Oort - Efficient Federated Learning via Guided Participant Selection
Offered By: USENIX via YouTube
Course Description
Overview
Explore an innovative approach to federated learning in this conference talk from OSDI '21. Delve into Oort, a guided participant selection system designed to enhance the efficiency of federated training and testing. Learn how Oort prioritizes clients with high-utility data and fast training capabilities to improve time-to-accuracy performance and final model accuracy. Discover techniques for interpreting results in model testing while meeting specific distribution requirements. Examine the challenges of identifying heterogeneous client utility, selecting high-utility clients at scale, and adapting selection processes. Gain insights into the anatomy of time-to-accuracy in training and the statistical performance improvements achieved through this novel approach to federated learning.
Syllabus
Intro
Emerging Trend of Machine Learning
Emerging Federated Learning on the Edge
Execution of Federated Learning (FL)
Challenges in Federated Learning
Existing Client Selection: Suboptimal Efficiency
Existing Client Selection: Unable for Selection Criteria
Oort: Guided Participant Selection for FL
Anatomy of Time to Accuracy in Training
Challenge I: Identify Heterogeneous Client Utility
Challenge 2: Select High-Utility Clients at Scale
Challenge 3: Select High-Utility Clients Adaptively
Time-to-Accuracy (TTA) Performance
Zoom into Statistical Performance
Taught by
USENIX
Related Courses
GraphX - Graph Processing in a Distributed Dataflow FrameworkUSENIX via YouTube Theseus - An Experiment in Operating System Structure and State Management
USENIX via YouTube RedLeaf - Isolation and Communication in a Safe Operating System
USENIX via YouTube Microsecond Consensus for Microsecond Applications
USENIX via YouTube KungFu - Making Training in Distributed Machine Learning Adaptive
USENIX via YouTube