Sensor Circuits - The Electronic Nose Artificial Olfaction Technology - page 200

Fig. 7.23 the basic structure

of SOM

in the input data space is preserved as faithfully as possible within the represen-
tation space. Thus, the similarity of the input data is reproduced to a very large
extent in terms of geographical area within the representation space.
A single computational layer of neurons organized in rows and columns is
arranged in feed forward structure. In the input layer all the source units are fully
connected to each neuron. The basic structure of SOM is shown in Fig. 7.23 .
For data clustering, the embedded competition paradigm is done by imposing
neighborhood limitation on the output unit, such that a certain topological property
in the input data is reflected in the outputs unit weights. SOM algorithm only
require sensors output. SOM is a network created by N neurons arranged as the
nodes of a planar grid.
The SOM Algorithm:
The aim is to study a feature map from the spatially continuous input space,in
which input vectors live, to the low dimensional spatially discrete output space,
which is formed by arranging the computational neurons into a grid.
The stages of the SOM algorithm that achieves this can be summarized as
follows:

1. Initialization—Choose random values for the primary weight vectors w j .
2. Sampling—Draw a sample training input vector x from the input space.
3. Matching—Find the winning neuron I(x)(best matching unit) with weight
vector closest to input vector, i.e., the minimum value of d j ðÞ¼ P

D

ð x i w ji Þ 2

i ¼ 1

4. Updating—Apply the weight update equation Dw ji ¼ g ðÞ T j ; I ðÞ ðÞð x i w ji Þ
Where T i ; I ðÞ ð t Þ = Gaussian neighborhood and g(t) is the learning rate.
5. Continuation—keep returning to step 2 until the feature map stops chang-
ing.(enough iterations for convergence)

The SOM is related to the category of competitive learning methods with
unsupervised learning rules. It basically carryout analysis of a topology preserving
projection of the data space onto a regular two-dimensional space where similar
samples are located together. On the other hand, the basic self-organizing map has
poor tracking capabilities when it is used with changing probability density of
data. With the complete knowledge of odors stored in a single self-organizing map
has the disadvantage of moving all the codebook vectors toward the new input