It is argued that the kind of adaptation that takes place in biologica neural networks must mainly be of the unsupervised type, on the lowest circuit levels at least, whereas supervised training may not be applied in biology until on a relatively high (behavioral) level, where the learning subject is already aware of the consequences of, say, reward and punishment. Therefore, the straightforward corrective learning strategies (e.g., the "delta rule" of the "backpropagation of errors") that have been used to tune artificial neural networks can hardly serve as any model for the true biological phenomena on the network level. This paper concentrates on some basic types of unsupervised learning, of which the so-called self-organizing map is believed to quantitatively reflect similar processes as those taking place in the brain during the developmentary phases and even later, under readjustment of the subject to the environment. In particular, a mechanism that is able to form geometrically organized two- dimensional maps for semantic data, similar to the categorical organizations experimentally found in the brain areas, is demonstrated.