Zhijin Qin

[intermediate/advanced] Token-Based Semantic Communications

Summary

Token-based semantic communication is an emerging communication paradigm that moves beyond bit-level fidelity and uses tokens as the basic units of semantic content. Starting from the semantic level identified in classical communication theory, recent work in semantic communication has shown that AI-native systems can transmit meaning, task-relevant information, and multimodal context more efficiently than conventional bit-centric designs in many scenarios. Building on this trend, recent token-centric frameworks treat text, image, speech, and multimodal representations as tokens, and study how these tokens can be selected, compressed, transmitted, recovered, and interpreted under wireless constraints. This tutorial introduces the foundations, representative architectures, and open research problems of token-based semantic communication, with emphasis on tokenization, joint semantic-channel processing, multimodal context utilization, and agent-to-agent communication.

Syllabus

• Motivation and background: From technical communication to semantic communication; why exact bit reconstruction is not always the right objective for intelligent systems.
• Foundations of semantic communication: Core concepts, semantic metrics, and representative early deep semantic communication systems such as DeepSC.
• From modality-specific systems to multimodal systems: Text, speech, and multi-task/multimodal semantic coding and communication, including U-DeepSC and MU-DeepSC.
• Tokenization and semantic representation: Tokens for text, vision, speech, and multimodal data; tokenizer design; discrete token/codebook representations; transformer-based token processing.
• Token-based transmission and coding: Token selection, token importance, joint token/channel coding, joint token coding and modulation
• Agent communication and open problems: Machine-language tokens, cross-modal context, robustness, semantic evaluation, and future research directions.

References

Pre-requisites

Basic knowledge of digital communications, probability, and linear algebra is expected. Familiarity with machine learning and deep learning is recommended. Prior exposure to transformers, representation learning, or multimodal models is helpful, but no prior research background in semantic communication is required.

Short bio

Dr. Zhijin Qin was with Imperial College London, Lancaster University, and Queen Mary University of London, U.K. from 2016 to 2022 as a Research Associate and Lecturer, respectively. She is currently with the Department of Electronic Engineering, Tsinghua University, Beijing, China as a tenured Associate Professor. Her research interests include semantic coding and token communications. She was the recipient of several best paper awards as the first/corresponding author, such as the 2017 IEEE GLOBECOM Best Paper Award, the 2018 IEEE Signal Processing Society Young Author Best Paper Award, the 2021 IEEE Communications Society SPCC Early Achievement Award, the 2022 IEEE Communications Society Fred W. Ellersick Prize, 2023 IEEE Signal Processing Society Best Paper Award, and the 2023 IEEE ICC Best Paper Award. She served as the symposium co-chair of various flagship conferences and the Area Editor of IEEE JSAC series. She is currently serving as an Associate Editor for IEEE Transactions on Wireless Communications, IEEE Signal Processing Magazine, as well as an Area Editor of IEEE Communications Letters.