Overfitting and Underfitting Analysis for Deep Learning Based End-to-end Communication Systems

In this paper, we study the deep learning (DL) based end- to-end transmission systems, then we present the analysis for the underfitting and overfitting phenomena which happen during the training of the neural networks (NNs). Different from the DL based image processing, the transmission systems require the bit error rate to be as low as 10– ⁴ to achieve high reliability, thus the training errors induced by the underfitting and overfitting may greatly degrade the transmission reliability performances of DL-based communications. In considered DL-based communication systems, we propose to use the average, variance and minimum of transmitted signals' minimum Euclidean distance to estimate the effects of underfitting and overfitting on the error rate performances in terms of the energy per bit to noise power spectral ratio E b / N o of signals. Furthermore, we propose to apply the regularization scheme to alleviate the overfitting issue. Simulations are performed to demonstrate the underfitting and overfitting analysis for transmission systems over the additive white Gaussian noise (AWGN) channel, and validate the improved performances by applying the regularization.