Deep Learning Enabled Semantic Communication Systems

Recently, deep learned enabled end-to-end communication systems have been developed to merge all physical layer blocks in the traditional systems, which make joint transceiver optimization possible. Powered by learning, natural language processing has achieved great success analyzing and understanding a large amount of texts. Inspired research results both areas, we aim provide new view on from semantic level. Particularly, propose learning based system, named DeepSC, for text transmission. Based Transformer, DeepSC aims at maximizing system capacity minimizing errors recovering meaning sentences, rather than bit- or symbol-errors communications. Moreover, transfer is used ensure applicable different environments accelerate model training process. To justify performance communications accurately, also initialize metric, sentence similarity. Compared with without considering information exchange, proposed more robust channel variation able achieve better performance, especially low signal-to-noise (SNR) regime, as demonstrated extensive simulation results.