Communication-Aware Dynamic Edge Computing (CONNECT)

物联网（IoT）正在慢慢渗透到我们生活的方方面面；然而，我们离拥有真正智能的物联网还很远。智能传感器不断产生大量数据；例如，一辆自动驾驶汽车预计每秒产生大约1gb的数据，但通常情况下，数据没有被系统地处理、存储或分析，以获得更好的推断。已经开发了许多专门的机器学习（ML）算法来从传感器测量中学习，但这些算法都假设集中设置，其中数据可在具有强大计算能力的中央处理器中获得。这种集中式方法假设将大量传感器数据传输到云中心，由于设备和通道的限制，这可能不可行，不符合大多数应用的严格延迟限制，例如控制自动驾驶汽车或用户的隐私要求。在CONNECT项目中，我们的目标是通过使边缘节点能够以协作的方式快速可靠地做出本地决策，从而开发真正的边缘智能。这将通过开发新的缓存、分布式计算和网络方法来实现，以启用联邦/分布式学习，同时考虑到网络动态和物理通道变化。开发的联合计算、缓存和通信框架将应用于车辆自组织网络（vanet）的分层异构架构。这不仅可以在移动节点之间实现高效可靠的学习，还可以通过将决策限制在当地社区来提高自动驾驶汽车的安全性和隐私性。高速缓存、计算和网络的集成将通过大规模模拟和小规模实施平台进行演示，该平台由两辆汽车和Koc大学的一个路边单元组成。该项目有望实现许多数据密集型边缘应用，从多媒体内容流到移动网络中的参与式数据收集，包括自动驾驶汽车、无人机、移动机器人和移动蜂窝用户。

Internet of things (IoT) is slowly permeating every aspect of our lives; however, we are far from having a truly intelligent IoT. Smart sensors generate massive amounts of data continuously; for instance, an autonomous vehicle is expected to generate about one gigabyte of data per second, but more often than not data is not systematically processed, stored, or analyzed for better inference. Many specialized machine learning (ML) algorithms have been developed to learn from sensor measurements, but these assume a centralized setting, where data is available at a central processor with powerful computation capabilities. This centralized approach assumes that the massive amount of sensor data is transmitted to a cloud center, which may not be feasible due to limitations of the devices and channels, not meet the stringent delay constraints of most applications, e.g., controlling an autonomous vehicle, or the privacy requirements of users. In the CONNECT project, our goal is to develop real edge intelligence by enabling edge nodes to make local decisions rapidly and reliably in a collaborative manner. This will be achieved by developing novel caching, distributed computing and networking methodologies to enable federated/ distributed learning taking into account the network dynamics and physical channel variations. The developed joint computing, caching and communication framework will then be applied to a hierarchical heterogeneous architecture for vehicular ad-hoc networks (VANETs). This will not only enable efficient and reliable learning across mobile nodes, but also improve the security and privacy of autonomous cars by limiting decision making to local neighborhood. Integration of caching, computing and networking will be demonstrated both through large-scale simulations, and on a small-scale implementation platform, consisting of two cars and a roadside unit at Koc University. This project is expected to enable many data intensive edge applications, from multimedia content streaming to participatory data collection in mobile networks, including autonomous cars, drones, mobile robots and mobile cellular users.