<ul><li>This paper introduces an adaptive framework for edge inference using a transformer-powered deep joint source channel coding architecture.</li><li>The approach allows efficient transmission of essential features for object detection in resource-constrained edge devices under varying network conditions.</li><li>Input data is tokenized into high-level semantic representations, refined by a transformer, and transmitted over noisy wireless channels.</li><li>The system incorporates a semantic token selection mechanism and a resource allocation algorithm to improve robustness and performance in dynamic network scenarios.</li></ul>

Adaptive Semantic Token Communication for Transformer-based Edge Inference

Discover more