Secure, Real-Time Decisions on Live Data

根据实时数据做出安全、实时的决策

• 批准号: 1730628
• 负责人: Ion Stoica
• 金额: $ 1000万
• 依托单位: University of California-Berkeley
• 依托单位国家: 美国
• 项目类别: Continuing Grant
• 财政年份: 2018
• 资助国家: 美国
• 起止时间: 2018-03-01 至 2024-02-29
• 项目状态: 已结题
• 来源: https://www.nsf.gov/awardsearch/showAward?AWD_ID=1730628&HistoricalAwards=false
• 关键词: Secure; Real; Time; Decisions; Live

A new era is rising in which AI systems will play an increasingly central role in people's lives. These systems will revolutionize healthcare through early identification of patients at risk, cell-level diagnosis and treatment using nanoprobes, and robotic surgery. They will reduce traffic congestion and help eliminate fatalities by powering autonomous vehicles and unmanned drones. And, they will make businesses safer by detecting and defending in real-time against financial fraud and internet attacks. More generally, these systems will transform how people sense and interact with the surrounding world making it more adaptive and responsive to our needs. In order to fulfill this vision, a new generation of AI systems is needed to power mission-critical applications where human safety and well-being are at stake, and can work in adversarial environments that change continually and unexpectedly. Besides being intelligent, these decision systems need to address four challenges. First, they must react in real-time (i.e., making decisions in seconds or even milliseconds) to support applications such as robotic surgery and self-driving cars. Second, AI systems need to learn continually on live data streams as their environments evolve chaotically. Third, these systems need to be secure, i.e., ensure privacy, data confidentiality, and decision integrity. Finally, as these systems make decisions on behalf of humans, their decisions need to be explainable to someone with limited understanding of AI. For example, if an AI system diagnoses a patient with a rare disease or deems a certain test unwarranted, the system should provide an explanation in terms of the patient's history and that of the larger population, and not point to the AI algorithm's internal computations. The goal of this Expedition project is to build AI decision systems to address these challenges by developing open source platforms, tools, and algorithms for Real-time, Intelligent, Secure, and Explainable (RISE) decisions. Achieving this goal requires a holistic approach that combines AI, security, systems, and hardware research. For example, to successfully deploy a fleet of delivery robots in a crowded city requires not only advances in AI (e.g., the ability to perceive and safely navigate complex urban environments), but also advances in systems (e.g., new hybrid edge-cloud systems able to coordinate vehicles in real-time), security (e.g., ensure the information collected by robots' sensors does not compromise customer's privacy), and computer architecture (e.g., hardware and software co-design to reduce power consumption and improve security). The RISE project aims to empower a large community of pioneers to build innovative applications and solutions based on the tools and ideas it will create, and broaden research participation, allowing students and researchers across many disciplines to contribute and build on its artifacts. Building and fostering a community around a common open platform for AI systems will enable the next decade of innovation centered around widespread, intelligent, and trustworthy computing. The key technical contribution is in the areas at the interface of systems, hardware, and security, which would enable real-time AI. In the Systems domain, there are two key ideas: 1) the design of micro-kernel to fundamentally transform the time scale at which decisions using deep models are made; and, 2) incorporating the ability to replay the state and the decision history of the system. In the security and hardware domain, investigators are designing general purpose systems capable of running on a variety of hardware and cloud platform with an added key feature of tunable security that provides a trade-off between security and performance. In the AI domain, the major contributions of the project are in developing real-time systems and hardware supports that would assign tasks suitably to back-end and edge for fast accurate decision making.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.

人工智能系统将在人们的生活中发挥越来越重要的作用的新时代正在到来。这些系统将通过早期识别处于危险中的患者，使用纳米探针进行细胞水平的诊断和治疗，以及机器人手术，彻底改变医疗保健。它们将通过为自动驾驶汽车和无人驾驶飞机提供动力，减少交通拥堵，帮助减少死亡人数。而且，它们将通过实时检测和防御金融欺诈和互联网攻击，使企业更安全。更一般地说，这些系统将改变人们感知和与周围世界互动的方式，使其更能适应和响应我们的需求。为了实现这一愿景，需要新一代人工智能系统来为人类安全和福祉受到威胁的关键任务应用提供动力，并且可以在不断变化且意想不到的敌对环境中工作。除了智能之外，这些决策系统还需要解决四个挑战。首先，它们必须实时做出反应（即在几秒钟甚至几毫秒内做出决定），以支持机器人手术和自动驾驶汽车等应用。其次，随着环境的混乱演变，人工智能系统需要不断地从实时数据流中学习。第三，这些系统需要安全，即确保隐私、数据机密性和决策完整性。最后，由于这些系统代表人类做出决定，它们的决定需要向对人工智能理解有限的人解释。例如，如果一个人工智能系统诊断出一个患有罕见疾病的病人，或者认为某个测试没有根据，系统应该根据病人的病史和更大的人群的病史提供解释，而不是指向人工智能算法的内部计算。探险项目的目标是通过开发开源平台、工具和算法，为实时、智能、安全和可解释（RISE）决策构建人工智能决策系统，以应对这些挑战。实现这一目标需要将人工智能、安全、系统和硬件研究相结合的整体方法。例如，要在拥挤的城市中成功部署送货机器人车队，不仅需要人工智能方面的进步（例如，感知和安全导航复杂城市环境的能力），还需要系统方面的进步（例如，能够实时协调车辆的新型混合边缘云系统）、安全性（例如，确保机器人传感器收集的信息不会损害客户的隐私）和计算机架构(例如，硬件和软件协同设计，以降低功耗和提高安全性)。RISE项目旨在使一个大型的先锋社区能够基于它将创建的工具和想法构建创新的应用程序和解决方案，并扩大研究参与，允许许多学科的学生和研究人员贡献并构建其工件。围绕人工智能系统的公共开放平台建立和培育一个社区，将使下一个十年的创新以广泛、智能和可信的计算为中心。关键的技术贡献是在系统、硬件和安全的接口领域，这将使实时人工智能成为可能。在系统领域，有两个关键思想：1)微内核的设计从根本上改变了使用深度模型做出决策的时间尺度；2)整合系统状态和决策历史的回放能力。在安全和硬件领域，研究人员正在设计能够在各种硬件和云平台上运行的通用系统，并增加了可调安全性的关键特性，从而在安全性和性能之间进行权衡。在人工智能领域，该项目的主要贡献在于开发实时系统和硬件支持，将任务适当地分配给后端和边缘，以实现快速准确的决策。该奖项反映了美国国家科学基金会的法定使命，并通过使用基金会的知识价值和更广泛的影响审查标准进行评估，被认为值得支持。

项目成果

Jiffy: elastic far-memory for stateful serverless analytics

• DOI: 10.1145/3492321.3527539
• 发表时间: 2022-03
• 期刊: Proceedings of the Seventeenth European Conference on Computer Systems
• 作者: Anurag Khandelwal;Yupeng Tang;R. Agarwal;Aditya Akella;I. Stoica

Gauss: program synthesis by reasoning over graphs

高斯：通过图推理进行程序综合

• DOI: 10.1145/3485511
• 发表时间: 2021
• 期刊: Proceedings of the ACM on Programming Languages
• 作者: Bavishi, Rohan;Lemieux, Caroline;Sen, Koushik;Stoica, Ion
• 通讯作者: Stoica, Ion

How Computer Science and Statistics Instructors Approach Data Science Pedagogy Differently: Three Case Studies

计算机科学和统计学教师如何以不同的方式处理数据科学教学法：三个案例研究

• DOI: 10.1145/3478431.3499384
• 发表时间: 2022
• 期刊: SIGCSE 2022: Proceedings of the 53rd ACM Technical Symposium on Computer Science Education
• 作者: Lau, Sam;Nolan, Deborah;Gonzalez, Joseph;Guo, Philip J.
• 通讯作者: Guo, Philip J.

Learning to Design Accurate Deep Learning Accelerators with Inaccurate Multipliers

学习使用不准确的乘数设计准确的深度学习加速器

• DOI: 10.23919/date54114.2022.9774607
• 发表时间: 2022
• 期刊: Automation & Test in Europe Conference & Exhibition (DATE
• 作者: Jain, Paras;Huda, Safeen;Maas, Martin;Gonzalez, Joseph E.;Stoical, Ion;Mirhoseini, Azalia
• 通讯作者: Mirhoseini, Azalia

Elastic Hyperparameter Tuning on the Cloud

云端弹性超参数调优

• DOI: 10.1145/3472883.3486989
• 发表时间: 2021
• 期刊: SoCC '21: Proceedings of the ACM Symposium on Cloud Computing
• 作者: Dunlap, Lisa;Kandasamy, Kirthevasan;Misra, Ujval;Liaw, Richard;Jordan, Michael;Stoica, Ion;Gonzalez, Joseph E.
• 通讯作者: Gonzalez, Joseph E.