Information from the external world goes through various transformatio
ns. The learning of the original neighbourhood relations of the world
using only the transformed information is examined in detail. An appro
ximate representation consists of a finite number of discretizing poin
ts and connections between neighbouring points. The goal here is to de
velop the theory of self-organizing approximate representations. Such
a self-organizing system may be considered as a generalization of the
Kohonen topographical map that we now equip with self-organizing neigb
ouring connections. For illustrative purposes an example is presented
for sensory fusion: the geometry of the 3D world is learned using the
outputs of two cameras.