Collaborative Research: NeTS: Medium: CAML: Communication-constrained Adaptive Machine Learning in Heterogeneous Edge Networks

About This Grant

This project investigates the principles and methodologies for Artificial Intelligence of Things (AIoT) devices to make decisions quickly and accurately, to meet the requirements for safety, accuracy, and latency even with limited communication and computing power. By developing new machine learning techniques that work well in the presence of communication resource limitations, this project will spur a new line of thinking for a variety of AIoT applications that face the challenges of communication-constrained machine learning, such as smart health, connected cars, augmented reality, and smart city, benefiting society at large. This project will also contribute to skilled workforce development in this area of national needs, by integrating the research findings into undergraduate- and graduate-level education and organizing various summer programs on data science, AI and machine learning to broaden the participation of K-12 students. This project aims to develop innovative Communication-constrained Adaptive Machine Learning (CAML) methods that adapt to varying communication bandwidth, computational power, and data size of edge devices, thereby enabling heterogeneous devices to work in concert. Specifically, the project will investigate CAML for two popular edge network settings: edge networks with a central coordinator and fully decentralized edge networks. For edge networks with a central server, this project studies distributed fine-tuning for on-device machine learning model adaptation through dynamic tier-based low-rank model adaptation, allowing each device to train a local adapter of a suitable rank. This approach ensures synchronized model updates across devices with varying capabilities, improving the training efficiency of large machine learning models. Cross-layer optimization schemes will be developed to speed up the learning process while ensuring accuracy. To address the scenarios with extreme communication resource limitations, the project introduces seed-aided machine learning model tuning algorithms that enable meaningful model updates with minimal data exchanges. For decentralized edge networks, the project develops communication-efficient workload balancing algorithms to reduce synchronization delays. The innovations in this project promise to advance federated and collaborative learning across diverse edge environments, enhancing scalability, efficiency, and robustness in real-world AIoT applications. 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.