CAREER: Toward A Machine Learning Framework for the Internet of Things

该项目将开发一种新的机器学习范式和工具，以应对物联网中大规模、地理分布的数据。关键的创新是利用物联网网络本身的设备和计算能力以可扩展、可靠的方式执行数据分析。物联网描述了一个设备网络，从RFID标签到智能温控器，再到灯泡，可以感知和交流信息。据预测，到2020年，物联网设备将达到250亿至500亿台。这一庞大的网络及其生成的数据将在环境管理、智能基础设施和医疗保健等广泛的关键领域实现新的应用。要实现这一愿景，能够快速分析生成的海量数据并从中学习至关重要。目前的大数据分析方法需要将数据完全传输到具有强大计算能力的平台，如云。考虑到设备数量和由此产生的数据生成率的预计爆炸性增长，这是不可行的。该研究在三个主要方面整合了机器学习、分布式计算和网络系统的工具和理论，包括：为算法设计和实现提供抽象的计算框架，该框架足够灵活地支持广泛的机器学习方法；框架实现，通过屏蔽物联网的设备异构性、设备故障和网络动态为算法开发人员提供稳定的平台；以及开发和实现适应网络和计算的技术，以支持算法执行并提供性能保证。一个重大贡献将是该框架的开放源码软件实施，允许其他网络专业知识有限的人开发自己的物联网应用程序。该项目还包括强有力的教育和外联部分，包括研究生和本科生研究、课程开发以及促进和支持妇女参与计算机科学的活动。

This project will develop a new paradigm and tools for machine learning that can cope with the massive-scale, geographically distributed data in the Internet of Things. The key innovation is to use devices and computing power within the Internet of Things network itself to perform data analysis in a scalable, reliable fashion.The Internet of Things describes a network of devices, from RFID tags, to smart thermostats, to light bulbs, that can sense and communicate information. It is predicted that by 2020, there will be 25 to 50 billion devices in the Internet of Things. This massive network and the data it generates will enable new applications in a wide range of critical domains including environmental management, smart infrastructure, and healthcare. To achieve this vision, it is crucial to be able to quickly analyze and learn from the massive amount of generated data. Current approaches for big data analytics require full data transfer to a platform with large computational power, such as the cloud. Given the projected explosion in the number of devices and the resulting data generation rate, this is not feasible. The proposed research integrates tools and theory from machine learning, distributed computing, and networked systems in three main thrusts that include; a computational framework that provides an abstraction for algorithm design and implementation that is flexible enough to support a wide collection of machine learning methods, a framework implementation that provides a stable platform for algorithm developers by masking device heterogeneity, devices failures, and the network dynamics of the Internet of Things, and development and implementation of techniques to adapt the network and computation to support algorithm execution with performance guarantees.The framework developed in this project will facilitate rapid development and deployment of Internet of Things applications. A significant contribution will be an open source software implementation of the framework allowing others with limited network expertise to develop their own applications for the Internet of Things. The project also includes robust educational and outreach components including graduate and undergraduate research, curriculum development, and activities to promote and support the participation of women in computer science.