Adaptive Edge Offloading for Image Classification Under Rate Limit

This article considers a setting where embedded devices are used to acquire and classify images. Because of limited computing capacity, rely on parsimonious classification model with uneven accuracy. When local is deemed inaccurate, can decide offload the image an edge server more accurate but resource-intensive model. Resource constraints, e.g., network bandwidth, however, require regulating such transmissions avoid congestion high latency. investigates this offloading problem when regulation through token bucket, mechanism commonly for purposes. The goal devise lightweight, online policy that optimizes application-specific metric (e.g., accuracy) under constraints bucket. develops based deep $Q$ -network (DQN), demonstrates both its efficacy feasibility deployment devices. Of note fact handle complex input patterns, including correlation in arrivals evaluation carried out by performing over testbed using synthetic traces generated from ImageNet benchmark. Implementation work available at https://github.com/qiujiaming315/edgeml-dqn .