Combining complex cell responses as described in the previous section, one obtains neural units sensitive only to local disparity. However, as discussed, any single disparity unit on its own is rather useless, since there is no way to decide whether a coded value is a true disparity estimate or just a random reading.

Actually, one can show that *any conceivable* disparity estimator, not
only the ones described above, will fail to estimate disparity values correctly
most of the time [45]! In the worst-case scenario, the range of
disparity values any basic estimator can faithfully calculate is *less
than the receptor spacing* in the retina.

At this point, the need for a special process arises which is able to derive a valid and stable estimate from the noisy data of a pool of disparity units. As we have argued, any weighted linear combination of unit outputs will fail, since such an estimator would require that the number of correctly coding units is high compared to the number of units with invalid estimates.