Unsupervised Learning
File names:
SOM.mbn and SOM.mbl
This example demonstrates a Self Organizing Map (SOM) of 10 x 10 output neurons
and 2 input neurons.
For unsupervised training the teacher 'Competitive with Momentum' is used. The following
teacher settings can be used for the example:
If you click on the button 'Advanced':
If you load the SOM example (SOM.mbn) it will look something like this:
Load the lesson SOM.mbl into the Lesson Editor and start the teacher. Ensure that the setting <View><Update
View during Teach> is activated.
You will then see how the patterns of the current lesson (named from "1" to "100") arrange on the SOM and find
locations in a way so that similar patterns locate close together in the SOM. Finally after the training the SOM
should look something like this:
Note that the ouput neurons of the SOM have been named according to the pattern for which they are the winner
neuron, i.e. the top left neuron has been named as '1' because this is the winner neuron for the pattern with name
'1' in the Lesson.
The lesson consists of patterns with X- and Y- input values ranging from 1 to 10. This correlates to rows and columns
in the SOM. The patterns are named in the Lesson from 1 to 100.
Since similar patterns are grouped together in a SOM after the training the pattern names with same X- resp. X-coordinate
value are located in the same row or column. Depending on the training run rows and columns may be flipped i.e. differ
from the above picture. This actually depends on the randomization start values for the weights of the links and also on
the random order of the patterns chosen during training.
Now check the menu option <View><Show Winner Neuron>, open the lesson editor and click on the button
<Think on Next Input> several times. You will see that the winner neuron of the corresponding input pattern
is visualized in the SOM by a blue cross:
You can use this function if you want to quickly identify the winner neuron of a SOM when applying new (untrained)
patterns. The location of the winner neuron in relation to the winner neurons named through the training will indicate
similarity between the new input pattern and the already trained patterns.
Note that the activation function for the SOM output neurons is 'MIN_EUCLID_DIST' which means 'Minimum Euclidian
Distance'. You can see this if you double click on one of the output neurons:
If you want to learns more about this activation function then check out MemBrain's help file.
What you should keep in mind is that you will need this activation function for the output neurons
of a SOM in order to obtain proper results.
Copyright © 2003 - 2007, Thomas Jetter