SOMICA and Geometric ICA

Geometric independent component analysis (ICA) uses a weight update rule that is very similar to the self-organizing map (SOM) learning rule in the case of a trivial neighborhood function. In this paper we use this fact and present a new geometric ICA algorithm that uses a SOM for learning. The separation qualiry is better in comparison to other geometric algorithms, but the computational cost is highe,: Furthermore. this new algorithm will pmvide insight in how to transfer theoretical results from the SOM area to geometric ICA.