Online Learning Meets Semantic Communication over Wireless Channels

One major challenge for semantic communication networks is to deploy the network in practice where channel environments change dynamically. When adjusting the network weights on a subframe basis to the changing channel environments, the online over-the-air (OTA) training data is extremely limited, m...

Ausführliche Beschreibung

Gespeichert in:
Bibliographische Detailangaben
Veröffentlicht in:MILCOM IEEE Military Communications Conference S. 478 - 483
Hauptverfasser: Xu, Jiarui, Saeed, Usama, Ashdown, Jonathan, Liu, Lingjia
Format: Tagungsbericht
Sprache:Englisch
Veröffentlicht: IEEE 30.10.2023
Schlagworte:
and limited training data
Computational modeling
Image communication
joint source and channel coding
Military communication
online adaptation
Online learning
reservoir computing
Reservoirs
semantic communication
Semantics
Training data
Wireless communication
ISSN:2155-7586
Online-Zugang:Volltext
Tags: Tag hinzufügen
Keine Tags, Fügen Sie den ersten Tag hinzu!
Beschreibung
Zusammenfassung:One major challenge for semantic communication networks is to deploy the network in practice where channel environments change dynamically. When adjusting the network weights on a subframe basis to the changing channel environments, the online over-the-air (OTA) training data is extremely limited, making it difficult for networks with a large number of learnable parameters to converge. In this paper, we present an online reservoir computing (RC) aided adaptation approach for the semantic joint source and channel coding (JSCC) network to adjust the network weights on a subframe basis according to the channel changes. Specifically, the semantic JSCC network is first trained offline to learn a set of the initial weights with sufficient training data. During the online stage, where the channel model differs from the offline stage and the OTA training data is scarce, RC is adopted for equalizing the channel and adapting the end-to-end network weights to the change of channel. Experimental results demonstrate the effectiveness of the introduced approach in improving the performance of the semantic JSCC network on a subframe basis when the channel models do not match during the offline and online stages. Our work provides a potential solution for deploying semantic communication networks in practice in dynamic channel environments with a limited amount of available training data.
ISSN:2155-7586
DOI:10.1109/MILCOM58377.2023.10356346