Improvement and Application of Sampling Neural Network Using Virtual Space

Abstract

Accurate nonlinear modeling underpins every layer of modern engineering. Sampling neural network (SNN) is a new fitting network, which has a simple structure, clear physical concept, concise algorithm, stable performance, and adopts a new error diffusion training method similar to the diffusion of neural stimuli in living organisms. To improve accuracy, an improved SNN using virtual space (VSNN) was introduced. Based on the spectral analysis, four types of VSNN algorithms are obtained through zero extension, mirror symmetry, center symmetry, and constant extension, respectively. The VSNNs demonstrated a significant reduction, that is, 23.8%, 66.9%, 63.9%, and 71.3% in average root-mean-square error compared to SNN, extreme learning machine, least squares method, and backpropagation, respectively, across 40 benchmark tests. The research on VSNNs can effectively supplement the SNN algorithm system, promoting the bionic research of artificial neural networks. It has unique advantages in online training or equidistant-sampling problems. The application value is proved by the experiments of the Negative Temperature Coefficient sensor and the Chebyshev filter.