Similarity analysis of molecules with self-organizing surfaces-An extension of the self-organizing map

The authors give an extension of the self-organizing map, called self-organizing surface, which has been developed in order to cope with the typical topological defects that self-organizing maps develop when mapping the surface of a 3D object to a 2D torus of neurons. The extension replaces the local grid with a parallel distance calculation on arbitrary topologies. The technique was first developed as an implementation technique of self-organizing maps on massively parallel computers. However, it has given rise to a conceptual extension of the self-organizing map where real-valued neuron positions can now be distributed over arbitrary non-convex and even disconnected areas. The extension naturally allows for an efficient dynamic change of the network size and density. The algorithm is demonstrated by mapping the surfaces of molecules to the surface of a sphere