Unsupervised Self-Organizing Feature Maps in Neural Networks free notes pdf
The unsupervised algorithm for training a SOM, recommended by Teuvo Kohonen, is mentioned in figure 4.22.After each and every input pattern is shown, the winner can be obtained along with the connection weights in its nearby area Nt increase, while the connection weights outside the location are kept as they are. a is a learning parameter. It is recommended that the training time moments t (cycles) tend to be more than 500 times the output neurons. If the training set contains a lesser number of instances compared to this number, then the whole training set is fed repeatedly into the network.