Information Technology Reference
In-Depth Information
Formal neuron consists of several inputs (x
1
,…, x
n
) and their connection weights (w
1
,…, w
n
),
formal input (x
0
) and its connection weight, so called bias or threshold (w
0
), neuron's own
body, where the computation of the output is made, and one output (y), which can be
further branching. The neuron function itself is divided into two steps. First, postsynaptic
potential (ξ), i.e. weighted sum of all inputs, including formal input, is calculated. In the
second step, so called activation function is applied to postsynaptic potential and its result is
the value of the neuron output.
The connection weight is a key element of every ANN. It expresses the relative importance
of each neuron input or, in other words, the degree of influence of the particular input to the
output. Connection weight represents storage of patterns learned from the input data. The
ability of learning lies exactly in the repeated refinement of the connection weights.
Because of their ability to learn and generalization, artificial neural networks are used in
many applications of prediction and data classification. Aburas (Aburas et al., 2010) use
neural networks to predict the incidence of confirmed cases of dengue fever. This prediction
is based on the observations of real parameters, such as average temperature, average
relative humidity, total rainfall, and the number of reported cases of dengue fever as a
response to these parameters.
Faisal, Ibrahim and Taib (Faisal et al., 2010) also deal with the issue of dengue fever disease.
They proposed a non-invasive technique to predict the health risks to ill patients via
combination of self-organizing maps and multilayer feed-forward neural networks.
Combining these techniques, they achieve 70% accuracy of forecasts.
Gil et al. (Gil et al., 2009) describe the use of artificial neural networks in the diagnosis of
urological disorders. To suppress the main neural networks drawbacks, so called over-
learning or over-fitting, they use a combination of three different ANN architectures, two
unsupervised and one supervised. This combination has provided decision support with
verified accuracy of almost 90%.
Both, Faisal (Faisal et al., 2010) and Gil (Gil et al., 2009), also mentioned the possibility of
increasing the accuracy of their systems by combining artificial neural networks with fuzzy
inference techniques. This approach uses Kannappan (Kannappan et al., 2010) for design the
system for prediction of autistic disorders using fuzzy cognitive maps with nonlinear Hebb
learning algorithm. Fuzzy cognitive maps combine the strengths and virtues of fuzzy logic
and neural networks.
As mentioned earlier, artificial neural networks are often used for prediction, or to predict
the probable progression of examined data. High-quality and credible prediction can be
derived only on the basis of a sufficiently large volume of data. Generally, the more relevant
input data is available, the more accurate is the prediction of their progression. A common
epiphenomenon of materials research, as well as the development of new products, is very
limited amount of data. In such cases, the neural networks must be appropriately modified
to achieve an acceptable accuracy of prediction based on small data sets.
Lolas and Olatunbosun (Lolas & Olatunbosum, 2008) used ANN to predict reliability behavior
of an automotive vehicle at 6000 km based solely on information from testing the prototype.
To this propose, they drawn up a three-phase optimization methodology for neural network
development. The proposed network can detect degradation mechanism of the vehicle and use
this knowledge to predict the trend of reliability throughout its life cycle. The overall error of
the whole neural network and the three output parameters were less than 9%.
Li (Li & Yeh, 2008) deals with the prediction of a product life cycle already in the initial
stages of manufacturing. For work with the small data sets, they developed nonparametric
Search WWH ::
Custom Search