 |
| Fig. 9.1 - Connection areas of the S-cell |
Each S-cell in the neocognitron evaluates outputs of C-cells (or receptor cells) from the certain connection areas from previous C-layer (or input layer).
Size of connection areas is the same for all S-cells in one S-layer and it is determined at construction of the network. One S-cell connection areas of size 3 by 3 cells are shown in figure 9.1.
|
| Fig. 9.1 - Connection areas of the S-cell |
Function of each S-cell is to extract the certain feature at the certain position in the input layer (i.e. in its receptive field). For extraction of this feature an S-cell uses only informations obtained from its connection areas and information about average activity in these areas obtained from corresponding V-cell. All S-cells in one S-plane always extract the same feature.
The feature extracted by S-cell is determined by weights for this cell. Weights and method for their adjusting will be described later in detail. For a better conception about function of weights we can compare them to a mask which is used at determination about feature presence for now.
The meaning of weights is obvious best for cells from layer US1. Each S-cell in this layer has only one connection area and this area is S-cell's receptive field at the same time. So weights (mask if you want) contain directly representation of the certain feature. In higher S-layers correspondency between extracted feature and its representation by the weights is already not so obvious.
Cell plane of S-cells designated for extraction of feature corresponding to vertical line is shown in figure 9.2. S-cell is activated only if this feature is present in S-cell's receptive field (it is identical with connection area here). When incorrect feature is presented the cell becomes inactive.
 |
| Fig. 9.2 - S-cell function |
S-cell output value is determined exactly by the equation described in mathematical description. However, for understanding of S-cell function simplified equation is sufficient :
The symbols used in this equation have the following meaning :
| Symbol | Denotes |
|---|---|
| us | S-cell output value |
 | non-linear function |
| E | excitatory part |
| a | a-weights |
| uc | output values of C-cells from connection areas |
| I | inhibitory part |
| r | selectivity |
| b | b-weight |
| uv | V-cell output value |
The S-cells ability to extract not only learned features but also deformed representations of these features is influenced by the choice of parameter denoted as selectivity to a great extent. On the following example we will demonstrate how selectivity affects the process of feature extraction in the neocognitron.