Biomedical Engineering Reference
In-Depth Information
Fig. 4
Neural network architecture composed of input, hidden and output layers
v i ¼ X
L
w m 1
ji
y m 1
j
þ b i
ð 2 Þ
j ¼ 1
where f is the activation function, L is the number of connections to the previous
layer, w m 1
ji
corresponds to the weights of each connection and b i
is the bias, which
represents the constant part in the activation function.
From among activation functions the sigmoid (logistic) function is the one most
usually employed in NN applications. It is given by:
fv i ¼
1
1 þ exp h v i
ð 3 Þ
ð
Þ
where h is a parameter defining the slope of the function.
It has been reported that the h parameter influences the speed of NN learning
and that the optimal value of the slope parameter is problem-dependent [ 29 , 44 ].
Nevertheless, a constant value of the h parameter of 1 (h ¼ 1) is generally applied
[ 17 , 29 , 44 ].
3.1 Training Algorithm
The training process in the NN involves presenting a set of examples (input
patterns) with known outputs (target output). The system adjusts the weights w m 1
ji
of the internal connections to minimize errors between the network output and
Search WWH ::




Custom Search