Applications of Emergent Phenomena - Systematic Design for Emergence in Cellular Nonlinear Networks - page 137

The CA array then the “reader” were invoked producing a database with 143
seven-dimensional feature vectors (corresponding to the same number of charac-
ters in the image). The first 50 were used for testing and the remaining 93 were
used for training a SVM (the simulator in [68] was used configured as C-SVM
with Gaussian kernels and the “gamma” parameter equal to 5). The confusion matrix
indicates that except characters “1” and “3” all others were all correctly recognized
on the test set. This quite good performance indicates that although simple, the pro-
posed feature extractor is robust and reliable. The character “1” has a recognition
rate of 75% and character “3” a rate of 90% on the test set. In fact one “1” and one
“3” were the only misclassified characters. When not properly recognized these
two characters are decided to be “4”.
A test on a distorted image as seen in Fig. 8.7, was then considered, with the
SVM trained on the database from Fig. 8.6. Here O was selected in the func-

tion FEATURE. The result shows that only 2 of the 16 characters (12.5%) were
misclassified, still a very good result given the distortions.

11

Fig. 8.7. A set of distorted characters were recognized correctly except the two arrow
pointed characters. The “2” was classified as “8” and the “1” was classified as 6

8.3 Excitable Membranes, for Temporal Sequence Classification

A growing number of applications are developing in the sense of rapidly proc-

essing large amounts of data. Quite often such data is readily available as signals
obtained from various sensors integrated in modern concepts such as “ambient
intelligence” .Many applications require to recognize such signals at high speeds
and with reasonable accuracy while being portable, with low power consumption.
Such applications include but are not limited at the recognition of various abnor-
mal states (e.g. epileptic state or other similar) from continuously available
bioelectric signals, or voice recognition for different purposes (either commercial
or biomedical, such in devices helping disabled people).
One of the most investigated area of research is that of speech recognition
where the aim is to recognize voice signals. But many other signals can be recog-
nized using similar approaches. Traditional solutions (e.g. those based on Hidden
Markov Models or HMM) are computationally intensive, requiring complex signal