The two support vector machines (SVMs) below are producers of support vectors.
The support vectors (SVs) that come from the the trained SVMs are sent to the
consumer node below the two producer nodes.

To run this demo, you need to perform some simple steps:
a) Generate som "+" and "-" instances on producer1 and producer2 by clicking in the coordinate system.
b) Click the "Go" button on each of producer1 and producer2.
c) Scroll down to the middle SVM named consumer1, which is a consumer of support vectors, and click "Go".
d) Scroll down below the five SVMs to read some hints about what happens.
Comments to the four steps above:
Ad. a) The "+" and "-" instances become the training set for the producer nodes.
Ad. b) The two SVMs are trained on the instances and then the support vectors are sent to the producer node.
Ad. c) The consumer node trains on the support vectors of the procucer nodes. Compare the decision boundary of the consumer node to the decision boundary of the producer nodes.

Variations and more fun:
*The two SVMs (producer3 and producer4) can also be used.
*Try to also create some new training data directly on consumer1. Then re-train.
*Try the Gaussian kernel... (use the drop down box)

Anyway, the point we are trying to make is to notice how the SVM can be used as a remote filter and send of its key information (support vectors) to other SVMs.


Rasmus Pedersen, Dept. of Informatics, Copenhagen Business School, Denmark, e-mail rup AT inf DOT cbs DOT dk.