Ultra-Low-Latency Edge Inference for Distributed Sensing (2407.13360v1)

Abstract: There is a broad consensus that AI will be a defining component of the sixth-generation (6G) networks. As a specific instance, AI-empowered sensing will gather and process environmental perception data at the network edge, giving rise to integrated sensing and edge AI (ISEA). Many applications, such as autonomous driving and industrial manufacturing, are latency-sensitive and require end-to-end (E2E) performance guarantees under stringent deadlines. However, the 5G-style ultra-reliable and low-latency communication (URLLC) techniques designed with communication reliability and agnostic to the data may fall short in achieving the optimal E2E performance of perceptive wireless systems. In this work, we introduce an ultra-low-latency (ultra-LoLa) inference framework for perceptive networks that facilitates the analysis of the E2E sensing accuracy in distributed sensing by jointly considering communication reliability and inference accuracy. By characterizing the tradeoff between packet length and the number of sensing observations, we derive an efficient optimization procedure that closely approximates the optimal tradeoff. We validate the accuracy of the proposed method through experimental results, and show that the proposed ultra-Lola inference framework outperforms conventional reliability-oriented protocols with respect to sensing performance under a latency constraint.