Biomedical Engineering Reference
In-Depth Information
the design of models in more complex networks.
It seems a logical approach to start the design of
these new models with simple CSs, and to orientate
the latest discoveries on astrocytes in informa-
tion processing towards their use in classification
networks, since the control of the reinforcement
or weakening of the connections in the brain is
related to the adaptation or plasticity of the con-
nections, which lead to the generation of activa-
tion ways. This process can therefore improve the
classification of the patterns and their recognition
by the ANGN (Porto et al., 2007).
are rather inefficient in fine-tuned local search
(Yao, 1999).
Moreover, numerous authors have considered
improving the efficiency of evolutionary training
by incorporating a local search procedure into
the evolution, i.e., combining EA's global search
ability with local search's ability to fine tune. For
instance, the hybrid Genetic Algorithms/Back-
propagation approach, used by Lee (1996) used
Genetic Algorithms (GAs) to search for a near-
optimal set of initial connection weights and then
used Backpropagation to perform local search
from these initial weights. Hybrid training has
been used successfully in many application areas
(Kinnebrock, 1994; Taha and Hanna, 1995; Yan et
al, 1997; Yang et al., 1996; Zhang et al., 1995).
Taking into account the above mentioned,
we designed a new hybrid learning method for
training ANGNs. Such method looks for optimal
connection weights in two phases. In one phase,
unsupervised learning, it modifies weights values
following rules based on the behaviour of brain
circuits studied. In the other phase, supervised
learning, the weights are searched by means of
GAs. The results (Porto et al., 2007) demonstrated
that the combination of these two techniques was
found to be efficient because GAs are good at
global search (Whitley, 1995; Yao, 1999) and the
use of this EA has enabled us to easily incorpo-
rate to CSs the artificial astrocytes which have
facilitated a fine-tuned local search. The basic
differences with regard to other hybrid methods
(Erkmen et al., 1997; Lee, 1996; Yan et al, 1997;
Zhang et al., 1995) are that we did not use the GAs
to search for initial connection weights and the
local search technique used is based on astrocytes
behaviour, which had never been implemented
until the present time. Details of this new hybrid
learning method can be seen further on.
Searching for a Learning Method
In order to design the integration of the astrocytes
into the Artificial Neural Networks (ANNs) and
elaborate a learning method for the resulting
ANGNs that allows us to check whether there is an
improvement in these systems, we have analysed
the main existing training methods. We have ana-
lysed Non-Supervised and Supervised Training
methods, and other methods that use or combine
some of their characteristics. After considering the
existing methods for training multilayer ANNs,
none of them seemed adequate to be exclusively
used. We have found that Backpropagation and
other based-gradient methods have difficulties
to train ANNs with more complex information
processing elements than those traditionally used
(Dorado, 1999; Dorado et al., 2000; Rabuñal et
al., 2004). Therefore, these methods could not be
used in order to achieve our essential aim, i.e. to
analyze the effect of including astrocytes in the
ANNs. Process elements's modifications in ANNs
have been made using Evolutionary Algorithms
(EAs) (Whitley, 1995) for the training and we
thought could be a good possibility to be taken
into account when it comes to training our CSs.
It's known EAs are particularly useful for dealing
with certain kind of problems. They are less likely
to be trapped in local minima than traditional
gradient-based search algorithms, although they
The Functioning of the ANGNs
The functioning of the ANGNs follows the steps
that were successfully applied in the construc-
Search WWH ::
Custom Search