CAREER: Towards a Communication Foundation for Distributed and Decentralized Machine Learning

职业：为分布式和去中心化机器学习建立通信基础

• 批准号: 2143559
• 负责人: Cong Shen
• 金额: $ 50万
• 依托单位: University of Virginia Main Campus
• 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: 2022-09-01 至 2027-08-31
• 项目状态: 未结题
• 来源: https://www.nsf.gov/awardsearch/showAward?AWD_ID=2143559&HistoricalAwards=false
• 关键词: CAREER; Towards; Communication; Foundation; Distributed

With the emerging paradigm shift towards moving the data collection and machine learning (ML) model training to the edge, distributed and decentralized ML has become increasingly critical to empowering many applications, such as autonomous driving, recommender systems, and Internet of Things (IoT). This trend imposes formidable challenges on the underlying communication design and catalyzes its evolution from connecting people and connecting things to connecting intelligence. This CAREER project develops fundamental communication technologies to enable distributed and decentralized ML in next-generation wireless systems. It transforms wireless communications from pure data transfer to intelligence transfer, building a synergy between communications and ML in a closely integrated fashion. In partnership with industry, results enabled by this project can be prototyped and integrated into real systems, potentially impacting 6G standardization and other future communication systems. The cross disciplinary nature of this project naturally translates into case studies and new development in a number of undergraduate and graduate level courses, by integrating ML and AI to the curriculum of communications and networking. The education and outreach activities will collectively promote a common thread of providing the best opportunities for diverse groups of bright young minds to develop into future scientists and engineers.This project aims at developing the theoretical foundation and novel communication algorithms for distributed and decentralized ML, thereby catalyzing a paradigm shift of wireless communications towards connecting intelligence. Towards this end, this project will develop a novel random orthogonalization principle that tightly integrates physical layer communications with ML. Additionally, the research will study the impact of fading and noisy channels on the performance of ML, and design communication methods to improve the accuracy and convergence of the ML tasks. Finally, a novel adaptive communication method for distributed and decentralized multi-armed bandits will be investigated, where coding and interleaving designs for online learning with adversarial communications will be studied. The proposed research promotes the fundamental understanding of the synergy between distributed and decentralized ML and communications, and will have broad applications beyond the specific problems studied in this project.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.

随着数据收集和机器学习 (ML) 模型训练向边缘转移的新兴范式转变，分布式和去中心化的 ML 对于支持自动驾驶、推荐系统和物联网 (IoT) 等许多应用程序变得越来越重要。这种趋势对底层通信设计提出了巨大的挑战，并促进了其从连接人、连接物到连接智能的演变。该 CAREER 项目开发基础通信技术，以在下一代无线系统中实现分布式和去中心化的机器学习。它将无线通信从纯粹的数据传输转变为情报传输，以紧密集成的方式在通信和机器学习之间建立协同作用。通过与业界合作，该项目的成果可以进行原型设计并集成到实际系统中，这可能会影响 6G 标准化和其他未来的通信系统。通过将机器学习和人工智能整合到通信和网络课程中，该项目的跨学科性质自然地转化为案例研究和许多本科生和研究生水平课程的新发展。教育和推广活动将共同促进一个共同点，即为不同群体的聪明年轻人提供发展成为未来科学家和工程师的最佳机会。该项目旨在为分布式和去中心化机器学习开发理论基础和新颖的通信算法，从而促进无线通信向连接智能的范式转变。为此，该项目将开发一种新颖的随机正交化原理，将物理层通信与机器学习紧密集成。此外，该研究还将研究衰落和噪声信道对机器学习性能的影响，并设计通信方法以提高机器学习任务的准确性和收敛性。最后，将研究一种用于分布式和去中心化多臂强盗的新型自适应通信方法，其中将研究用于对抗性通信的在线学习的编码和交织设计。拟议的研究促进了对分布式和去中心化机器学习与通信之间协同作用的基本理解，并将具有超出本项目研究的具体问题的广泛应用。该奖项反映了 NSF 的法定使命，并通过使用基金会的评估进行评估，被认为值得支持。智力价值和更广泛的影响审查标准。

项目成果

Teaching Reinforcement Learning Agents via Reinforcement Learning

通过强化学习教授强化学习代理

• DOI: 10.1109/ciss56502.2023.10089695
• 发表时间: 2023-03
• 期刊: 2023 57th Annual Conference on Information Sciences and Systems (CISS
• 作者: Yang, Kun;Shi, Chengshuai;Shen, Cong
• 通讯作者: Shen, Cong

Provably Efficient Offline Reinforcement Learning with Perturbed Data Sources

使用扰动数据源进行可证明高效的离线强化学习

• 发表时间: 2023-07
• 期刊: International Conference on Machine Learning
• 作者: Shi, C.;Xiong, W.;Shen, C.;Yang, J.
• 通讯作者: Yang, J.

Nearly Minimax Optimal Offline Reinforcement Learning with Linear Function Approximation: Single-Agent MDP and Markov Game

具有线性函数逼近的近极小极大最优离线强化学习：单智能体 MDP 和马尔可夫博弈

• 发表时间: 2023-01
• 期刊: International Conference on Learning Representations (ICLR
• 作者: Xiong, Wei;Zhong, Han;Shi, Chengshuai;Shen, Cong;Wang, Liwei;Zhang, Tong
• 通讯作者: Zhang, Tong

Near-optimal Conservative Exploration in Reinforcement Learning under Episode-wise Constraints

逐集约束下强化学习的近最优保守探索

• 发表时间: 2023-07
• 期刊: International Conference on Machine Learning
• 作者: Li, D.;Huang, R.;Shen, C.;Yang, J.
• 通讯作者: Yang, J.

Federated Learning via Indirect Server-Client Communications

通过间接服务器-客户端通信进行联合学习

• DOI: 10.1109/ciss56502.2023.10089783
• 发表时间: 2023-03
• 期刊: Annual Conference on Information Sciences and Systems (CISS
• 作者: Bian, Jieming;Shen, Cong;Xu, Jie
• 通讯作者: Xu, Jie