Decentralized Machine Learning
summary
Claims that models train on-device without centralizing raw data and that privacy holds up when updates are Secure Aggregation. Demonstrations use Federated Averaging on non‑IID mobile batches, a recurrent model for Gboard glide typing, and simulations with TensorFlow Federated and LEAF: loss decreases in trials but production readiness is unresolved. Notes concrete costs: reduced communication via quantization, threshold decryption only after ~30 clients, severe challenges from skewed data, intermittent connectivity, and that Nubank (13 million customers) is still evaluating, not yet deploying.
AI-written summary · about
chapters (14)
- 0:00Introduction
- 1:12Speaker background
- 2:57Federated learning overview
- 4:22Privacy advantages of federated learning
- 7:13Real-world case
- 10:24Three-step process
- 13:54Communication reduction
- 15:42Aggregation methods
- 18:10Federated averaging
- 20:38Secure aggregation
- 22:50Workflow comparison
- 25:01TensorFlow Federated demo
- 28:36Adoption and outlook
- 33:01Optimization and Q&A
related talks
41mFederated Learning: Machine Learning on Decentralized Data
42mLearning Together, Distributed: An Introduction to Federated Learning
48mFederated ML for Edge Applications
26mMachine Learning without the Data #federatedlearning
29mBalancing Thin Line Between Data Intelligence & Privacy
34mTraining without sharing: Is Federated Learning the answer to data privacy in AI?
40mDemystifying Decentralized Fair Models for Federated Machine Learning
31mFederated Machine Learning with Kubernetes