As an example of the concepts discussed in Section 2, in this section a neural network is presented which solves the task of depth perception from stereoscopic images.
Depth perception by animals through stereo vision requires overlapping, binocular visual fields. With both eyes looking at the same part of a scene, but from two slightly different perspectives, objects in the left retinal image will be shifted relative to their position in the right retinal image. This shift will depend on the distance of the objects from the observer. If one fixates a point at infinity, the amount of shift turns out to be inversely proportional to the distance of objects. Objects at infinity will not be shifted, while nearer objects will have larger horizontal shifts. Estimating these shifts (disparities) between the two stereo views makes it possible to recover the three-dimensional structure of the scene.
The network developed in this section performs the task of disparity estimation by a combination of rate- and spike-coding neurons. It consists of two parts: (1) layers of simple and complex cells, calculating raw disparity estimates out of the input images; and (2) a coherence detecting network, combining the raw, but noisy estimates from the initial network layers into a final, stable disparity map. The layers of simple and complex cells are modeled by rate-coding neurons, but the coherence-network needs exact spike timing to function properly.