Meta-Modeling for Autoencoder-Based End-to-End Communications Systems

With the rapid development of neural network in various applications, it has also attracted great attention in communication. However, neural network is a black box, it is not clear why neural networks can effectively improve the performance of traditional communication technologies. In this paper, we use the meta-model to approximate the autoencoder, and obtain the symbolic explanation of autoencoder, which consists of incomprehensible neural network. Most importantly, we use the symbolic explanation to explain the principle of autoencoder-based transceiver, and give some reasons why the performance of autoencoder is better than that of traditional coding technology. So that users can understand and trust the complex neural network better. In addition, this paper also gives the symbolic explanation of autoencoder trained under different SNR, and analyzes the influence of different SNR on autoencoder.