Beruflich Dokumente
Kultur Dokumente
1
experience. The subsequent parts of the
paper elucidate this fact in more details.
2
supervised manner. A character is for all i=1 to x
presented to the system and is assigned a {
particular label. Several variant patterns for all j=1 to y
of the same character are taught to the {
network under the same label. Hence the Wk (i, j ) = Wk (i, j ) + M (i, j )
network learns various possible }
variations of a single pattern and } (1.2)
becomes adaptive in nature. During the
training process, the input to the neural Here x and y are the dimensions of the
network is the input matrix M defined as matrix Wk (and M).
follows:
The following figure shows the
If I (i, j ) = 1 Then M (i, j ) = 1 digitization of three input patterns
Else: representing S that are presented to the
If I (i, j ) = 0 Then M (i, j ) = −1 (1.1) system for it to learn.
3
1 3 3 3 3 1 [4.] NETWORK ARCHITECTURE:
3 3 −3 −3 −1 −1
The overall architecture of the
3 −1 −3 −3 −3 −3 recognition system is shown in Fig. (4).
3 3 1 −1 −1 −1 In this system, the candidate pattern I is
WS = the input. The block ‘M’ provides the
−1 3 3 3 3 3
input matrix M to the weight blocks Wk
−3 −3 −3 −3 −3 3 for each k. There are totally n weight-
3 −3 −3 −1 1 3 blocks for the totally n characters to be
3 3 3 3 3 1 taught (or already taught) to the system.
Fig. (3)
4
(4.2) Ideal Weight-Model Score ( µ ): • Conclude that the candidate
This statistic simply gives the sum total pattern does not exist
of all the positive elements of the weight within the knowledge base
matrix of a learnt pattern. It may be OR
formulated as follows (with • Teach the candidate pattern to
µ (k ) initialized to 0 each time). the network till a satisfactory
value of Q(k ) is obtained.
for i=1 to x 4. Conditionally, identify the input
{ candidate pattern as being akin to the
for j=1 to y kth learnt pattern OR proceed with
{ the training for better performance.
if Wk (i, j ) > 0 then
{ In Fig. (4), the selector gives an output k
by making the best selection as in Step 4
µ (k ) = µ (k ) + Wk (i, j )
of the aforementioned algorithm. The
} adaptive performance of the network can
} easily be tested by an example: we
} (1.4) submit two hand-drawn patterns
representing S and P respectively to the
(4.3) Recognition Quotient (Q): This system that has already learnt only the
statistic gives a measure of how well the character S. The recognition quotient
recognition system identifies an input yielded by the trained system is
pattern as a matching candidate for one mentioned alongside.
of its many learnt patterns. It is simply
given by:
ψ (k )
Q(k ) = (1.5)
µ (k )
5
from prior experience in neural The dimensions of the input matrix need
networks. to be adjusted for performance. Greater
the dimensions, higher the resolution and
better the recognition. This however
increases the time-complexity of the
system which can be a sensitive issue
with slower computers. Typically,
32X32 matrices have been empirically
found sufficient for the recognition of
English handwritten characters. For
Fig. (6) intricate scripts, greater resolution of the
matrices is required.
The system however dismisses the
candidature of the pattern representing P As already illustrated in the previous
in Fig. (5) by a yielding a low value example, efficient supervised teaching is
of Q(= 0.21) . It can be observed by essential for the proper performance.
regular teaching, that the system Neural expert systems are therefore
develops on its ability to identify a typically used where human-centered
matching pattern and reject non- training is preferred against rigid and
matching patterns. Thus, regular inflexible system-rules.
supervised teaching marks enhanced
performance of the system
[6.] CONCLUSION:
6
same program. It is available for
download at the following web address:
http://neil.fraser.name/software/recog
[8.] REFERENCES: