Performance Analysis of Normality Test Loss for Intelligent RSCNN Denoiser Design With Application to Channel Decoding
Deep learning based signal processing technology has been applied in communication systems for enhancing information transmission performances, whose learning capability depends on the loss function using different normality tests. In this paper, we investigate the impacts of different normality test methods, including Jarque-Bera, Shapiro-Wilk, Kolmogorov-Smirnov and Anderson-Darling schemes, on the learning performance of the deep neural network used for the denoising and channel decoding. Explicitly, the Residual Shrink Convolutional Neural Network(RSCNN) is applied to increase the Signal-to-Noise Ratio (SNR), while the normality test is combined with the Mean Square Error(MSE) loss to keep the residual noise following Gaussian distribution. Besides, a weighting factor is added between two loss items to provide the overall normalization and prevent overpenalization of data with too much variance. Simulation results show that with the improved loss function, the transmission reliability performance can be enhanced, while the computational complexity of the deep learning network based on the non-parametric test is relatively higher
Chapter, 2022