Agriculture Reference
In-Depth Information
validation the curve decreases to a min-
imum and then starts to increase. Therefore,
the minimal point in the validation learning
curve may be used as a sensible criterion to
stop training (Haykin, 2001).
The error
e
j
(n) is minimized by a method,
and the general delta rule is applied when
differential and non-descendent activation
functions are applied such as the logistic
sigmoidal function and hyperbolic tangent
function. The logistic sigmoidal function
and hyperbolic tangent function are defined,
respectively, as:
reported for the identification of standards
and for solutions to non-linearity problems.
Because of the easy access to software pro-
grams and hardware, as well as the capacity
of neural networks to develop models com-
bined with other techniques, connectionist
models are powerful and modern tools that
attempt to understand and seek solutions
for specific problems in all fields of know-
ledge. Although the number of artificial in-
telligence applications in animal science
has increased in the last few years, this
number is still low. It may be explained by
the absence, until recently, of disciplines
that include connectionist models in under-
graduate and graduate courses. Today, some
new curricula have included this subject
and this will surely motivate animal scien-
tists to apply connectionist models.
Biological systems are surprisingly flex-
ible in processing information from the real
world. Some biological organisms have a cen-
tral processing unit called a brain. The human
brain contains 10
11
neurons and is capable of
intelligent processing in a precise and sub-
jective manner. Artificial intelligence (AI)
tries to bring to the digital processing world
the heuristics of biological systems in a var-
iety of manners, but still a lot needs to be done.
ANNs and fuzzy logic have been shown to
be effective in solving complex problems
using the heuristics of biological systems.
The number of AI applications in animal pro-
duction systems has increased significantly in
the last few years.
Huang (2009) mentions that ANNs have
been extensively studied and applied in
several fields during the past three decades.
Research on back-propagation training algo-
rithms for multilayer perceptron networks
has stimulated the development of training
algorithms for other neural networks includ-
ing the radial basis function, recurrent network
and Kohonen's self-organized competitive and
non-supervised networks. These networks,
particularly the multilayer perceptron net-
work with back-propagation training algo-
rithm, have been used in research and
applied in several scientific fields, and in
engineering. These networks were inte-
grated with other advanced methods, such
as fuzzy logic and wavelet, to increase data
1
(7.16)
y
=
,
0
<
<
y
1
1
+
e
-
x
-
x
1
1
-
+
e
e
y
=
,
-
< <
1
y
1
(7.17)
-
x
Data need to be normalized because the val-
ues presented to the neural network are
under the domain of the restricted image set
activation functions. Data can be normal-
ized by different functions, but the most
commonly used is:
z
xx
xx
-
-
=
i m
(7.18)
i
Mm
Where
z
i
is the normalized value of
x
i
, and
x
m
and
x
M
are the minimum and the maximum
values of dataset
X
.
Training Multilayer Artificial Neural
(MLP) networks with back-propagation may
demand several steps in the training set and
consequently training time is long. When a
local minimum is found, the error of the train-
ing set stops decreasing and is stuck at a higher
than acceptable value. The learning rate can be
increased without oscillation by changing the
general delta rule including the momentum, a
constant that determines the effect of previous
weight changes on the current direction of the
movement in weight spaces. Momentum is
useful in error spaces with long gorges, sharp
curves or valleys with smooth declines.
Multilayer Artificial Neural Networks
in Animal Science (MLP)
In animal science research, several applica-
tions using connectionist methods have been
