Robust and light-weight overlay management for decentralized learning

A growing number of companies are extracting value from the digital data produced by our modern society using Machine learning (ML) techniques. Most of these companies rely today on centralized or tightly coupled ML systems hosted in data centers or in the cloud [5, 6]. This is problematic as this concentration poses strong risks to the privacy of users, and limits the scope of ML applications to tightly integrated datasets under unified learning models. To address these limitations, this PhD proposes to explore an alternative approach inspired by peer-topeer networks in which users control their own system, and only exchange a limited amount of information to construct local machine learning models. This strategy is more amenable to preserving user privacy, and respecting the constraints possibly imposed on sensitive data-sets (such as health records, or personal financial data), and holds the potential for highly scalable and robust learning systems. This project aims to study the challenges raised by this strategy in terms of distribution and overlay management.