2- BACK PROPAGATION nodes varying from 70 subgroups listed in Table 9,
IMPLEMENTATION
to 200 nodes. From the the first pair of group no. 1
To prove the feasibility of table, it is found that 90 (words of 2, 4, 8, 9, 10) with
the selected features, the hidden nodes give the 88 %, and group no. 2 (0,
BP algorithm is applied to best
recognition
rate 1, 3, 5, 6, 7) with 70 % are
test the new features. The (78.18 %). It is also seen developed. The second
network included 45 input that the 135 nodes give pair group is the even (with
nodes, one output node, a high percentage (74.55 86.67 %) and the odd (with
and
varying
number %) but less than this of 90 72 %) groups. The last pair
hidden nodes. If the nodes. It is noticed that is the first five words (0: 4)
number of hidden nodes 90, and 135 are a multiple gives 76 %, and the other
less than that of the input of 45 which is the number six words (5: 10) gives 86.67
nodes, the network will not of the input nodes, and %. The control network was
converge. With the same at 180 hidden nodes it the same BP classifier with
as input nodes (45 nodes), gives a rate of (69.09 %). varying hidden nodes.
the network gives after It can be concluded that For the first pair group the
3000 trials a sum square the percentage rate is classifier is tested with 10 to
error (SSE) of 30.1457 but raising at each multiple 100 hidden nodes where
for 60 hidden nodes, the of the number of the the results are given in Fig.
network gives 10.772 SSE, input nodes than those 8. It is seen that the highest
So the tests with number of the other numbers recognition rate of the INN
of hidden nodes greater in its neighbor (Fig. 7). is that of the classifier of
than 60 nodes would be
10, and 40 hidden nodes
(74.5 %), which is less
than that obtained from
the BP network. For the
second pair, the results
(Fig. 9) proves that it fails
to exceed 70 %. For the
last pair (Fig. 10), it is found
that the INN with the 50
hidden
nodes
control
network gives the best
recognition rate 80 %. This
rate is higher than that
needed. The results are 3- INN SUBGROUPS
reached in the work with
listed in Table 5 for hidden Training the network on the the LPC features.
Research
36

---