Pattern Recognition Techniques - Biomedical Signal Analysis: Contemporary Methods and Applications - page 183

elements K jk ,

h jk
σ j ∗

K jk = (6.25)

σ k

are the correlation coecients h jk and σ j the standard deviation of the

i th shape matrix.
For h jk we choose: h jk =1for j = k ,and

1otherwise.
4. Forward computation of output layer's activations: Calculate the
values of the output nodes according to

| h jk |≤

f oj = ϕ j =
i

w ji ψ i (6.26)

5. Updating: Adjust weights of all neurons in the output layer based on
a steepest descent rule.
6. Continuation: Continue with step 2 until no noticeable changes in the
error function are observed.

The above algorithm assumes that the locations and the shape of a
fixed number of radial-basis functions are known a priori. RBF networks
have been applied to a variety of problems in medical diagnosis [301].

Design considerations
The RBF network has only one hidden layer, and the number of basis
functions and their shape are problem-oriented and can be determined
online during the learning process [151, 206]. The number of neurons
in the input layer equals the dimension of the feature vector. Likewise,
the number of nodes in the output layer corresponds to the number of
classes.
The success of RBF networks as local approximators of nonlinear
mappings is highly dependent on the number of radial-basis functions,
their widths, and their locations in the feature space. We are free to
determine the kernel functions of the RBF networks: they can be fixed
or adjusted through either supervised or unsupervised learning during
the training phase.
Unsupervised methods determine the locations of the kernel func-
tions based on clustering or learning vector quantization. The best-
known techniques are hard c -means algorithm, fuzzy c -means algorithm