Semantic communication is a promising communication paradigm that utilizes Deep Neural Networks (DNNs) to extract the information relevant to downstream tasks, hence significantly reducing the amount of transmitted data. In current practice, the semantic communication transmitter for a specific task is typically pre-trained and shared by all users. However, due to user heterogeneity, it is desirable to use different transmitters according to the available computational and communication resources of users. In this paper, we first show that it is possible to dynamically adjust the computational and communication overhead of DNN-based transmitters, thereby achieving adaptive semantic communication. After that, we investigate the user association and resource allocation problem in a multi-cell network where users are equipped with adaptive semantic communication transmitters. To solve this problem, we decompose it into three subproblems involving the scheduling of each user, the resource allocation of each base station (BS), and the user association between users and BSs. Then we solve each problem progressively based on the solution of the previous subproblem. The final algorithm can obtain near-optimal solutions in polynomial time. Numerical results show that our algorithm outperforms benchmarks under various situations.

翻译：语义通信是一种利用深度神经网络提取与下游任务相关信息的新兴通信范式，可显著减少传输数据量。当前实践中，针对特定任务的语义通信发射机通常采用预训练方式，并由所有用户共享。然而，由于用户存在异质性，根据用户可用的计算与通信资源采用不同的发射机更为理想。本文首先证明，基于深度神经网络的发射机可动态调整其计算与通信开销，从而实现自适应语义通信。随后，我们研究了配备自适应语义通信发射机的多小区网络中的用户关联与资源分配问题。为解决该问题，我们将原问题分解为三个子问题：每个用户的调度问题、每个基站的资源分配问题、以及用户与基站间的关联问题，并基于前一子问题的解逐步求解各子问题。所提算法可在多项式时间内获得接近最优的解。数值结果表明，该算法在多种场景下均优于基准方案。