Goal-Oriented Multi-Task Semantic Communication for Efficient Wireless Transmission under Dynamic Channels in a Resource Constrained Environment
February 19 @ 6:00 pm - 7:01 pm
Emerging latency-sensitive applications such as autonomous driving, immersive extended reality, and large-scale sensing are expected to impose more stringent and heterogeneous requirements on beyond-5G/6G communication systems. These applications simultaneously demand ultra-low
latency, high reliability, and task-specific semantic fidelity under tight bandwidth and computational constraints, which are challenging to address with conventional bit level communication paradigms. Additionally, as current wireless communication systems approach the Shannon capacity limit, coupled with the explosive growth of connected devices with data-intensive applications, making further improvements in transmission efficiency is becoming increasingly critical for future wireless networks. In recent years, semantic communication (SemCom) has emerged as a promising paradigm to bridge this gap by transmitting semantic information (SI) that is most relevant to the downstream task or communication goal and discarding task-irrelevant information to improve transmission efficiency and reduce communication overhead. However, due to the complex task objectives and dynamic environments, the unification of SemCom into current wireless networks is plagued by a plethora of challenges, including but not limited to meeting the diverse needs of heterogeneous users in dynamic conditions and intelligent resource allocation to ensure efficient utilization especially in resource constrained scenarios. Accordingly, this thesis develops novel SemCom frameworks to address these <a href="http://challenges.
Firstly” target=”_blank” title=”challenges.
Firstly”>challenges.
Firstly, to address the challenges posed by diverse channel conditions, changing user task performance demands, and limited computational capabilities amongst users, this thesis proposes CAAR-MTSC (Context-Aware, Adaptive-rate, Multi-Task Semantic Communication), a
deep learning-enabled SemCom architecture that supports concurrent downstream tasks. Central to the design is a Semantic Mask Module (SMM) with its two-fold functionality to preserve task-salient SI and discard redundant SI (semantic compression) depending on the dynamic channel conditions and the user’s perceptual need. We then introduce a novel composite rate–distortion– perception (RDP) objective that finds the tradeoff between perceptual quality and data reconstruction. Then, a knowledge distillation (KD) paradigm is introduced to train a lightweight receiver for resource-constrained devices, while relatively retaining the high downstream task performance of the larger computationally intensive receiver. Simulation results prove that the proposed framework is able to achieve high concurrent multi-task accuracy under dynamic channel conditions with changing user needs on resource-constrained <a href="http://devices.
Secondly” target=”_blank” title=”devices.
Secondly”>devices.
Secondly, to address the combinatorial challenges faced with heterogeneous users with diverse multi-tasking and latency needs in a resource constrained wireless network, we introduce a multi-task multi-user semantic communication system to allocate wireless resources and select appropriate degree of semantic compression to meet the needs of heterogeneous users. A novel Segmentation-Guided Dual-Branch Encoder for Multi-Task Semantic Communication (SegDB-MTSC) architecture is proposed to individually extract SI from regions of interest (ROI) and regions of non-interest (RONI). A latency-aware multi-task semantic utility (LAMTSU) term is derived to maximize the comprehensive system performance for diverse users, including their multi-task performance and their latency constraints. A deep reinforcement learning (DRL) algorithm is then used to perform resource allocation to maximize long term system utility. Simulation results illustrate the superiority of the proposed scheme to meet the latency and task needs of <a href="http://users.
Virtual:” target=”_blank” title=”users.
Virtual:”>users.