Scale-Adaptive Growing Neural Network Based on Distortion Error Stability and its Application in Image Topological Feature Extraction

Abstract

Self-organizing neural networks are characterized by topology preservation, dynamic adaptation, clustering, and dimensionality reduction, which prompt their wide application in data mining, knowledge extraction, and image processing. However, existing self-organizing neural networks fail to automatically generate an output space that contains an appropriate number of neurons according to the data input. To address this problem, this paper proposes a growing neural gas (GNG) algorithm with adaptive output network scale, which is called scale-adaptive GNG (SA-GNG) algorithm. The learning process of SA-GNG is divided into two stages: growth and convergence. At the growth stage, distortion error stability is introduced to objectively judge the degree of approximation of the output network to the input space, so that SA-GNG can grow neurons on demand until significant improvement is no longer made for distortion errors. In the convergence stage, neurons are not allowed to be produced, and the similarity between the output network and the input data is improved through continuous learning. SA-GNG promises to autonomously generate an appropriate number of neurons according to the size of the input data, with no need of determining the total number of neurons to be generated in advance, thereby greatly improving its adaptability. As such, the algorithm is especially suitable for the application scenarios where the amount of the data to be input is unknown. The validity and feasibility of the algorithm proposed in this paper are verified by experiments.