Neural joint source-channel coding: benefits of bayesian perspective and its limitation

Abstract

In this paper, we studied a method of transmitting information using neural joint source-channel coding in a noisy communication channel environment. According to the source-channel separation theorem defined by Shannon, it was possible to construct a communication system that separately optimizes source-channel coding. These has limitations in the assumption that infinite-blocklength regime. Through the previous studies that optimize the joint source-channel coding based on machine learning, the encoding and decoding performance has been improved. The neural joint source-channel network still did not show remarkable performance due to the limitation of optimizing the discrete code to compress it into fixed-length codewords and correct errors. In this paper, unlike previous studies that implemented neural joint source-channel coding that maximizes mutual information between data and codewords, the coding process is reparameterized using the Gumbel-Max trick. For baysesian perspective, optimizing variational autoencoder (VAE) by log-likelihood function method is proposed. The experimental results show higher compression and error correction performance in same datasets. By introducing more realistic channel model, we clarify the limitations of the overall model and show the overcome possibilities of our model by using model-agnostic meta-learning (MAML).