Joint Source-Channel Coding for a Multivariate Gaussian over a Gaussian MAC using Variational Domain Adaptation

With the development of the distributed learning and edge computing, servers must often receive information from multiple terminal devices; thus, the importance of source-channel coding for distributed sources over multiple access channels (MACs) becomes more and more significant. This study present...

Ausführliche Beschreibung

Gespeichert in:
Bibliographische Detailangaben
Veröffentlicht in:IEEE transactions on cognitive communications and networking Jg. 9; H. 6; S. 1
Hauptverfasser: Li, Yishen, Chen, Xuechen, Deng, Xiaoheng
Format: Journal Article
Sprache:Englisch
Veröffentlicht: Piscataway IEEE 01.12.2023
The Institute of Electrical and Electronics Engineers, Inc. (IEEE)
Schlagworte:
Adaptation
Adaptation models
Coding
Correlation coefficients
Decoding
Deep joint source-channel coding
Delays
domain adaptation
Domains
Edge computing
Encoding
Feature extraction
Learning
multiple-access channel
Multivariate analysis
multivariate Gaussian
Regularization
Symbols
variational autoencoder
ISSN:2332-7731, 2332-7731
Online-Zugang:Volltext
Tags: Tag hinzufügen
Keine Tags, Fügen Sie den ersten Tag hinzu!
Beschreibung
Zusammenfassung:With the development of the distributed learning and edge computing, servers must often receive information from multiple terminal devices; thus, the importance of source-channel coding for distributed sources over multiple access channels (MACs) becomes more and more significant. This study presents a deep joint source-channel coding (JSCC) design for a multivariate Gaussian source over a Gaussian MAC. The widely used autoencoder based deep-JSCC cannot perform stably under such conditions due to their easiness to fall into local optimum. Therefore we propose the variational domain adaptation (VDA)-JSCC scheme. Firstly, the loss function with an additional regularization term is introduced through variational analysis. The crucial prior distribution related to this item is obtained by domain adaptation, which is a transfer learning method. The proposed fine-tuning technique during the training process yields further performance improvement. Experiment results show that VDA-JSCC can always learn reasonable coding structures without artificial design and outperforms other state-of-the-art methods under different channel signal-to-noise ratios (CSNRs). We have also analyzed the reason why the performance of VDA-JSCC deteriorates in high CSNR range and then replace the encoder of VDA-JSCC with Mixture-of-Experts to improve its performance in high CSNR range. Finally, VDA-JSCC exhibits considerable robustness when the channel quality or correlation coefficient varies.
Bibliographie:ObjectType-Article-1
SourceType-Scholarly Journals-1
ObjectType-Feature-2
content type line 14
ISSN:2332-7731
2332-7731
DOI:10.1109/TCCN.2023.3294754