Asymmetric Weight Pruning for Resource-Limited IoT Devices in Semantic Communications

In this paper, we explore the application of semantic communication in Internet of Things (IoT) networks, taking into account the limited computing resources of IoT devices. We propose a resource-efficient asymmetric autoencoder framework that leverages weight pruning and quantization techniques to compress the encoder model on IoT devices. Our simulation results demonstrate that the proposed encoder-only pruning model achieves up to 98% model compression without any performance degradation. Furthermore, even with 8-bit quantization applied, the pruned model exhibits no performance loss.