Oort: Efficient Federated Learning via Guided Participant Selection
Talkat 15th USENIX Symposium on Operating Systems Design and Implementation (OSDI '21) · 2021-07-14..07-16·
summary
Oort is a participant-selection framework that selects clients by a combined loss-over-duration utility score, using exploration/exploitation and uncertainty from last-observation age. Prototype integrates with TensorFlow Federated and PyTorch, runs on a 68-GPU testbed simulating millions of clients, and prioritizes fast, high-utility clients per round. Evaluation shows concrete benefits: 1.2X–14.1X time-to-accuracy improvements and 1.3%–9.8% final accuracy gains while enforcing developer-specified test distributions.
AI-written summary · about
chapters (14)
- 0:00Intro
- 0:21Emerging Trend of Machine Learning
- 0:56Emerging Federated Learning on the Edge
- 1:30Execution of Federated Learning (FL)
- 2:56Challenges in Federated Learning
- 4:19Existing Client Selection: Suboptimal Efficiency
- 5:33Existing Client Selection: Unable for Selection Criteria
- 6:32Oort: Guided Participant Selection for FL
- 7:09Anatomy of Time to Accuracy in Training
- 8:19Challenge I: Identify Heterogeneous Client Utility
- 9:53Challenge 2: Select High-Utility Clients at Scale
- 11:01Challenge 3: Select High-Utility Clients Adaptively
- 13:23Time-to-Accuracy (TTA) Performance
- 14:22Zoom into Statistical Performance
related talks
51mExploiting Similarity in Federated Learning
06mPREFER: Point-of-interest REcommendation with efficiency and privacy-preservation via Federated ...
01h 00mFedOpt for LLMs
42mLearning Together, Distributed: An Introduction to Federated Learning
31mFrom Theory to Practice: Advances in Real-World Federated Learning
04mProject Lightning Talk: Empowering Federated Learning with Multi-Cluster Management for Privacy and Efficiency
31mFederated Machine Learning with Kubernetes
34mTraining without sharing: Is Federated Learning the answer to data privacy in AI?