As the rapid evolution of smart devices and real-time applications, many new kinds of computation-intensive services have been emerged successively and the corresponding requirements have been growing dramatically. Extended from cloud computing, mobile edge computing (MEC) is a novel technology which can provide powerful computing resource at the proximity of resource-restrained mobile devices. Thus, it enables collaboration between edge server and mobile device, which can improve the quality of experience for users. In this article, we propose an intelligent collaborative inference (ICI) approach for real-time computation-intensive services in MEC network, which can achieve intelligent service partitioning and partial task offloading. Since machine learning algorithms have been applied in many applications with the advancement of big data and computing power, we focus on the services based on deep-learning. Particularly, we research a service based on Pose-Net model to achieve human pose estimation in the field of computer vision. And we design relevant ICI algorithm to achieve fine-grained video stream processing in consideration of video service requirement, deep neural network (DNN) model structure, mobile device capability, wireless network condition, and cooperative server workload. Based on Python programming language and TensorFlow library, we test the ICI approach with some practical simulation parameters on real hardware platforms. The experiment results show that the presented ICI approach have superior performance in terms of service frame rate and client energy consumption than other benchmark approaches.

中文翻译：

---

移动边缘计算网络中基于深度学习的服务划分和任务卸载的智能协同推理方法

随着智能设备和实时应用的快速演进，许多新的计算密集型服务相继出现，相应的需求也急剧增长。移动边缘计算（MEC）是从云计算扩展而来的一种新技术，它可以在资源受限的移动设备附近提供强大的计算资源。因此，它可以实现边缘服务器和移动设备之间的协作，从而可以提高用户的体验质量。在本文中，我们提出了一种针对 MEC 网络中实时计算密集型服务的智能协同推理 (ICI) 方法，该方法可以实现智能服务划分和部分任务卸载。由于随着大数据和计算能力的进步，机器学习算法已经应用在许多应用中，因此我们专注于基于深度学习的服务。特别是，我们研究了一种基于 Pose-Net 模型的服务，以实现计算机视觉领域的人体姿态估计。并综合考虑视频业务需求、深度神经网络（DNN）模型结构、移动设备能力、无线网络状况、协同服务器工作负载等，设计相关ICI算法实现细粒度视频流处理。基于 Python 编程语言和 TensorFlow 库，我们在真实的硬件平台上使用一些实用的仿真参数来测试 ICI 方法。