Efficient stereo coding in the primary visual cortex and its experimental tests by optical imaging and single cell data

Recoding sensory inputs to remove the input redundancy has been advocated as a sensory preprocessing goal(Barlow 1961). An overwhelming source of redundancy in visual inputs is the pair-wise image pixel correlation, and the binocular correlation in particular. It is redundant to transmit the correlated inputs from the two eyes independently to the cortex, and it has been proposed that V1 cells de-correlate the binocular inputs to improve coding efficiency (Li and Atick 1994). This means, the V1 cells recode or combine the retinal inputs from the two eyes such that the cell outputs would be less correlated while preserving input information. In particular, this leads to the following consequences. (1) when the inputs from the two eyes are almost identical, such as arising from horizontal bars which typically have near zero vertical disparities, they are summed into binocular cells rather than redundantly and separately coded by two monocular cells; (2) when the inputs from the two eyes are more different, such as arising from bars oriented vertically, monocular or ocularly unbalanced V1 cells will extract the difference between the two retinal inputs and thus the stereo information; (3) when the retinal inputs are very weak, e.g., when received by cells of smaller receptive fields or tuned to higher spatial frequencies, they are summed into binocular cells in order to increase signal-tonoise. This efficient code predicts that V1 cells tuned to horizontal orientation or to higher spatial frequencies are more likely binocular than cells preferring other orientations or spatial frequencies. To test these predictions, we examined the data obtained by optical imaging (Hubener et al 1997) and electrophysiological recordings (Anzai et al 1995) of the cat V1. By presenting visual stimulus of different orientations and spatial frequencies through different eyes, the preferred orientations, spatial frequencies, and ocular dominance were obtained for single cells, or the local cortical areas of the optical images. In the optical images from four cats, we examined the proportion of image pixels that are more binocular among all the pixels tuned to particular orientation or spatial frequency and confirmed both predictions. In the electrophysiological data from 136 V1 cells, we obtained distributions of ocular balance indices for cells tuned to different ranges of orientations. We confirmed the predicted association between horizontal orientation preference and binocularity. Our experimental results confirmed the theoretical predictions, thus supporting the hypothesis of efficient stereo coding as one of V1ís the role