Biomimetic Semantic Communication with LLM Integration for IoT Systems

Traditional communication systems focus on trans-mitting data accurately rather than conveying the meaning. In this work, we propose a biomimetic semantic communication for internet of things (IoT) systems that communicate meaning rather than raw data. The nature-inspired semantic encoder running a quantized neural network (QNN) generates semantic descriptions for the events occurring on live video streams. These semantics are transmitted through a low-cost transceiver on a 600 kbps link. The semantic decoder generates the prompt from the received description, to reconstruct the event in several modalities, such as text, audio, image, and video. The decoder is integrated with large language models (LLMs), which recreates the event. The system’s robustness is measured using a simple metric, the semantic score, which is the ratio of correctly received letters to the total number of letters transmitted. Experiment results show that the receiver can reliably generate a prompt to reconstruct video content with up to 16% packet loss, while the image reconstruction success rate is up to 19% loss. Beyond 20% loss, direct text display is recommended, unless the system is augmented with advanced text-recovery techniques. This approach is a step towards energy-efficient communication among IoT devices, bridging the gap between machine perception and human interpretation.