Biomedical Engineering Reference
In-Depth Information
tion and use of CSs: design, training, testing and
execution.
The design is based on feed-forward multi-
layer architectures which are totally connected,
without back propagation or lateral connections,
and oriented towards the classification and rec-
ognition of patterns.
ANGNs hybrid training method combines
non-supervised learning (first phase) with the
supervised learning that uses the GAs evolution-
ary technique (second phase).
The first phase is based on the behaviour of
astrocytes (Perea & Araque, 2005). This behaviour
has been incorporated to the ANNs as control ele-
ments (CE). The ANGNs have one CE per layer
which is responsible for monitoring the activity
of every neuron in that layer (Figure 1).
Since the GAs requires individuals, the first
phase creates a set of individuals to work with.
Each individual of the GA consists of as many
values as the connection weights existing in the
CS, and each arbitrary set of values of all the
weights constitutes a different individual. We
study the CS functioning with all the individu-
als. Every individual is modified as each training
pattern passes on to the network, according to
how the activity of the neurons has been during
the passage of that pattern. For each individual,
every pattern or input example of the training set
is presented to the network during a given number
of times or iterations. These iterations allow the
modification of the individual by means of the
application of the rules based on brain circuits
behaviour, and these iterations constitute a pattern
cycle. The number of iterations can be changed
and can be established for any cycle. As regards
these modifications over the connection weights,
we selected a set from all the different possibili-
ties analyzed in our computational models (Porto,
2004; LeRay et al., 2004) that will be explained
in the following section.
Going on with the training, when a pattern
cycle has finished, the error for that pattern is
calculated and stored. It will be the difference
between the CS output obtained and the one
desired. Later on, when all the training patterns
have been passed on to the CS, the mean square
error (MSE) of that individual is calculated. We
have opted for the MSE because it gives a rela-
tive measure to the patterns that are fed to the
network to compare the errors between different
architectures and training sets. Also, the square
in the numerator favours the cases of individuals
for which the output of the network is close to the
optimal values for all the examples.
The process is the same for all the individuals.
This phase constitutes a non-supervised training,
because the modifications of the connections
Figure 1. CS with artificial neurons (AN) and control elements (CE)
A N
A N
A N
A N
A N
A N
A N
C E
A N
C E
C E
Search WWH ::
Custom Search