Federated Learning: Machine Learning on Decentralized Data
summary
Federated learning trains models across phones without moving raw data, using TensorFlow Federated to simulate rounds of on-device training and server aggregation. Production techniques include secure aggregation so the server cannot read individual updates, and clipping plus noise for differential privacy. Empirical results: language model converges after ~1,000 rounds (~1 week) and boosts next-word accuracy by 25% and prediction-strip clicks by 10%; emoji accuracy +7% and clicks +4%, with 11% more users sending emoji; GIF suggestion model cuts unhelpful prompts by 47%.
AI-written summary · about
chapters (11)
- 0:00Intro
- 0:43Agenda
- 1:21Decentralized Data
- 3:51Federated Computation
- 7:49Federated Computation vs Decentralized Computation
- 12:53Secure Aggregation
- 14:53Federated Learning Workflow
- 19:23Federated Learning vs Traditional Distributed Learning
- 21:43Language Modeling
- 27:23New Words
- 29:36Dont Memorize
related talks
34mDecentralized Machine Learning
42mLearning Together, Distributed: An Introduction to Federated Learning
48mFederated ML for Edge Applications
38mFederated Learning and Its Application...
31mFederated Machine Learning with Kubernetes
42mFederated Learning: Unlocking the Values of Discrete Data for Enterprise- 31mFederated Machine Learning with Kubernetes
31mFrom Theory to Practice: Advances in Real-World Federated Learning