#mainTextBox As you should have expected, with two PEs, the competitive network simply divided the input data into two. The classification is horrible since the clustering does not correspond to the classification. We have added a file component to the competitive synapse which will now do LVQ. The LVQ adjusts the weights of the winning PE based upon whether it is the right classification or not. This will force the PEs towards the center of the CLASSES (not the CLUSTERS) and away from the other classes. Run the network and you should see much better performance.
#subtitleTextBox Disjoint (or Multimodal) Data
#mainTextBox We have now slightly modified the input data such that the horizontal class has been divided into two clusters, one on either side of the vertical class. This will cause problems because the center of the horizontal class is also the center of the vertical class. The center of the disjoint class does not allow good classification. In fact, often the PEs will move to strange locations as they are constantly pushed and pulled by correct and incorrect classifications.
#subtitleTextBox Summary
#mainTextBox This example showed how LVQ allows you to use competitive learning to do classification. The key point, however, is that it only works in relatively simple cases. Disjoint or multimodal data often creates difficulties for the LVQ algorithm. In fact, there are many versions of the LVQ algorithm (LVQ2, LVQ3, etc.) which try to alleviate some of these problems.
