Biomedical Engineering Reference
In-Depth Information
the masks that code each of the parameters (name of the parameters), and then to
make the conjunction with the mask that codes the numerical value of this parame-
ter. The conjunctive terms of the parameters obtained are united disjunctively for
the parameter set, forming the nonnormalized code of the set of numerical para-
meters. Examples of such coding will be given in the description of concrete tasks.
A quantity of 1s in the coding masks must be selected in such a way that the total
quantity of ones in the nonnormalized code of the parameter set exceeds the sizes of
the ensembles, since all operations of normalization assume a decrease in the total
quantity of 1s in the codes.
5.2.1.2 Code normalization
Codes are normalized in order to obtain the necessary sizes of the ensembles
formed in the associative fields of higher levels. We will examine the normalization
procedures of the vectors that code the feature sets.
Normalization procedures have to meet several requirements. One basic require-
ment consists of equal representation of any feature in the normalized code. This
means that the normalized code has to contain approximately equal numbers of “1”
drawn from the codes of the features that were combined. The second requirement
states that during normalization of the codes for different feature sets, the repre-
sentatives of the same feature must be different, i.e., the same representatives must
not always be selected. The third requirement is that during repeated codings of the
same feature set, the normalization must preserve the same representatives of each
feature.
Let
X
be the binary vector subject to normalization and
m
be the quantity of unit
elements that must remain in the vector after normalization. Let us select the set of
the entire random numbers
s
[
1
],
s
[
2
],
,
s
[
k
], uniformly distributed in the interval
(0,
n
), where
n
is a quantity of neurons in the buffer field, and let us treat this set as
the characteristic of the concrete buffer field in which the normalization procedure
is produced. Let us denote as
Y
(
s
) the vector
Y
cyclically shifted on
s
digits. Then
the normalization procedure will appear as follows:
1. Form the working vector
Y
=
X
, where
...
X
is the bitwise inversion of vector
X
,
and
i
=1.
2. Perform the operation
X
¼
X
&
Y
ð
s
½
i
Þ:
(5.10)
3. Check the condition
S
ð
X
Þ <
m
;
(5.11)
where
S
(
X
) is a quantity of unit elements in the vector
X
.
