随着现代物联网系统越来越趋向于大型化、复杂化和智能化，对边缘处计算设备的感知能力和计算能力提出越来越高的要求，使协同低功耗高实时的边缘计算与低成本高性能的云计算成为计算智能发展的未来。为实现边云之间的“协同”，实现边缘计算与云计算应用价值的最大化，提出了一种边云系统中面向边缘模型情景任务自适应的计算框架。首先，以云侧模型推理与边缘侧模型推理间的差异性为基础，构建了边云协同机制模型；然后，采用少量边缘应用侧任务样本对迁移的云侧模型进行增量学习，提升云侧模型到边缘侧模型迁移的自适应性能；最后，实时监测边侧推理结果，实现边侧模型性能的保持。

As modern IoT systems tend to become increasing larger, complex and intelligent, which makes requirements become higher on the perception and computation capabilities of calculation devices at the edge. The future development of cloud computation focuses on enabling the collaboration of low-power, high-real-time edge computation with low-cost and high-performance. In order to realize the collaboration between edge and cloud to maximize the application value of the advantages of edge and cloud computation, this paper proposed a computing framework for edge model scenario adaptive tasks in edge-cloud system. Firstly, the edge-cloud collaboration mechanism model is constructed based on the difference between inference results of cloud-side model and edge-side model. Secondly, a small amount of edge application-side task samples are employed to incrementally re-train the cloud-side model to the adaption performance of edge-side transferred model from the cloud-side model. Finally, the performance of the edge-size can be maintained by the real-time monitoring of the side-side reasoning results.

国家自然科学基金资助项目