Biomedical Engineering Reference
In-Depth Information
small segments. Since each large segment contains
k
small segments, and the whole
code contains
k
segments, we have
k
k
points of absorbtion.
Let us consider the coordinate
Y
= 1. In this case, all small segments are shifted
to the right with the absorption of unit elements in all small segments that fall on the
boundary of large segments. Since the small segment contains
k
single elements,
and the large segment contains
k
small segments, the number of absorbed points is
equal to
k
k
, which corresponds to the foregoing case with the coordinate
X
=1.
Thus, if the code is divided into equal fractions, i.e., when the small segments
contain
k
elements and the large segments contain the same small segments
k
, the
coordinates
X
and
Y
are represented equally.
If, during the laying out of the code, the number of small segments inside the
large segment does not correspond to the number of points inside the small
segment, then a seemingly different scale of coordinates is realized. For an expla-
nation of this property, the neural code can be represented not in the form of a one-
dimensional sequence but in the form of a three-dimensional parallelepiped where
the points of small segments are plotted along axis
X
and the small segments are
plotted along axis
Y
, forming the front plane of the parallelepiped along the axis
Z
-
the large segments, which in this case are the planes that intersect the parallelepiped
(Fig.
5.12
).
During coding of the
X
-coordinate, the shift of the code (Fig.
5.12
) is accom-
plished along the coordinate
X
, and everything that is advanced beyond the limits of
the volume of the parallelepiped is absorbed. If volume 1 is the neural code, then
volume 2 is the absorption after shifting along the
X
axis. Accordingly, while
coding coordinate
Y
, the shift is accomplished in the direction of this coordinate.
This structure of the code makes it possible to code even a third coordinate
Z
;
however, we reserve this possibility for other purposes.
Let us examine the behavior of the neuron code representing the object that is
described by the feature set extracted from the image. Each feature is shifted
correspondingly to its position. The object code is formed as the bitwise disjunction
of the codes of all features. This method of neural code formation leads to the fact
that with the shift of the entire object on the image, there is no need for new coding
of the feature positions as in the code of local connected coding. In order to code the
new position of the object, it is sufficient to move its corresponding code, which
Fig. 5.12 Properties of neural
codes
