Channel-aware Partial Fine-Tuning for AirPooling based Semantic Communications

Recently, multiple devices in the Internet of Things collaborate to perceive the environment, offering richer information for intelligent decisions. However, due to the limited network resources, the aggregation of massive sensory data over the dynamic channel environments presents a critical challenge to the real-time learning performance. In this work, we propose a channel-aware partial fine-tuning scheme for the AirPooling based semantic communication system, enabling rapid delivery of accurate models in dynamic channel environments. To achieve efficient information aggregation, Over-the-Air Pooling (AirPooling) is leveraged to facilitate the pooling of multiple single-view features during concurrent transmission, with channel encoding and decoding performed via autoencoder-based neural networks. In addition, to address the performance degradation caused by the mismatch between the channel conditions during inference and those during pre-training, a partial fine-tuning method is proposed, which freezes part of the network parameters to reduce computation overhead while maintaining real-time performance. Simulation results show that in dynamic channel environments, the proposed scheme achieves high learning performance with lower fine-tuning overhead compared to baselines.