Pattern Recognition Techniques - Biomedical Signal Analysis: Contemporary Methods and Applications - page 165

p 0

w 0

g

f

p 1

w 1

m

Σ

w i p i

i=0

w m

p m

Figure 6.3
Propagation rule and activation function for the MLP network.

The input vector is usually the result of a preprocessing step of a
measured sensor signal. This signal is denoised, and the most relevant
information is obtained based on feature extraction and selection. The
MLP acts as a classifier, estimates the necessary discriminant functions,
and assigns each input vector to a given class. Mathematically, the
MLP belongs to the group of universal approximators and performs a
nonlinear approximation by using sigmoid kernel functions. The learning
algorithm adapts the weights based on minimizing the error between
given output and desired output.

The steps that govern the data flow through the perceptron during
classification are the following [221]:

R l to the perceptron, that is,

1. Present the pattern p =[ p 1 ,p 2 ,...,p l ] ∈

l .
2. Compute the values of the hidden layer nodes as is illustrated in figure
6.3:

set x i = p i for 1 ≤ i ≤

1

1+exp

w 0j + i=1 w ij x i 1

h j = ≤ j ≤ m (6.1)

−