RTML: Large: Continuous Adaptation for Decision Streams

RTML：大：决策流的持续适应

• 批准号: 1937301
• 负责人: Oyekunle Olukotun
• 金额: $ 150万
• 依托单位: Stanford University
• 依托单位国家: 美国
• 项目类别: Standard Grant
• 财政年份: 2019
• 资助国家: 美国
• 起止时间: 2019-10-01 至 2022-09-30
• 项目状态: 已结题
• 来源: https://www.nsf.gov/awardsearch/showAward?AWD_ID=1937301&HistoricalAwards=false
• 关键词: RTML; Large; Continuous; Adaptation; Decision

Systems that can efficiently make real-time decisions based on large-scale data streams will impact broad areas of daily life, including autonomous vehicles, personal assistants, medicine, and fraud detection tools. Such systems have become critical as machine learning is increasingly tasked with making richer decisions over constantly-changing large data streams for both consumer and industry applications. This project seeks to develop hardware-software systems capable of making such real-time decisions over large data streams while flexibly and continuously adapting to changes in their environment. This project will also support redesign of courses on hardware accelerators and parallel computing at Stanford University, with large-scale data streaming systems as a central driver. These courses are designed to provide students with a sufficiently strong background to engage in systems and machine learning research, thus enabling a diverse and much desired US workforce in an area of technology of current importance.This project will create tools and techniques for large-scale data streaming systems by producing innovations spanning software, hardware, and machine learning. Modern machine learning requires a vast amount of labeled data, and streaming scenarios only increase this requirement. To handle the need for more data, techniques for automatically labeling data (and in particular, temporal data) under real-time constraints will be developed. Because the environment for data streaming systems is constantly changing, the proposed project seeks to continuously adapt and specialize models to the current environment, leading to vastly improved efficiency. Real-time data streaming systems will require hardware that achieves both exceptional efficiency, as well as provides sufficient flexibility to support both real-time model training and inference; this project will develop such hardware. These innovations will be demonstrated and evaluated on two applications: next-generation video stream processing in autonomous vehicle, medical and industrial domains and smart routers for networking systems. The project will also collaborate with a synergistic DARPA program for related hardware development.This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.

能够基于大规模数据流高效地做出实时决策的系统将影响日常生活的广泛领域，包括自动驾驶汽车、个人助理、医药和欺诈检测工具。随着机器学习越来越需要对消费者和行业应用程序不断变化的大型数据流做出更丰富的决策，这样的系统变得至关重要。该项目旨在开发硬件-软件系统，该系统能够在灵活和持续地适应环境变化的同时，对大型数据流做出这样的实时决策。该项目还将支持斯坦福大学硬件加速器和并行计算课程的重新设计，以大规模数据流系统为核心驱动因素。这些课程旨在为学生提供从事系统和机器学习研究的足够强大的背景，从而使美国劳动力在当前重要的技术领域具有多样化和渴望的能力。该项目将通过产生横跨软件、硬件和机器学习的创新，为大规模数据流系统创造工具和技术。现代机器学习需要大量的标签数据，而流媒体场景只会增加这一要求。为了处理对更多数据的需求，将开发用于在实时约束下自动标记数据(特别是时间数据)的技术。由于数据流系统的环境不断变化，拟议的项目寻求不断调整和专业化模型以适应当前环境，从而极大地提高效率。实时数据流系统将要求硬件既能实现卓越的效率，又能提供足够的灵活性来支持实时模型训练和推理；本项目将开发这样的硬件。这些创新将在两个应用上进行演示和评估：自动驾驶汽车、医疗和工业领域的下一代视频流处理以及网络系统的智能路由器。该项目还将与DARPA相关硬件开发的协同计划合作。该奖项反映了NSF的法定使命，并通过使用基金会的智力优势和更广泛的影响审查标准进行评估，被认为值得支持。