Competitive Learning Methods

Introduction

This applet, DemoGNG, implements several methods related to competitive learning. It is possible to experiment with the methods using various data distributions and observe the learning process. A common terminology is used to make it easy to compare different methods. Hopefully, experimentation with the models will increase one's intuitive understanding and make it easier to judge their particular strengths and weaknesses. Here is a link on the updated version of the applet.

The following algorithms are available:

• Growing Neural Gas (Fritzke)
• Hard Competitive Learning (standard algorithm)
• Neural Gas (Martinetz and Schulten)
• Neural Gas with Competitive Hebbian Learning (Martinetz and Schulten)
• Competitive Hebbian Learning (Martinetz and Schulten)
• LBG (Linde, Buzo, Gray)
• Growing Grid (Fritzke)
• Self-Organizing Map (Kohonen)

Credits

The original applet was written by Hartmut S. Loos and Bernd Fritzke and slightly modified by Olivier Michel and Sébastien Baehni.

Applet

Questions

1. Hard Competitive Learning: explain the behavior of the iterative LGB method (K-means). Test this algorithm on the ring distribution and on the HiLo density.
2. Kohonen (Self Organizing Map): what are the principles of this algorithm? Test it with the ring distribution and the HiLo density. Explain the experimental results according to the theory.
3. Configure the Kohonen algorithm with a grid size of 1x30 and choose the rectangular shape. What is the role of the sigma_f parameter ? What is the value of the sigma_f parameter for which the representation changes from a straight line to an oscillatory line ?