High-speed Learning in a Supervised, Self-growing Net.

A supervised, self-growing neural net classifier is presented, in which categories are formed for known classes of patterns. The level of supervision is minimal. The net is unsupervised and self-organising except when an erroneous classification is detected, whereupon a new unit is recruited. The number of units on a recognition layer is strictly limited, but the net is allowed to grow in a tree-like way by issuing recognition sublayers from the main recognition layer. Thus, a sufficient number of categories are formed to correctly classify all of the training set patterns.