机器学习任务目前已经成为边缘计算网络中的主要负载之-现有的边缘计算和云计算网络中的资源调度器通常按用户指定的资源数划分，并没有很好地结合机器学习任务对资源的敏感性。我们在这篇工作中设计了一个自优化的资源调度器，可同时优化任务的部署和细粒度的资源调度。为了学习机器学习任务的资源敏感程度的不确定，我们使用了多智能体强化学习的基本框架，在此基础上为我们的资源调度问题定制化地设计了多个模块。此外，我们还融合了图神经网络来学习任务与边缘服务器等网络重要部分之间的关系。我们设计的机制较现有的较先进的调度算法可实现更低的任务完成时间。

Congestion management, including congestion control and flowcontrol, plays a critical role in the performance of datacenternetworks. Due to the growth of port bandwidth and the scale ofdata center applications, new opportunities appear in this classicresearch area. A large number of researches are published recentlyIn this talk, l will first briefly review the status quo.l will thenpresent our recent progresses in this field, which stick to our ownmotto "Congestion Control is Dead, Long Live Flow Control"At last l will talk about possible future directions of congestionmanagement, especially those beyond datacenter networks, e.g.,inter-datacenter andintra-routernetworks.

We study a fair resource scheduling problem, where a set of interval jobsare to be allocated to heterogeneous machines controlled by intellectualagents.Each job is associated with release time, deadline, and processingtime such that it can be processed if its complete processing period isbetween its release time and deadline. The machines gain possibly differentutilities by processing different jobs, and all jobs assigned to the samemachine should be processed without overlap.We consider two widelystudied solution concepts, namely, maximin share fairness and envyfreeness.For both criteria, we discuss the extent to which fair allocationsexist and present constant approximation algorithms for various settings

Artificial intelligence models will continue to be pervasively deployed to support diverse intelligent Internet of Things (IoT) applications in the 5G/6G era. Many such applications rely on deep neural networks (DNN) for object classification. In this presentation, DNN inference uses a pre-trained DNN model to process an input data sample such as raw sensing data, and generates a classification result. We will discuss when to offload DNN inference computation from resource constrained IoT devices to the edge and how to incorporate different contributions from multiple random DNN inference results to improve task classification accuracy, while achieving high transmission, computation, and energy resource utilization.

Wireless sensing, which applies wireless signals to sense objects, events, or phenomenon without physical contact, has been widely applied for many applications (e.g., gesture recognition, vital sign monitoring, localization). Though promising, existing wireless sensing methods mostly work in controlled environments. Enabling wireless sensing for complicated environments in industrial and agricultural applications still faces many challenges. In this talk, two of our wireless sensing systems designed respectively for industrial and agricultural applications will be presented. The first part of this talk focuses on mRotate, a novel and practical mmWave radar-based rotation speed sensing system that enables accurate rotation speed sensing from a safe distance, and is robust to the illumination conditions and the target object’s light reflectivity.

The recent success of machine learning, especially deep learning, stems from the availability of big data and strong computational power brought by cutting-edge hardware (e.g., GPUs and TPUs). Facing massive computational loads, it is inefficient or sometimes impossible to train models on a single machine, driving attention towards distributed machine learning on multiple machines. This talk reviews the recent research efforts about distributed systems for machine learning. It starts from data centers, and present concepts and motivations of distributed machine learning, paying special attention on machine learning job scheduling on heterogeneous GPUs. Then, it walks out of data centers to show how to extend the distributed machine learning ideas to edge devices. Finally, it presents new challenges of training distributed unstructured data, graphs, and possible solutions.

We study a fair resource scheduling problem, where a set of interval jobs are to be allocated to heterogeneous machines controlled by intellectual agents.Each job is associated with release time, deadline, and processing time such that it can be processed if its complete processing period is between its release time and deadline. The machines gain possibly different utilities by processing different jobs, and all jobs assigned to the same machine should be processed without overlap.We consider two widely studied solution concepts, namely, maximin share fairness and envy-freeness.For both criteria, we discuss the extent to which fair allocations exist and present constant approximation algorithms for various settings.

随着数据成为生产要素，发展数字经济上升为国家战略，数据安全可信流通是数字经济发展的迫切需求，而数据权属困境、监管合规困境与技术标准困境，导致数字经济中主体不愿共享、不敢共享和不能共享，为应对上述困境，本讲座重点介绍以区块链与隐私计算作为可信数字底座，实现数据所有权和使用权分离， 构建“数据可用不可见，用途可控可计量”的可信流通模式，重点讨论链下扩容技术以及以及离线支付等关键问题。

随着广告技术的演进，程序化交易广告愈发成为主流。作为全球顶尖的程序化广告技术公司之一，萃弈（The Trade Desk）利用自动化流量采买、优化竞价、实时监测等方面，致力于让广告主在合适的时间，将合适的广告信息投放给合适的目标群体。本次报告将详细介绍如何基于全球海量数据，使用数据驱动的方式，建立机器学习模型预估点击或行动并指导实时竞价，以优化广告投放效果。

在如今的数字时代，激活数据要素潜能，以数字化转型整体驱动生产方式、生活方式和治理方式的变革已经成为主流趋势并进入了快速发展阶段，数字化应用场景和数字化信息技术已经贯穿了包含政府、企业、城市建设在内的众多行业。然而，在数字经济发展浪潮中，传统行业的数字化依然面临着较为明显的数字鸿沟，本报告将结合浪潮云实践从新型基础设施、数据要素和治理保障体系三个层面阐述浪潮云如何助力行业跨越数字化鸿沟。