In this tutorial we deal only with the version of the neocognitron which uses learning with a teacher and therefore we will describe only this principle of learning here.
Learning in this version of the network is controlled by a teacher. His task is to determine what features shall be extracted in particular stages of the network and to prepare corresponding training patterns before beginning of learning.
Learning of the neocognitron proceeds stage by stage from the lowest stage of the network and it inheres in adjusting of modifiable weights (i.e. a-weights and b-weights) according to the response of already learned parts of the network to presented training patterns. For each S-plane in the network one training pattern is usually used and this pattern is usually necessary to present to the network only once.
On the beginning of learning teacher have to set all a-weights and b-weights in the network to zero. Then he selects S-plane from layer US1 and in this cell plane he selects one of cells, so-called seed cell. Presentation of training pattern given for this S-plane into the input layer U0 is the next step. Finally teacher adjusts weights of the seed cell according to the equations mentioned in mathematical description of learning. Since weight sharing is used in the neocognitron adjusting of weights of all the other S-cells in the cell plane occurs simultaneously. If more training patterns for the selected S-plane exist then they are presented subsequently and process repeats. In opposite case we move to learning of the next S-plane.
The learning process of the neocognitron is demonstrated on the following example in detail.