Dora: Online Scheduling of Dual-Model Inference and Retraining for Multi-View Edge Systems

Combining observations from several viewpoints can improve prediction quality, but running these pipelines at the edge creates competing demands on scarce computing capacity. Changing input conditions also affect the feature extraction and fusion models differently, making coordinated retraining difficult. Dora estimates how training choices and resource allocations influence accuracy, then uses these estimates to select execution and retraining configurations as requests arrive. It breaks a long-horizon scheduling objective into decisions for individual time slots, balancing immediate responsiveness with accuracy over time. The scheduler comes with theoretical performance guarantees. Experiments on an edge testbed report accuracy gains of up to 40% and latency reductions of up to 60% relative to the evaluated baselines.

Recommended citation: Wenxiao Liu, Ruiting Zhou, Zhenyu Xu, Kun Tian, Yifan Zeng, and Chengyang Wang, "Dora: Online Scheduling of Dual-Model Inference and Retraining for Multi-View Edge Systems," IEEE Transactions on Mobile Computing, early access, September 1, 2026, pp. 1-13, doi: 10.1109/TMC.2026.3731470.
Download Paper