 |
| Fig. 13.1 - Weights in the neocognitron |
The neocognitron is characteristic not only by large number of cells but also by large number of connections. These connections serve for transfer of informations between cells in adjoining layers. Particular cell obtains by means of connections informations from all cells which are located in its connection areas.
For each connection there is a weight by means of it we can affect amount of transferred information. If we imagine a connection as a pipeline with a valve we can compare weight assigned to the connection to a degree of opening of this valve.
Four types of weights (a-weights, b-weights, c-weights and d-weights) exist in the neocognitron. Each of these types of weights is used for connections between two layers of different types. It is shown schematically in figure 13.1.
|
| Fig. 13.1 - Weights in the neocognitron |
Weight sharing is the next term being connected with weights. By this term we designate the fact that all cells in one cell plane use the same weights for connections leading from cells in their connection areas. By the means of weight sharing it is guaranteed that all cells from one cell plane always extract the same feature.
 |
| Fig. 13.2 - Weight sharing |
We can split the weights shown in figure 13.1 according to the way which they are adjusted :
The a-weights are the first type of weights modified by learning. These weights are used for connections between S-cells and C-cells which belong to their connection areas. Features extracted by S-cells are encoded in these a-weights. Adjusting of a-weights is performed during learning of the network according to the presented training patterns.
 |
| Fig. 13.3 - a-weights |
The b-weights are the second type of weights modified by learning. These weights are used for connections between S-cells and corresponding V-cells. Adjusting of b-weights is performed during learning of the network according to the presented training patterns as well.
 |
| Fig. 13.4 - b-weights |
Fixed c-weights are used for connections between V-cells and C-cells which belong to their connection areas. Values of c-weights are determined at construction of the network. These weights are most often set up in such a way that they mostly reduce transfer of information from the periphery of connection area and towards to the center of area the degree of reduction decreases.
 |
| Fig. 13.5 - c-weights |
Fixed d-weights are used for connections between C-cells and S-cells which belong to their connection areas. As well as c-weights also d-weights are determined at construction of the network and again in such a way so as to reduce transfer of information from periphery of connection areas mostly.
 |
| Fig. 13.6 - d-weights |