On weight adjustment of self-organizing feature maps

This paper proposes an alternative adjustment of weights for Kohonen learning that provides a better match of topology and an improved representation of the input data. Based on the adjustment of weights, a new weighted energy function is defined. Then we propose a new scheme that applies the multiple training concept to the weighted energy function. The new scheme results in a better directional adjustment of weights and increases the training speed of the network dramatically. Finally experimental results demonstrate that much better recognition rate is obtained.