Hello,
I would like to confirm the problem setting (especially of task 1). In my understanding, in task 2, the training program generates four models, each model is trained by using the data from each center. In task 1, can the training program access to data from all centers to generate a global model ? Or can the training program to generate a global model access to only the local model information as the federated learning setting ?
Created by SKJP Hi,
All models (the global model and each client's local model) should be generated and stored during the training with your federated learning setup.
**Task 1:**
The global model refers to the server's model, which is formed by aggregating updates from local models trained on each center's data. The server cannot access raw data from the centers directly. Instead, it receives local model updates (e.g., gradients or parameters) from each center, consistent with the federated learning setting. The global model's performance is evaluated on data from a new, previously unseen client, measuring its generalization ability.
**Task 2:**
In federated learning, each center performs multiple training rounds independently using only its own data. After each round, the model is sent to the server, which aggregates the updates and generates an updated global model to be used as the initial model for the next round. You can store the best model for each training center. The performance of these local models is assessed on their respective datasets, reflecting how well each model performs on its own local data.
**Summary:**
- **Task 1:** The training program can only access local model updates from each center to generate the global model, following the federated learning approach.
- **Task 2:** Each center trains its own local model using its own data based on the global model, and performance is evaluated locally on each center?s data.
I hope that makes the task clearer.