Evaluation of the Performance of O(log2M) Self-Organizing Map Algorithm without Neighborhood Learning
Hiroki Kusumoto, Yoshiyasu Takefuj · 2006
In this paper, we evaluate the performance of O(log2M) self organizing map (SOM) algorithm previously presented by us. Our algorithm was developed for reduction of the computational costs and is the fastest one in SOM algorithm. The order of the computational costs is log2M where the size of a feature map is M 2. Our algorithm does not require neighborhood learning and thus tuning of parameters is relatively simple. The performances of the basic SOM developed by Kohonen and our algorithm were tested using the benchmark of a central nervous system (CNS) tumor patient dataset which comprises five groups. The simulation results show our algorithm can map the input dataset more appropriately than the basic SOM with constant or changing parameters of the neighborhood function. Our algorithm is able to contribute to various research fields using the SOM algorithm.