Sieves for Selection of Genes According to Desired Behaviors - Systematic Design for Emergence in Cellular Nonlinear Networks - page 105

As seen in Chap. 4, a large percentage of cells (almost one third) belong to the
category of “imploding” patterns, potentially useful for feature extraction and
other such spatial filtering processes. By imposing the additional “large transients”
sieve, we hope to reduce the number of genes significantly enough to allow an in
depth investigation of their properties.
Several examples of CA cells selected from the above double sieve are depicted
in Fig. 6.12. We have selected only those cases that may be interpreted as feature
extraction. In each case the final steady state (representing a transform of the
initial state) was obtained after a relatively large number of iterations. Particularly
interesting is the case of ID = 5607 which appears to emphasize any initial state
cluster with no or less curvatures in its shape while it removes most of the pixels
of the shapes that have certain curvatures.
Such cells may be used to build a classifier capable to recognize the figure “1”
in a written text. Another interesting cell is the one with ID = 3619, where some
parts of the initial state patterns are removed. An in-depth further analysis would
probably reveal some practical applications. Many other cell genes are also on the
output list of the above double sieve that may deserve further investigation. Such
genes can do things such as removing circles or rectangles, edge extraction and so
on. In evaluating the properties of such genes it is useful to use an initial state con-
taining various shapes and figures so that human observers may associate a meaning
to the transformed state obtained as the terminal state of the cellular automata.
Fig. 6.12. Three examples of feature extractors selected by the sieve in Fig. 6.11. Different
values of alpha (parameter in the sieve definition) were used