Representing 2D Objects. Comparison of Several Self-Organizing Networks
Francisco Flórez‐Revuelta, José M. Amigó, José García‐Rodríguez, Antonio Hernández · 2002
Abstract: Self-organizing networks have been generally considered as topology preserving of an input space. However, following recent definitions of topology preservation, not every self-organizing model has this quality. In this work, we study the topology preservation capability of four different self-organizing models: Self-Organizing Maps, Growing Cell Structures, Neural Gas and Growing Neural Gas. We use the topographic product to determine if these networks preserve the topological features of several bidimensional objects. We conclude that neuronal gases obtain better results since they can modify the topology of the network during the learning phase. Also, within these, the Growing Neural Gas has a smaller complexity than the Neural Gas, reason why it is candidate to be used in the representation of 2D objects. In future works, we will employ this network to represent 2D objects and to extract geometrical features.