NSF-AoF: CNS Core Small: Lean-NextG: Learning to Network the Edge in Next Generation Wireless Networks

Yale University, New Haven CT

Investigators

Abstract

Artificial Intelligence (with its subset known as Machine Learning, ML) and wireless networking (in its most recent generation, 5G, and beyond) are two disruptive technologies paving the way for the world of tomorrow. The synergy of these technologies is expected to benefit a wide range of application domains, such as healthcare and environmental sustainability, that attract increasing attention as the world moves on from the COVID-19 pandemic. In spite of this potential, the existing frameworks for distributed ML used by these applications fail to fully utilize the different types of network resources when managing the data they use for training models. Furthermore, judiciously selecting important subsets of data and making them available where they are needed the most continues to be open an problem. This project aims to develop the theoretical foundations for the in-network management of the training data to address the above challenges and enhance the performance of the distributed ML models and corresponding applications. Towards this, it combines tools from supervised and reinforcement learning, deep neural networks, and network optimization. Building on ongoing testbed and experimentation activities, the theoretical framework is implemented and experimented with real-life wireless testbed components. The PIs will also engage in various outreach activities targeting K-12 students. This project proposes a new distributed Machine Learning (ML) methodology that departs significantly from the current practice of Federated Learning (FL) by enabling devices at the network edge to share training data to each other as part of an overall fivefold data management strategy that includes collecting, discarding, caching, processing and transferring of data. These are additional degrees of freedom in optimizing data management that can help improving the performance of the ML models and the corresponding applications while at the same time exploring the entire gamut of tradeoffs between the different types of resources in the edge network. Specifically, the following three interdependent research thrusts are proposed: (i) an optimization framework for data management that shifts the attention from "where the data is collected"' to "where the data is processed" this way facilitating the co-design of solutions to related learning and networking problems, (ii) an extension of these solution methods by learning the importance of the available data and optimizing their management decisions accordingly, and (iii) a testbed implementation and experimentation with various applications including one for environmental sensing and another for network slicing aimed at augmented reality and virtual reality used in tele-education. 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.