2.1 Federated learning and terminology
What is federated learning?
Federated learning has the following features
- Training private data on models locally, so the data will NOT upload to the cloud. The cloud global model only receive the gradients from each local model to update itself.
- Ensemble dataset might performs better than mixing all dataset.
- Unbalanced and Non-IID (non-independent and non-identical distribution) problems among different edges might be solved by certain algorithms.
You can refer to the paper Communication-Efficient Learning of Deep Networks from Decentralized Data that first propose the concepts of federated learning for more details.
Roles
As the previous chapter introduced, we briefly summarize as the below table:
| Roles | Example | Tasks order | Dashboard |
| Edge Admin | Hospital IT Team | 1. Create a project and assign it to PI | Aggregator Dashboard |
| PI | Project Owner | 0. wrap code into an image 2. Invite Co-PIs to join the project 3. Upload image 4. Setup a FL/FV plan 6. Start training/validation 7. See training/validation result |
Aggregator Dashboard |
| Co-PI | Project participants | 5. Upload dataset to the plan | Edge Dashboard |
More hints:
- Task 0: means that PI should modify the code to fit the interface of our platform then wrap it into a docker image. It will be fully introduced in 2.3 The input and output interfaces.
- Task 1: Projects are initiated by Edge Admin and assigned to PI.
- Task 3: A project can have multiple images, PI can choose one for a plan.
- Task 4: PI is the project host who can create multiple plans under the project.
- Task 4: A plans can be training or validation.
- Task 4: Plans are mutually independent which hold different parameters just as conducting experiments with different parameters.
What is AI Labs Federated Learning Dashboard?
AI Labs Federated Learning Dashboard is a platform for federated learning developed by AI Labs. The GUIs contains two types. One is Aggregator Dashboard for Edge Admin and PIs, and the other is Edge Dashboard for Co-PIs.
What is Flavor?
PI needs to upload a docker image that involves customized training or inference code that fits the format into the interface of AI Labs Federated Learning Dashboard. Flavor is a library installed in a docker image which helps user to fit the format that our platform needs.
Federated Learning (FL) and Federated Validation (FV)
A complete machine learning includes training steps and inference steps. To avoid confusion, the training stage is called Federated Learning (FL) and the validation stage is called Federated Validation (FL) from now on.
More references
This document bridge the gap between the following guides: