Collaborative Learning for Task-Oriented Semantic Communications: Overcoming Data Mismatch Between Transceivers

doi:10.1109/OJCOMS.2025.3586462

Collaborative Learning for Task-Oriented Semantic Communications: Overcoming Data Mismatch Between Transceivers

2025-01-01

0

OA

AI

DOI：10.1109/OJCOMS.2025.3586462

Original

Original request for help

Save

Abstract

En 中文

Task-oriented semantic communication (ToSC) enhances efficiency and performance by leveraging task-specific data representations and end-to-end learning, which are more compact and effective than traditional reconstruction-oriented methods. However, current deep learning (DL)-based ToSC schemes heavily rely on empirical data, with the assumption that both the source and destination transceivers are trained using identical empirical data. In practical scenarios, the distribution of transmitted data can differ significantly from that of the empirical data at the destination. Such discrepancies may lead to the destination’s inability to accurately interpret the received semantic information, consequently impacting task performance. To address this challenge, we propose a Transceiver Collaborative Learning-aided Semantic Communication (TCLSC) framework. This framework facilitates the collaborative training of semantic encoders and decoders by transceivers, leveraging their local datasets, and periodical sharing parameters of the semantic encoder and decoder. To reduce the communication overhead caused by collaborative learning, we innovatively employ a model update quantization mechanism. Subsequently, we analyze the convergence of TCLSC and propose an adaptive scheme for determining the frequency of transceiver collaborative learning. Furthermore, considering that transceiver A may communicate simultaneously with transceiver B having mismatched data and transceiver C with identical data, we propose a personalized TCLSC (P-TCLSC) framework based on prompt learning. This approach aims to solve the problem of data mismatch while preserving the personalization of each transceiver. Extensive experiments demonstrate that, compared with the baseline methods, the proposed approaches are more adept at tackling data mismatch issues between the source and destination, thereby ensuring robust communication performance.

Keywords:

Task-oriented semantic communications

transceiver cooperative learning

data discrepancy

model quantization

prompt learning

Collaborative Learning for Task-Oriented Semantic Communications: Overcoming Data Mismatch Between Transceivers

Abstract

Journal

Organization