Growing neural gas efficiently

• Published in: Neurocomputing (Amsterdam) Vol. 104; pp. 72 - 82
• Main Authors: Fišer, Daniel, Faigl, Jan, Kulich, Miroslav
• Format: Journal Article
• Language: English
• Published: Amsterdam Elsevier B.V 15.03.2013; Elsevier
• Subjects: 3-D surface reconstruction; Algorithms; Applied sciences; Artificial intelligence; Classification; Computer science; control theory; systems; Connectionism. Neural networks; Constraining; Data processing. List processing. Character string processing; Exact sciences and technology; Growing neural gas; Memory organisation. Data processing; Nearest neighbor search; Neural networks; Optimization; Pattern recognition. Digital image processing. Computational geometry; Preserves; Reconstruction; Self-organizing map; Software; Three dimensional; Nearest neighbor search; 3-D surface reconstruction; Growing neural gas; Self-organizing map; Cluster analysis; Nearest neighbour; Neural network; Kohonen algorithm; Optimization; Image reconstruction; Self organization; Classification; Internal structure; Tridimensional image; Reduced order systems; Large scale
• ISSN: 0925-2312; 1872-8286
• DOI: 10.1016/j.neucom.2012.10.004

This paper presents optimization techniques that substantially speed up the Growing Neural Gas (GNG) algorithm. The GNG is an example of the Self-Organizing Map algorithm that is a subject of an intensive research interest in recent years as it is used in various practical applications. However, a poor time performance on large scale problems requiring neural networks with a high amount of nodes can be a limiting factor for further applications (e.g., cluster analysis, classification, 3-D reconstruction) or a wider usage. We propose two optimization techniques that are aimed exclusively on an efficient implementation of the GNG algorithm internal structure rather than on a modification of the original algorithm. The proposed optimizations preserve all properties of the GNG algorithm and enable to use it on large scale problems with reduced computational requirements in several orders of magnitude.