Revisiting Outage for Edge Inference Systems

The article presents a theoretical framework for edge inference systems in 6G networks, introducing the concept of inference outage (InfOut) probability to quantify the risk of sub-threshold end-to-end (E2E) inference accuracy. This framework highlights the tradeoff between communication overhead and inference reliability, enabling optimization under E2E latency constraints. Experimental results indicate that this approach outperforms traditional communication-focused methods, which is crucial for developing reliable AI-driven applications in latency-sensitive environments like autonomous driving and industrial automation.