CAREER: "Adapt, Learn, Collaborate" — Closing the Pervasive Edge AI Loop with Liquid Intelligence

随着海量边缘计算技术的成熟，人们越来越期待人工智能(AI)的边缘部署，并希望实现一个能够从各种边缘设备提供及时的AI服务并进行边缘学习以实现可持续服务定制的“普适边缘AI”生态系统。然而，考虑到边缘设备的巨大异构性，一系列的挑战也随之而来：(1)如何高效地生成自适应的AI模型来适应巨大的边缘系统异构性？(2)如何同时克服边缘学习协作中的AI模型异构性？(3)如何统一异构的边缘计算系统并提供可扩展的边缘AI支持？本项目通过三项主要研究工作解决了这些基本挑战：推力1深入研究了不同深度神经网络(DNN)算子结构的硬件资源消耗，并设计了一种新的神经体系结构搜索(NAS)方法来解决体系结构级别的边缘异构性。它显著提高了边缘AI模型生成和自适应的有效性和效率。推力2通过揭示潜在的结构-信息关联问题，提高了联邦学习技术在异质AI模型中的能力。因此，它能够以最佳的融合和通信性能实现极端异类边缘协作。Struts 3通过硬件抽象和软件虚拟化革新了全栈EDGE AI系统支持，增强了普适EDGE AI开发的系统可扩展性和易开发性，完成这些推力后，本项目将实现深度学习、计算系统、EDGE协作技术和相关EDGE AI应用的联合创新。考虑到其灵活性和可行性的显著优势，我们将该框架命名为液体智能。随着这个项目的成功，一个无处不在、持续不断的边缘AI生态系统可能成为目标，推动下一波AI应用、边缘计算以及其他技术。它还将为许多社会挑战做出贡献，如智能城市、医疗保健信息学、工业基础设施等。该教育计划通过将嵌入式系统、移动计算和机器学习方面的跨学科模块与新开发的教学实践相结合，增强了现有的课程和教学方法。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。

With the maturity of massive edge computing technologies, people are increasingly looking forward to the edge deployment of artificial intelligence (AI) and hope to realize a “pervasive edge AI” ecosystem that could provide timely AI services from various edge devices and perform edge learning for sustainable service customization. However, given the vast heterogeneity in edge devices, a series of challenges also emerge: (1) How to efficiently generate adaptive AI models to accommodate vast edge system heterogeneity? (2) How to overcome the AI model heterogeneity across edge learning collaboration at the same time? (3) How to unify heterogeneous edge computing systems and provide scalable edge AI support? This project addresses these fundamental challenges with three major research thrusts: Thrust 1 thoroughly investigates the hardware resource consumption of different deep neural network (DNN) operator structures and designs a novel neural architecture search (NAS) method to address the edge heterogeneity down to the architecture level. It significantly improves the effectiveness and efficiency of edge AI model generation and adaptation. Thrust 2 advances the federated learning technique’s capability with heterogeneous AI models, by revealing the underlying structure-information correlation problem. And therefore, it enables extreme heterogeneous edge collaboration with optimal convergence and communication performance. Thrust 3 renovates the full-stack edge AI system support with hardware abstraction and software virtualization to enhance the system scalability and ease of development for pervasive edge AI development.With these thrusts completed, this project could achieve a joint innovation with deep learning, computing system, edge collaboration techniques, and related edge AI applications. Regarding its significant advantages of flexibility and feasibility, we name this framework as “Liquid Intelligence”. With the success of this project, a pervasive and continuous edge AI ecosystem could be targeted, boosting the next wave of AI applications, edge computing, as well as other technologies. It will also contribute to many societal challenges, such as smart cities, healthcare informatics, industrial infrastructures, etc. The education plan enhances existing curricula and pedagogy by integrating interdisciplinary modules on embedded systems, mobile computing, and machine learning with newly developed teaching practices.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.