Biomedical Engineering Reference
In-Depth Information
Table 5.1
Y
X
-20
-10
0
10
20
-20
26,36
26,24
22,21
26,11
17,40
26,19
12,87
26,10
7,48
26,09
-10
26,48
21,79
21,79
21,48
16,97
21,46
12,80
21,26
7,71
21,21
0
26,59
17,03
21,72
16,65
16,86
16,55
12,73
16,43
7,68
16,43
10
26,66
11,26
21,41
11,38
16,47
11,43
12,62
11,56
7,75
11,57
20
26,48
6,46
21,34
6,54
16,58
6,42
12,50
6,40
7,54
6,25
5.2.2.2 Application of shift coding
The basic associative field of shift coding can be used for image recognition
because it provides recognition independently of object position. Cyclic shift
coding without absorption makes it possible to obtain the recognition invariant to
the rotation of an object on the image, and only invariance to the scale of the image
must be obtained by additional methods (for example, learning of the image
simultaneously on several scales).
One positive quality of shift coding is that the smaller neural ensembles
corresponding to the frequently occurring fragments of the image are formed
automatically and simultaneously with the formation of the neural ensembles
corresponding to the entire object. Thus, if we train this neural network for face
recognition, the ensembles that correspond to eyes, nose, mouth, and so forth must
be formed in the network. This quality can be used also for treating natural-
language texts. If we code separate letters, and their positions in the word are
coded by the corresponding shift, then the neural ensembles corresponding to the
most frequent combinations of letters, including the roots of the words, the prefixes,
and the suffixes, must be formed automatically. Similarly, at the training level of
phrases, the ensembles corresponding to the most frequent word combinations must
be formed. This property can be useful when creating information storage and
retrieval systems. Local connected coding does not possess these properties, so it is
necessary to use artificial methods to form analogous “small” ensembles.
The specific application of shift coding will be described in the following
paragraphs since shift coding actually is the basis of multi-float coding. The
classifiers built using multi-float coding showed sufficiently good results, which
will be described below.
5.2.3 Functions of Neural Ensembles
The vectors formed by the method described above (for example, with the use of
local connected coding) are used to train the associative field. Each unit in the
vector corresponds to the active neuron, and 0 corresponds to the neuron that is not
excited. During training, two active neurons are connected with binary connection.
