Exploring Topology Preservation of SOMs with a Graph Based Visualization

摘要: The Self-Organizing Map (SOM), which projects a (high-dimensional) data manifold onto lower-dimensional (usually 2-d) rigid lattice, is commonly used learning algorithm. However, postprocessing --- that often done by interactive visualization schemes necessary to reveal the knowledge of SOM. Thanks SOM property producing (ideally) topology preserving mapping, existing are designed show similarities local lattice without considering topology. This can produce inadequate tools investigate detailed structure and what extent preserved during learning. A recent graph based visualization, CONNvis [1], exploits underutilized topology, be suitable tool for such investigation. paper discusses represent on despite grid structure, hence preservation violations.