Collaborative edge and cloud learning: Potentials and solutions

今天的个人计算设备产生和贡献的数据，从文本、语音到图片和视频，都是巨大的，包含着值得发现的宝贵信息。深度神经网络的最新进展在探索其中隐藏的信息方面显示出巨大的潜力。深度学习依靠强大的计算能力来处理海量数据，这是由机器集群或更一般的现代数据中心提供的典型功能。虽然以云为中心的学习对数据中心中可用的数据很有效，但从全球来源收集数据不可避免地会产生高流量，更重要的是延迟，这对相机网络中的人脸识别和人类跟踪等实时学习应用提出了挑战。边缘计算的概念最近被认为是对云计算的补充。它将应用程序、数据和计算内容从集中式数据中心推送出去。因此，它可以通过减少传输到云的流量和各种深度学习应用程序的推理来显著加快培训过程。华为是构建用于边缘计算的通信和计算基础设施的行业先驱，并已开始建设其公共云基础设施。在这个项目中，我们将共同了解协作边缘和云学习的最新发展。我们将确定如何将数据预处理和特征提取推向网络边缘的机会和挑战，然后我们将开发新的解决方案，最大限度地减少网络流量和推理延迟，同时保持深度学习所需的准确性。

The data generated and contributed by today's personal computing devices, ranging from text, voice, to pictureand video, are enormous and contain invaluable information worth discovering. Recent advances in deep neuralnetworking have shown great potentials in exploring the hidden information therein. Deep learning relies onstrong computation power to process the massive amount of data, which is typical offered by machine clusters,or more general, modern data centers. While cloud-centric learning works well for the data that are available indatacenters, gathering the data from worldwide sources unavoidably incurs high traffic and, more importantly,latency that challenges such realtime learning applications as face recognition and human tracking in cameranetworks. The concept of edge computing has been recently advocated as a complement to cloud computing. Itpushes applications, data, and computing content away from the centralized data centers. As such, it cansignificantly accelerate the training process by reducing the traffic transferred to the cloud and the inferencelatency for a broad spectrum of deep learning applicationsHuawei is an industrial pioneer in building the communication and computation infrastructure for edgecomputing, and has started building its public cloud infrastructure as well. In this project, we will worktogether to understand the state-of-the-art of collaborative edge and cloud learning. We will identify theopportunities and challenges on how to push data pre-processing and feature extracting to the network edge.We will then develop novel solutions that minimize the network traffic and inference latency yet with desiredaccuracy for deep learning.