Based On Non-local Diffusion Properties Of Nitric Oxide Neurotransmitters Self-organizing Neural Gas Algorithm And Applications

In recent years, intrinsic Nitric Oxide (NO) is recognized as a new neurotransmitter, which plays an important role in brain neural activities, and attracts wide attention. NO provides some new inspiration about some of the traditional tenets in the domain of brain science, artificial neural network and neural signal processing, and it is valuable to introduce NO mechanism into self-organizing neural network models.This paper is focused on neurocomputing ability of NO and improving on self-organizing methods, which supported by National Natural Science Foundation of China (60171003). Firstly some typical self-organizing neural networks are described and their characteristics are analyzed and evaluated with some simulations. Secondly, the dynamic of NO diffusion is modeled based on the latest researches and conclusions, and then a novel growing self-organizing model called Diffusing and Growing Self-Organizing Maps (DGSOM) is presented by combining the NO diffusion mechanism with growing topology-unpreserving network. The DGSOM model increases units through competition mechanism, generates and updates the topology of network using Competitive Hebbian Learning (CHL) fashion and aging mechanism, uses NO diffusion model and topology connection model to build the dynamic balance of the network, and adopts short-range competition and long-range cooperation as fine-tuning manner. A lot of simulations indicate that the DGSOM model can not only compartmentalize input space correctly, but also reflect the important topological relations and estimate the intrinsic dimensionality in a given input signals. Its superiority over the other self-organizing methods is the plasticity and flexibility of the network, which makes it excellent for the DGSOM model to handle various non-stationary input distribution. Finally, the applications of the DGSOM model in nonlinear system identification and data clustering are analyzed, the expectations of NO applied to information processing and neurocomputing are discussed, and directions and targets in our further work are suggested.