Managed federated learning on AWS: A case study for healthcare
summary
Reproduces federated training on eICU (200k stays, ~200 hospitals) to predict in‑hospital mortality, compares centralized vs federated runs, and evaluates open source stacks (TensorFlow Federated, PySyft/PyGrid, FedML). Implements FedML on AWS with VPC peering and gRPC for cross‑silo experiments, uses custom partitioned data loaders, and tracks per‑round bytes, rounds, and AUC/loss. Notes operational gaps: orchestration, instance sizing, decentralized logging and debugging, and privacy needs addressed via differential privacy and secure aggregation; centralized AUC ~0.89, federated ~0.88.
AI-written summary · about
chapters (15)
- 0:00Intro
- 1:31Team
- 5:26Privacy-preserving federated learning (PPFL)
- 10:38What makes this hard to run on AWS?
- 14:16Need for generalizability in healthcare models
- 21:20Important criteria for selecting FL frameworks
- 23:16TensorFlow Federated
- 28:37FedML architecture
- 29:49System architecture
- 31:39Prerequisites Single account - multiple instances Instances
- 32:24Multiple accounts: Cross-silo FL clients with multi-GPU distributed training
- 35:27What you will see
- 42:57Evaluation metrics: Performance metrics
- 43:15Evaluation metrics: Model metrics
- 44:01Acknowledgments
related talks
31mFrom Theory to Practice: Advances in Real-World Federated Learning
31mFederated Machine Learning with Kubernetes
18mEnabling Medical AI through Federated Learning
26mMachine Learning without the Data #federatedlearning
42mFederated Learning: Unlocking the Values of Discrete Data for Enterprise- 31mFederated Machine Learning with Kubernetes
34mTraining without sharing: Is Federated Learning the answer to data privacy in AI?
32mInnovation in the Age of Regulation: Federated Learning with Flower