NSF-AoF: Lean-6G: Learning to Network the Edge in 6G

Background

This project aims at developing theoretical foundations for in-network data management to exploit different types of network resources, enhance the performance of the distributed machine learning models and corresponding applications. Distributed machine learning leverages resources at the network edge, near the data sources, for model training and inference, thereby offloading from these tasks the centralized cloud platforms and satisfying demands of applications for low latency, scalability and privacy. Training ML models at realistic 5GIoT network scales requires processing enormous volumes of data generated by the end-user devices, thus leading to huge amount of data communications between network users and the central cloud platform. As a consequence, the data uploading and training procedure in real-time applications (e.g., ARVR) causes significant performance degradation. Finally, since the training data are collected by network users, preserving the privacy of network users with affordable accuracy loss is an important issue in distributed ML. Developing data aware resource allocation algorithms in distributed ML systems is a complex and open problem. Existing works on distributed ML can be classified into two categories: (i) one focusing on how to run the ML tasks (model training and inference) at the edge devices in a collaborative and resource-optimized manner, and (ii) another focusing on how to exploit ML to find a better solution to the optimization problems in the edge network. It has been illustrated that data quality is a critical factor in machine learning algorithms, and data importance-aware network resource allocation and training algorithms can improve ML systems performance. Therefore, a unified treatment of data selection, communication and computing is required to improve the performance of distributed ML systems. More importantly, protecting against data leakage in the distributed resource-sharing process is an unresolved practical issue in protocol realization.

Goals

This project aims to address the above challenges in theoretical research and healthcare applications that arise in distributed machine learning systems. Overall, the major goals in this project can be summarized as follows:

• Develop a systematic framework for describing the relationship between communication load, training accuracy and privacy leakage in distributed machine learning systems with heterogeneous data.

• Propose in-network data management strategies that enables devices at the network edge to share training data to each other as part of an overall data management strategy that includes discarding, caching, processing and transferring of data. We will start with a framework for optimizing such data management decisions under limited availability of devices’ resources and volatility of network ant then extend our framework for handling learning tasks with unlabeled data.

• Quantify data importance for machine learning tasks based on data quality, locality and relationship with the ultimate learning goals. Develop data importance-aware network resource management strategy and training algorithms to improve the machine learning accuracy.

• Develop methods to support concurrent learning applications each with its own subset of important data.

• Verity the theoretical results of previous algorithms on proof-of-concept testbeds for key application use cases: (i) real-world urban sensor testbeds in Helsinki, Finland for environmental sensing, and (ii) a lab 5G-enabled mobile edge computing testbed at the Yale University campus, USA for AR/VR aimed at tele-education.

Contact

Prof. Leandros Tassiulas Yale University
Prof. Sasu Tarkoma University of Helsinki