nonlinearity suggest that neurons with odd-symmetric tuning have more correlation-based response properties than those with even-symmetric tuning. To examine whether later processing stages beyond V1 have this nonlinear property, we analyzed neural responses to correlated and anti-correlated random dot stereograms (cRDSs and aRDSs) in macaque areas V2, V3 and V3A, which receive disparity signals from V1. The correlation-based representation should have inverted tuning functions for aRDSs relative to those for cRDSs; the match-based representation should lose disparity selectivity for aRDSs. We recorded neural responses from 54, 63 and 59 disparity selective neurons in V2, V3 and V3A while two monkeys performed a fixation task. Proportions of V2, V3 and V3A neurons with inverted tunings for aRDSs were close to 50%. In V2 and V3A, but not V3, neurons with odd-symmetric tuning tended to show inverted tuning curves for aRDSs more often than neurons with even-symmetric tuning. Thus, V2 and V3A were consistent with the simulation of binocular disparity computation with the nonlinearity. These findings suggest that responses in V2 and V3A, but not in V3, have the nonlinearity for transforming the correlation-based representation into the match-based representation.