Learning properties of Modular Network SOMs

The modular network self-organizing map (mn-SOM) is a generalization of SOMs, where each node represents a parametric function such as a multi-layer perceptron or another SOM. Since given datasets are fewer than the nodes in general, some nodes never win in competition and have to update their parameters from the winners in the neighborhood, which can be regarded as interpolation. This study derived the interpolation curve between winners in a simple case and discussed the distribution of winners based on the neighborhood function.