Biomedical Engineering Reference
In-Depth Information
frequently will be incorrect, and then the procedure of training, which includes
three stages, begins.
During the first stage, training with the correct name is carried out. For this
purpose, the computer
MC
introduces the correct name of that texture into the
coding block
CU
1
, which corresponds to the current position of the window. The
coding device generates the mask of the correct name and transfers it into the buffer
field
BF
, where it is united with the code of texture features and is transferred to
create the corresponding ensemble in the associative field
AF
. The creation of the
ensemble is achieved due to a change in the synaptic weights according to the
formula:
;
W
ij
0
¼
W
ij
U
q
i
&
q
j
&
h
ij
(6.2)
where
W
ij
'
is the synaptic weight of the associative connection from neuron
j
to
neuron
i
after training;
W
ij
is the synaptic weight of this connection before training;
q
i
is the output of the neuron
i
;
q
j
is the output of the neuron
j
;
h
ij
is the binary ran-
dom value with an average value that equals the signal of reinforcement (
M
(
h
ij
)=
l
);
U is disjunction; and & is conjunction.
In the second stage, the ensemble with the incorrect name is partially destroyed.
To complete this stage, the
computer
gives the command to repeat calculation,
coding, and normalization of the texture features, after which it sends the name of
the texture that erroneously was indicated by the recognition system into the coding
block
CU
1
. The coding device finds the mask of this erroneous name and transfers it
into the buffer field
BF
, at which point the code of the feature and the mask of the
erroneous name are transferred to the associative field
AF
, where a change of the
synaptic weights is produced with the negative signal of reinforcement according to
the formula:
;
W
ij
¼
W
ij
&
q
i
&
q
j
&
h
ij
(6.3)
where the horizontal line above the parentheses is negation and the remaining
designations are the same as in formula (6.2), except that the mathematical expec-
tation of the binary random variable
h
ij
is equal to the absolute value of the signal of
reinforcement (
M
(
h
ij
)=
).
This operation is necessary because we need to decrease the total quantity of
synaptic connections from the code of the analyzed texture to the mask of the
incorrectly identified name in order to decrease the probability of this error in
the future. It is expedient to form the ensembles in such a way that the name of the
texture would serve as the nucleus. Then, while calculating associative field
activity, the excitation is moved to the nucleus, and the corresponding texture
name is identified easily. Partial destruction of an ensemble executed in the second
stage has side-line action. It partially destroys the nucleus, which corresponds to the
incorrectly identified name. Since this is undesirable, the third stage, resolution of
the nucleus, is carried out. For this purpose, the buffer field,
BF
, is cleaned and only
the mask of the incorrect name is transferred to it. Without the change, this mask
j
l
j
