The process of feature extraction is influenced by selectivity to a great extent. For each S-layer in the neocognitron we can set different amount of selectivity at construction of the network.
By the change of selectivity we change the effect of inhibitory part on the S-cell output value. Decreasing of selectivity causes decreasing of effect of inhibition part. Decreased S-cell ability to distinguish learned feature exactly is the result of it. In other words it means that S-cell considers also more deformed features to be correct.
Example represents one S-cell from layer US1 and its connection area which is its receptive field here as well. This S-cell extracts feature corresponding to vertical line in the centre of the receptive field.
Select one of the prepared patterns with using of mouse, set up desired amount of selectivity and observe the effect on excitatory part E, inhibitory part I and S-cell output value as well.
We reminder that S-cell output value is exactly determined by equation described in mathematical description of its behaviour but simplified form of this equation will suffice for us :
From this example it is obvious that excitatory part E is influenced only by cells which correspond to the mask marked by gray color. Inhibitory part I is influenced by all cells from connection area and by amount of selectivity as well. S-cell becomes inactive if inhibitory part is greater or equal to excitatory part.