An energy-based SOM model not requiring periodic boundary conditions
Alexander Gepperth · 2017
We present a SOM model based on a continuous energy function derived from the original energy-based model developed in [9]. Due to the convolution that is contained in the energy function, this model can only be applied when periodic boundary conditions are imposed (toroidal SOM), leading to markedly higher quantization errors, especially for small map sizes. We introduce a simple strategy, based on the assumption of homogeneous long-term averages for input-prototype distances, that allows to operate the energy-based SOM model without periodic boundary conditions, and demonstrate that its quantization errors are consistently lower especially for small map sizes. Simple experiments are conducted showing the worth of a continuous energy function, namely for novelty detection and automatic control of SOM parameters.