I run in coordination with Alexandru-Andrei Iacob the Lab sessions of this course.

Objectives

This course aims to extend the machine learning knowledge available to students in Part I (or present in typical undergraduate degrees in other universities), and allow them to understand how these concepts can manifest in a decentralized setting. The course will consider both theoretical (e.g., decentralized optimization) and practical (e.g., networking efficiency) aspects that combine to define this growing area of machine learning.

At the end of the course students should:

• Understand popular methods used in federated learning
• Be able to construct and scale a simple federated system
• Have gained an appreciation of the core limitations to existing methods, and the approaches available to cope with these issues
• Developed an intuition for related technologies like differential privacy and secure aggregation, and are able to use them within typical federated settings
• Can reason about the privacy and security issues with federated systems
• Lectures
• Course Overview. Introduction to Federated Learning.
• Decentralized Optimization.
• Statistical and Systems Heterogeneity.
• Variations of Federated Aggregation.
• Secure Aggregation.
• Differential Privacy within Federated Systems.
• Extensions to Federated Analytics.
• Applications to Speech, Video, Images and Robotics.
• Lab sessions
• Federating a Centralized ML Classifier.
• Behaviour under Heterogeneity.
• Scaling a Federated Implementation.
• Exploring Privacy with Federated Settings