SplitNet: learning of tree structured Kohonen chains

This work introduces a tree structured neural network model for topology preserving vector quantization with one-dimensional Kohonen chains. The leaves of the tree are the chains, each of which quantizes a subspace of the input space. Topological defects can effectively be detected and splitting of the chain at that location results in a growing of the tree structure and increase of topology preservation. Additionally, the chains are able to grow and shrink in order to approximate user defined criteria. Advantages over existing dynamic network models are the flexible tree structure, the total lack of global parameters or calculations as well as the simulation and retrieval speed due to the network structure. Different levels of generalization and prototypicality are naturally observed