Page 412 - Introduction to Information Optics

7.6. Pattern Classification 39'

Fig. 7.36. The AT-maxnet.

The threshold 0(t) is controlled by the number of nonzero output nodes.
Underthresholding would give rise to multiple nonzero nodes, while over-
thresholding would reject all the nodes. Therefore, the maximum hidden node
can be extracted by simply adjusting the threshold value 6(t) in each iteration.
Let us denote 6 min and $ max to be the rejection and the maximum threshold
levels, respectively. By denoting d(t) the threshold level at the t th iteration, and
k the number of nonzero output nodes, as given by

(7.43)

the adaptive thresholding strategy can be formed as follows:

1. Set 0(0.) = e m{n.
(a) Terminate the process if k = 0, which means the input pattern does
not belong to any of the stored exemplars.
(b) Terminate the process if k = 1, which means that the maximum node
has been found.
(c) Set t = 1 and go to step 2 if k > 1.
2. Set

then the threshold value for the fth iteration is
6(t - 1) + <50(r), k > 1,
0(r-l)-<50(f), fe = 0