Kohonen Net Clustering

The Kohonen (1984) net has much in common with connectionist learning ( Boltzman machines (Hinton and Sejnowski 1986), Hopfield networks (1982)). The Kohonen net can be thought of as two layers of nodes. The input layer has a node for each input. Thus if objects are described by 3 attributes there will be 3 input nodes. There are a larger number of output nodes as shown in figure 3.12. Each input node is fully connected to every output node. There is a weight associated with each connection from the input layer to the output layer. The positions of the output nodes are determined by the values of the weights.

 

 

Figure 3.12 : Full connection between the input and output layer in a Kohonen net

 

Whenever an input vector is presented to the network the weights are amended in line with the algorithm presented below. The effect of this is to move the nodes to form clusters where there are many inputs.

 

  • CHOOSE random values for all weights on all nodes.
    •
  • CHOOSE a positive value for K.
    •
  • CHOOSE a value between 0 and 1 for c.
    •
  • WHILE K and c are not negligible
    •
  • FOR each training instance COMPUTE the distance of each node from the instance
    •
  • SELECT the closest node
    •
  • FOR each node within distance K of the selected node UPDATE the weights by a factor based on c
    •
  • REDUCE K and c
    •

     

    The effect of applying the algorithm can be seen in figure 3.13.

     

    Figure 3.13 : The Kohonen net correlates input by forming clusters of nodes

     

    Table of Contents

    References

    Authored by Serengul Smith

    E-mail to: serengul1@mdx.ac.uk
    School of Computing Science Middlesex University
    Revised: September 1998