Biomedical Engineering Reference
In-Depth Information
down the use of the RANN to solve static and
dynamic problems.
However, the chances of RANN are very big
compared to the powerful of feedforward ANN.
For the dynamic or static pattern matching, the
RANN developed until now offer a better per-
formance and a better learning skill.
Most of the studies that have already been done
about RANN, have been center in the develop-
ment of new architectures (partial recurrent or
with context layers, whole recurrent, etc.) and
to optimize the learning algorithms to achieve
reasonable computer times. All of these studies
don't reflect changes in the architecture of the
process elements (PE) or artificial neurons, that
continue having an input function, an activation
function and an output function.
The PE architecture has been modified, basing
our study in biological evidences, to increment
the RANN powerful. These modifications try to
emulate the biological neuron activation that is
generated by the action potential.
The aim of this work is to develop a PE model
with activation output much more similar to the
biological neurons one.
process. The ANN's are organized according to
training methods for specific applications.
There are two types of ANN's, the first one
with only feed forward connections is called feed
forward ANN, and the second one with arbitrary
connections without any direction, are often called
Recurrent ANN (RANN). The most common
type of ANN consists on different layers, with
some neurons on each of them and connected
with feed-forward connections and trained with
the back propagation algorithm (Johansson et
al., 1992).
The numbers of neurons contained in the input
and output layers are determined by the number
of input and output variables of a given problem.
The number of neurons of a hidden layer is an
important consideration when solving problems
usign multilayer feed-fordward networks. If there
are fewer neurons within a hidden layer, there may
not be enough opportunity for the neural network
capture the intricate relationships between the
inputs and the computed output values. Too many
hidden layer neurons not only require a large
computational time for accurate training, may
also result in overtraining situation (Brion et al.,
1999). A neural network is said to be “overtrained”
when the ANN focuses on the characteristics of
individual data points rather than just capturing
the general patterns in the entire training set.
The optimal number of neurons in a hidden layer
can be estimated as two-thirds of the sum of the
number of input and output neurons.
An ANN has a remarkable ability to derive
meaning from complicated or imprecise data.
The ANN can be used to extract patterns and
detect trends that are too complex to be noticed
by either humans or other computer techniques.
“Training” of an ANN model is a procedure by
which the ANN repeatedly processes a set of
test data (input-output data pairs), changing the
values of its weights according to a predetermined
algorithm in order to improve its performance.
Backpropagation is the most popular algorithm
for training feed-forward ANN's (Lippman,
bACkGROUND
Artificial Neural Networks
An Artificial Neural Network (ANN) (Lippmann,
1987; Haykin, 1999) is an information-processing
system that is based on generalizations of human
cognition or neural biology and they are electronic
or computational models based on the neural
structure of the brain. The brain basically learns
from experience. An Artificial Neural Network
consists on various layers of parallel procesing
elements or neurons. One or more hidden layers
may exist between the input and the output layer.
The neurons in the hidden layer(s) are connected
to the neurons of a neighboring layer by weight-
ing factor that can be adjusted during the training
Search WWH ::
Custom Search