Revisiting Outage for Edge Inference Systems
A new research paper introduces the concept of "inference outage" (InfOut) probability to better evaluate the reliability of edge inference systems. This metric quantifies the likelihood that end-to-end inference accuracy falls below a target threshold, addressing limitations of existing communication-centric reliability measures. The proposed framework establishes a trade-off between communication overhead and inference reliability, using a Gaussian approximation for optimization. Experimental results suggest this approach offers superior end-to-end inference reliability compared to traditional methods. AI
IMPACT Introduces a new metric to improve the reliability of AI models deployed on edge devices.