LBG-m: a modified LBG architecture to extract high-order neural structures

Neural networks that learn in an unsupervised way and generate their topology during learning can be useful to build topology representing structures. These networks can be used for vector quantization and clustering each time it is necessary to characterize the topology of the underlying data distribution. A drawback of these networks is that the structures created have the same complexity as the input data, so a simplification of the structure is needed to allow the user to visualize and manipulate these representations. The aim of the proposed algorithm is to simplify the graph structure created by these kinds of neural networks. The LBG-m algorithm takes the position of the nodes and the adjacency matrix of the graph as input and builds an over-imposed graph that clusters the graph nodes and tries to reproduce the “shape” of the input graph