Decentralized Federated Machine Learning: Empowering Edge Devices with Kubernetes
Talkat Kubernetes on Edge Day North America 2024 · on Tuesday, 2024-11-12· at 15:20· Salt Palace | Level 2 | 250 A·
summary
Deploys FLARE/Q-FL on Kubernetes (Minikube) with a server pod and two client pods, uses CIFAR-10, Docker image from Docker Hub, and Federated averaging for aggregation. Runs five federated rounds over ~10,000 eval images and ~1,563 local training examples per client; five rounds take ~45 minutes on the given Minikube resources and raise accuracy from ~30% to ~52%. Demonstrates working end-to-end training, shows CPU/memory observability, and requires explicit Minikube CPU/memory tuning for practical runtimes.
AI-written summary · about
chapters (16)
- 0:00Decentralized federated ML intro
- 1:47Traditional ML challenges
- 3:33Federated learning benefits
- 5:19Federated learning process
- 8:09FATE architecture overview
- 10:40FATE Kubernetes integration
- 11:21FLARE framework intro
- 13:27FLARE Kubernetes deployment
- 14:56Demo setup and resources
- 17:05Building Q FLARE images
- 18:31Server and client config
- 20:26Global model init
- 21:39Client training and eval
- 22:42Aggregation and updates
- 23:51Results and accuracy
- 25:17Q&A
related talks
31mFederated Machine Learning with Kubernetes- 31mFederated Machine Learning with Kubernetes
21mWhat, Why and How of Federated Machine Learning – Implementation Using...
27mFederated Learning with Azure Machine Learning NVIDIA FLARE and MONAI
31mBuilding and Managing a Centralized ML Platform with Kubeflow at...
48mFederated ML for Edge Applications
46mAccelerating Machine Learning DevOps with Kubeflow
34mDecentralized Machine Learning