CAREER: Computing-Aware Network Optimization for Efficient Distributed Data Analytics at the Wireless Edge

职业：计算感知网络优化，用于无线边缘的高效分布式数据分析

• 批准号: 2110259
• 负责人: Jia Liu
• 金额: $ 52.41万
• 依托单位: Ohio State University
• 依托单位国家: 美国
• 项目类别: Continuing Grant
• 财政年份: 2020
• 资助国家: 美国
• 起止时间: 2020-10-01 至 2025-09-30
• 项目状态: 未结题
• 来源: https://www.nsf.gov/awardsearch/showAward?AWD_ID=2110259&HistoricalAwards=false
• 关键词: CAREER; Computing; Aware; Network; Optimization

In recent years, machine learning (ML) and artificial intelligence (AI) applications are quickly finding their ways into our everyday life. All of these applications generate and inject a massive volume of data into the network for a wide range of complex ML/AI data analytics tasks, including but not limited to the training and/or inferences in computer vision, natural language processing, recommendation systems, etc. However, most of the existing wireless network control and optimization algorithms rarely take the new characteristics of ML/AI data analytic traffic into considerations. Likewise, most ML/AI data analytics algorithms oversimplify the underlying wireless networks as "bit pipes" and ignore their complex networking and physical layer constraints, hence leading to poor overall data analytics efficiency. The overarching theme of this CAREER research program is to bridge the gap between the rapidly growing ML/AI data analytics demands and the existing networking and communication technologies. The principal investigator (PI) explore a cross-disciplinary understanding between wireless networking and data analytics through a unified research program, which consists of the development of tractable theoretical models, exploration of theoretical performance bounds and limits, and the development of low-complexity distributed algorithms and protocols that are easy to implement in practice.In this CAREER program, the PI will develop networking-computing co-designs to facilitate ML/AI data analytics with data and model parallelisms in wireless edge networks. The PI will focus on three complementary research thrusts, each of which addresses one key aspect in supporting distributed data analytics at a different protocol layer: (i) communication-efficient distributed optimization at the physical layer; (ii) joint-queueing-computing scheduling at the medium access control layer; and (iii) admission control and resource virtualization at the transport layer. Collectively, the results in this research contribute to a new direction of wireless network control and optimization theory and systems design. The proposed research will serve as a foundation of the next-generation wireless networking that supports a plethora of data analytics and ML/AI applications. Due to its unique scientific and engineering challenges, this research program encompasses strong and holistic expertise in mathematical modeling, optimization, control, queueing theory, stochastic analysis, as well as deep knowledge of ML/AI system operations in practice. The proposed research will support not only the networking, communications, control, and machine learning research communities, but also the general public, by developing new optimization technologies for substantially improved network and data analytics performances.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.

近年来，机器学习（ML）和人工智能（AI）应用正迅速进入我们的日常生活。所有这些应用程序生成并注入大量数据到网络中，用于各种复杂的ML/AI数据分析任务，包括但不限于计算机视觉、自然语言处理、推荐系统等方面的训练和/或推理。然而，现有的大多数无线网络控制和优化算法很少考虑ML/AI数据分析流量的新特点。同样，大多数ML/AI数据分析算法将底层无线网络过度简化为“位管道”，忽略了其复杂的网络和物理层约束，从而导致整体数据分析效率低下。这个CAREER研究项目的首要主题是弥合快速增长的ML/AI数据分析需求与现有网络和通信技术之间的差距。首席研究员（PI）通过统一的研究计划探索无线网络和数据分析之间的跨学科理解，该研究计划包括开发易于处理的理论模型，探索理论性能界限和限制，以及开发易于在实践中实现的低复杂性分布式算法和协议。在该CAREER项目中，PI将开发网络计算协同设计，以促进无线边缘网络中数据和模型并行的ML/AI数据分析。PI将专注于三个互补的研究重点，每个重点都解决了在不同协议层支持分布式数据分析的一个关键方面：(i)物理层的通信高效分布式优化；（ii）介质访问控制层联合排队计算调度；（三）传输层的准入控制和资源虚拟化。本文的研究成果为无线网络控制与优化理论和系统设计开辟了新的方向。拟议的研究将作为支持大量数据分析和ML/AI应用的下一代无线网络的基础。由于其独特的科学和工程挑战，该研究项目在数学建模，优化，控制，排队理论，随机分析以及ML/AI系统实际操作方面拥有强大而全面的专业知识。拟议的研究不仅将支持网络、通信、控制和机器学习研究社区，还将通过开发新的优化技术来大幅提高网络和数据分析性能，从而支持公众。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

Low Sample and Communication Complexities in Decentralized Learning: A Triple Hybrid Approach

• DOI: 10.1109/infocom42981.2021.9488686
• 发表时间: 2021-05
• 期刊: IEEE INFOCOM 2021 - IEEE Conference on Computer Communications
• 作者: Xin Zhang;Jia Liu;Zhengyuan Zhu-;E. Bentley

Taming Fat-Tailed (“Heavier-Tailed” with Potentially Infinite Variance) Noise in Federated Learning," in Proc. NeurIPS, New Orleans, LA, Dec. 2022

Taming Fat-Tailed (“Heavier-Tailed” with Potentially Infinite Variance) Noise in Federated Learning，载于 Proc. NeurIPS，路易斯安那州新奥尔良，2022 年 12 月

• 发表时间: 2022
• 期刊: Proc. NeurIPS
• 作者: Yang, H.;Qiu, P.;Liu, J.
• 通讯作者: Liu, J.

Federated Learning with Fair Worker Selection: A Multi-Round Submodular Maximization Approach

• DOI: 10.1109/mass52906.2021.00033
• 发表时间: 2021-07
• 期刊: 2021 IEEE 18th International Conference on Mobile Ad Hoc and Smart Systems (MASS)
• 作者: Fengjiao Li;Jia Liu;Bo Ji

GT-STORM: Taming Sample, Communication, and Memory Complexities in Decentralized Non-Convex Learning

• DOI: 10.1145/3466772.3467056
• 发表时间: 2021-05
• 期刊: Proceedings of the Twenty-second International Symposium on Theory, Algorithmic Foundations, and Protocol Design for Mobile Networks and Mobile Computing
• 作者: Xin Zhang;Jia Liu;Zhengyuan Zhu-;E. Bentley

Adaptive Multi-Hierarchical signSGD for Communication-Efficient Distributed Optimization

用于通信高效分布式优化的自适应多层次符号SGD

• DOI: 10.1109/spawc48557.2020.9154256
• 发表时间: 2020
• 期刊: Special Session on Distributed Signal Processing for Coding and Communications
• 作者: Yang, Haibo;Zhang, Xin;Fang, Minghong;Liu, Jia
• 通讯作者: Liu, Jia