Task-Oriented Multi-User Semantic Communication With Lightweight Semantic Encoder and Fast Training for Resource-Constrained Terminal Devices

This letter studies efficient task-oriented multi-user semantic communication, with resource-constrained terminal devices considered. We design a joint training architecture for semantic encoders and the semantic decoder, denoted as Fed-CL-KD. These encoders are used for semantic information extraction and the decoder is for inference tasks. Through federated learning (FL), the semantic decoder can be trained without uploading local data, while knowledge distillation (KD) can effectively compress the semantic encoder size via knowledge transfer. By integrating curriculum learning (CL), we design a training data reordering method that re-orders the training samples fed to the semantic encoders. By starting with more promising examples, this method achieves faster convergence for both semantic encoders and the semantic decoder. Experimental results demonstrate that the proposed architecture exhibits excellent Top-1 accuracy performance for image classification under three channels, achieving faster convergence while requiring less computing power and memory.