Bandwidth Management in Semantic Communications: A Tradeoff Between Data Sensing and Transmission

Semantic communication (SemCom) is a novel paradigm that exploits sophisticated deep learning tools to distil semantic features from source data at transmitter and recover the meaning of the source data based on these semantic features at the receiver. To train proficient semantic coding models with minimum semantic ambiguity, it is of paramount importance to sense abundant data for both model training and background knowledge construction. However, if interminable sensing is performed for getting highly capable SemCom coding models, a large amount of bandwidth resources should be occupied, which may reduce semantic spectrum efficiency (SSE). In this paper, we investigate how to achieve an efficient tradeoff between data transmission and data sensing in SemCom. Specifically, we first formulate an optimization problem to jointly allocate bandwidth for sensing and data transmission with the objective of maximizing SSE, while subject to resource budgets and service quality requirements. To solve this problem, we start with a simplified resource allocation policy where bandwidth is equally allocated to all users. Inspired by the solution to equal allocation, we then derive the bandwidth allocation policy under a generic non-equal allocation scenario, based on the projected gradient method considering the variables sensing bandwidths and data transmission bandwidths are interdependent. In addition, we extend to the scenario of multi encoder models, and divide the optimization into two tractable subproblems. Furthermore, numerical simulations verify the effectiveness and superior performance of our proposed algorithm on the SSE, compared with several baselines.