Contrastive Embedding Multiplexing for Multi-User Semantic Communication Systems

Citations

WEB OF SCIENCE

0
Citations

SCOPUS

0

초록

This letter proposes contrastive embedding multiplexing (CEM), which multiplexes users in semantic communication systems by designing the signal-space geometry of a shared embedding space instead of allocating dedicated physical resources. CEM projects each user onto a soft subspace of that space through user-specific positional masking and refines inter-user separation with a normalized temperature-scaled cross-entropy (InfoNCE) objective trained jointly with the reconstruction loss. Through the alignment-uniformity duality, the contrastive term promotes inter-user near-orthogonality in expectation, a soft counterpart of explicit orthogonalization. Simulations over Rayleigh fading channels with 4 to 16 users show a consistent symbol error rate reduction that increases with user density and exceeds two orders of magnitude at 16 users at single-user inference complexity, corroborated by embedding-geometry, ablation, and text-transmission analyses.

키워드

contrastive learning; learning-based signal processing; multi-user multiplexing; Semantic communication; signal-space methods; Transformer
제목
Contrastive Embedding Multiplexing for Multi-User Semantic Communication Systems
저자
Lee, Ki-Ho; Choi, Hyun-Ho; Lee, Jung-Ryun
DOI
10.1109/LSP.2026.3733303
발행일
2026
유형
Article
저널명
IEEE Signal Processing Letters
권
33
페이지
3806 ~ 3810