Collaborative Research: CNS Core: Small: Edge AI with Streaming Data: Algorithmic Foundations for Online Learning and Control

许多新兴应用，如智能医疗、自动驾驶和增强现实，都依赖于将实时人工智能(AI)应用于不断在线生成的流数据。Edge AI将AI服务移动到靠近生成数据流的最终用户和设备的网络边缘，对于减少延迟和通信瓶颈并实现快速准确的推理决策至关重要。然而，由于流数据的不可预测的动态以及网络边缘有限的计算/通信能力，在线流数据的边缘人工智能带来了巨大的挑战。该项目通过开发将复杂的学习方法与先进的边缘网络控制相结合的新理论模型，以及显著提高边缘AI服务对流数据的准确性和及时性的实用算法来应对这些挑战。具体地说，该项目将重点关注三个密切相关的主题：(I)将开发用于模型选择的在线学习策略，以快速确定哪些机器学习模型应该动态地部署在边缘服务器以获得最佳推理精度，同时考虑到异构性的切换和反馈成本；(Ii)将开发分布式在线迁移学习方法，以在新的流媒体数据下快速地重新训练边缘处的新机器学习模型；以及(Iii)将开发基于部分索引的边缘网络控制策略，以在资源紧张的情况下优化交互边缘人工智能服务的及时性。边缘网络和人工智能都被认为是下一代无线网络的关键要素。该项目将直接惠及部署和运营EDGE-AI系统的网络运营商和服务提供商。具体地说，结果将帮助他们自动化此类系统的端到端协调所需的复杂决策过程，并提高EDGE-AI服务的准确性和及时性，尽管环境不断变化。该项目还将惠及由EDGE AI支持的新兴应用的最终用户，改善他们的用户体验和福祉。更广泛地说，在这个项目中开发的学习/控制协同设计的理论和算法不仅将改变边缘人工智能，而且还将在显著的动态性和不确定性下惠及具有类似优化要求的其他学科。最后，该项目将为多个本科生和研究生课程贡献教学和培训材料，并将通过接触当地学校来吸引女性和代表性不足的少数族裔学生。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。

Many emerging applications, such as smart healthcare, autonomous driving, and augmented reality, rely on applying real-time Artificial Intelligence (AI) to streaming data that are constantly generated online. Edge AI, which moves AI services to the network edge close to the end users and devices where data streams are generated, is crucial for reducing latency and communication bottlenecks and enabling fast and accurate inference decisions. However, edge AI for online streaming data poses significant challenges due to the unpredictable dynamics of the streaming data and the limited computation/communication capability at the network edge. This project addresses these challenges by developing both new theoretic models that integrate sophisticated learning methods with advanced edge-network control, and practical algorithms that significantly improve the accuracy and timeliness of edge AI services for streaming data. Specifically, the project will focus on three closely-related thrusts: (i) online learning policies for model selection will be developed to quickly identify which machine-learning models should be dynamically deployed at the edge servers for best inference accuracy, while accounting for the heterogeneous switching and feedback costs; (ii) distributed online transfer learning methods will be developed to quickly retrain new machine learning models at the edge upon new streaming data; and (iii) partial-index based edge-network control policies will be developed to optimize the timeliness of interactive edge-AI services under tight resource constraints.Both edge networks and AI are considered crucial elements of next-generation wireless networks. This project will directly benefit network operators and service providers that deploy and operate edge-AI systems. Specifically, the results will help them automate the complex decision-making process required for the end-to-end orchestration of such systems, and improve the accuracy and timeliness of the edge-AI services despite the constantly-changing environments. This project will also benefit the end users of emerging applications powered by edge AI, improving their user experience and well-being. More broadly, the theories and algorithms developed in this project for learning/control co-design will not only transform edge AI, but also benefit other disciplines with similar requirements for optimization under significant dynamism and uncertainty. Finally, this project will contribute teaching and training materials to multiple undergraduate and graduate courses, and will engage women and underrepresented minority students by reaching out to local schools.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.