Abstract
Task-oriented semantic communications represent a novel paradigm designed to optimize performance for specific semantic tasks. Traditional approaches often extract common semantic features to balance transmission rate and performance distortion across multiple tasks. However, these methods tend to underperform on individual tasks and struggle to adapt to dynamically changing task requirements. To address these short-comings, we propose a task-adaptive semantic communication system with feedback (TASC-f), which leverages a unified, end-to-end adaptable codec model capable of handling diverse tasks. Specifically, we formulate the task-adaptive semantic communication problem as a conditional rate-distortion optimization, incorporating task-specific semantic prompts as side information. We derive an analytical solution to minimize this optimization problem and introduce a classical iterative algorithm to compute the rate-distortion function. Building on this solution, we design a trainable semantic prompt module (SPM) that dynamically generates side information reflecting task-specific features. To further enhance adaptability, the system integrates a feedback mechanism that enables real-time adjustments to the encoding strategy in response to changing task demands. This unified network architecture supports variable transmission rates and seamlessly adapts to different channel conditions within a single model. Experimental evaluations across two distinct scenarios involving three semantic tasks demonstrate the superior performance of TASC-f. Not only does the proposed system outperform competing data-driven communication methods, but it also matches the performance of dedicated single-task systems for fixed tasks and surpasses multi-task communication systems when adapting to new tasks.
Funding Information
  • National Natural Science Foundation of China (62571159)
  • National Natural Science Foundation of China (61871147)
  • National Key Research and Development Program of China (2020YFB1806403)
  • Major Key Project of PCL (PCL2024A01)