Joint Channel Coding and Modulation via Deep Learning

Yihan Jiang, Hyeji Kim, Himanshu Asnani, Sreeram Kannan, Sewoong Oh, Pramod Viswanath

Abstract: Channel coding and modulation are two fundamental building blocks of physical layer wireless communications. We propose a neural network based end-to-end communication system, where both the channel coding and the modulation blocks are modeled as neural networks. Our proposed architecture combines Turbo Autoencoder together with feed-forward neural networks for modulation, and hence called TurboAE-MOD. Turbo Autoencoder was introduced in [1] and consists of a neural network based channel encoder (convolutional neural networks with an interleaver) and a neural network based decoder (iterations of convolutional neural networks with interleavers and de-interleavers in between). By allowing joint training of the channel coding and modulation in an end-to-end manner, we demonstrate that TurboAE-MOD performs comparable to modern codes stacked with canonical modulations for moderate block lengths. We also demonstrate that TurboAE-MOD learns interesting modulation patterns that are amenable to meaningful interpretations.