CNS: Medium: Ground up Adaptive Learning System for Heterogeneous Environments

CNS：中：异构环境的基础自适应学习系统

• 批准号: 2211882
• 负责人: Tianqi Chen
• 金额: $ 120万
• 依托单位: Carnegie-Mellon University
• 依托单位国家: 美国
• 项目类别: Continuing Grant
• 财政年份: 2022
• 资助国家: 美国
• 起止时间: 2022-10-01 至 2025-09-30
• 项目状态: 未结题
• 来源: https://www.nsf.gov/awardsearch/showAward?AWD_ID=2211882&HistoricalAwards=false
• 关键词: CNS; Medium; Ground; up; Adaptive

Artificial intelligence (AI) is becoming ubiquitous, with applications running on smartphones, IoT devices, self-driving cars, and the cloud. These applications are supported by different types of computer hardware, including hardware specially designed to accelerate AI applications. To operate as intended, these applications require a training process, which is usually performed using a shared infrastructure containing a variety of hardware. Training and deploying AI applications in diverse, real-world computing environments takes significant human effort. This project aims to realize adaptive AI computations, which enable AI programs to move across different hardware types and react quickly to resource and task changes in shared computing infrastructures – and to do so automatically. The project will build an end-to-end solution that (1) coordinates when and where to run AI programs for a given set of need; (2) automatically partitions, distributes, and configures AI tasks across different hardware resources; and (3) builds AI programs that can seamlessly migrate across hardware platforms.Modern society relies on AI systems and applications. Considering adaptivity and accessibility as first-order objectives when building AI systems can have far-reaching impacts on how people develop and deploy AI applications. For example, with the help of this project, developers will be able to run AI applications on their platform of choice without being locked into a single platform or vendor. This project will also enable coordination among multiple AI applications on shared resources to reduce the overall cost of running AI applications. These improvements will make AI systems more affordable and accessible to small organizations and individual practitioners. Finally, this project will also increase the resource and energy efficiency of AI applications and will directly contribute to reducing the burden that data centers place on energy infrastructure and, thereby, the environment. The project also includes activities to engage undergraduate researchers and to broaden the participation of underrepresented groups in computing.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（AI）正变得无处不在，应用程序可以在智能手机、物联网设备、自动驾驶汽车和云上运行。这些应用程序由不同类型的计算机硬件支持，包括专门设计用于加速人工智能应用程序的硬件。为了按预期操作，这些应用程序需要一个培训过程，该过程通常使用包含各种硬件的共享基础设施来执行。在各种真实的计算环境中训练和部署人工智能应用程序需要大量的人力。该项目旨在实现自适应人工智能计算，使人工智能程序能够在不同的硬件类型之间移动，并对共享计算基础设施中的资源和任务变化做出快速反应，并自动完成。该项目将构建一个端到端解决方案，其中：(1)协调何时何地运行人工智能程序以满足给定的需求；(2)跨不同硬件资源自动分区、分配和配置AI任务；(3)构建可以跨硬件平台无缝迁移的人工智能程序。现代社会依赖于人工智能系统和应用。在构建人工智能系统时，将适应性和可访问性作为首要目标，可能会对人们如何开发和部署人工智能应用程序产生深远的影响。例如，在这个项目的帮助下，开发人员将能够在他们选择的平台上运行人工智能应用程序，而不会被锁定在单一平台或供应商上。该项目还将在共享资源上实现多个人工智能应用程序之间的协调，以降低运行人工智能应用程序的总体成本。这些改进将使小型组织和个人从业者更能负担得起人工智能系统。最后，该项目还将提高人工智能应用的资源和能源效率，并将直接有助于减轻数据中心对能源基础设施的负担，从而减轻对环境的负担。该项目还包括吸引本科生研究人员的活动，并扩大未被充分代表的群体在计算机领域的参与。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

Federated Learning under Distributed Concept Drift

• DOI: 10.48550/arxiv.2206.00799
• 发表时间: 2022-06
• 期刊: ArXiv
• 作者: Ellango Jothimurugesan;Kevin Hsieh;Jianyu Wang;Gauri Joshi;Phillip B. Gibbons

ED-Batch: Efficient Automatic Batching of Dynamic Neural Networks via Learned Finite State Machines

• DOI: 10.48550/arxiv.2302.03851
• 发表时间: 2023-02
• 作者: Siyuan Chen-;Pratik Fegade;Tianqi Chen;Phillip B. Gibbons;T. Mowry

Sia: Heterogeneity-aware, goodput-optimized ML-cluster scheduling

• DOI: 10.1145/3600006.3613175
• 发表时间: 2023-10
• 期刊: Proceedings of the 29th Symposium on Operating Systems Principles
• 作者: Suhas Jayaram Subramanya;Daiyaan Arfeen;Shouxu Lin;Aurick Qiao;Zhihao Jia;G. Ganger

SparseTIR: Composable Abstractions for Sparse Compilation in Deep Learning

• DOI: 10.1145/3582016.3582047
• 发表时间: 2022-07
• 期刊: Proceedings of the 28th ACM International Conference on Architectural Support for Programming Languages and Operating Systems, Volume 3
• 作者: Zihao Ye;Ruihang Lai;Junru Shao;Tianqi Chen-;L. Ceze