Attention-Aware Neuromorphic Semantic Communications

Spiking neural networks (SNNs) are characterized by their high energy efficiency and event-driven signal processing, which is gaining increasing research interest as a new computational framework for semantic communication systems in 6G and beyond. This paper introduces Att-NeuroSC, an end-to-end semantic communication system based on SNNs. The proposed system utilizes SNNs for joint source and channel coding, efficiently extracting and encoding semantics using a minimal number of spikes, thereby significantly reducing energy consumption and bandwidth costs. Furthermore, to compensate for the performance loss caused by sparse computation, we introduce a novel attention mechanism that enables the network to focus on important semantics in both temporal and spatial dimensions. Additionally, we introduce a spiking rate loss to trade-off the performance and costs. The experiments on representative neuromorphic datasets show that our proposed systems outperform conventional schemes in terms of task performance, energy efficiency, and inference latency. The ablation study demonstrates the effectiveness of the introduced attention modules.