Biomedical Engineering Reference
In-Depth Information
Fig. 4
Kolmogorov network 1 (K1)
Fig. 5
Kolmogorov network 2 (K2)
structure is derived from that of a benchmark test of classifying tumors based on cell
descriptions gathered by microscopic examination [
31
].
The structure of the network (C1) is given in Fig.
6
:
Training
In light of the relatively small sizes of the networks considered in this work and conse-
quentially, the small number of weights, the Levenberg-Marquardt (LM) algorithm
is chosen as training function. Resilient back propagation was also considered in sim-
ulations. In the resilient back-propagation training, only the sign of the derivatives
of the training function is used to update the weights. The size of the weight update
is determined by another value. This significantly improves the training speed but
simulations showed that its performance was not ideal for the small error tolerance
desired.
The ratio of input data to be used for training, validation and testing has also been
left at the default of 0.6, 0.2 and 0.2, respectively.
Results
The accuracy (
ACC
) of a test is the proportion of the true positives (
TP
) and true
negatives (
TN
) to the number of estimated positives (
P
) and negatives (
N
). This can
be expressed as
ACC
N
) where
P
and
N
are the total numbers
of positives and negatives classified by the binary classifier.
=
(
TP
+
TN
)
/
(
P
+
Search WWH ::
Custom Search