Federated Meta-Learning combines the principles of federated learning with meta-learning to train models that can quickly adapt to new tasks on distributed clients.
In federated learning, local models are trained on devices without centralizing the data to ensure privacy.
Meta-learning, on the other hand, focuses on learning learning algorithms or initializations that enable rapid adaptation to new tasks.
By combining these, the systems learn to build a shared knowledge base from the heterogeneous data of many clients, which can be efficiently transferred to new local tasks.
Typical algorithms use methods such as Model-Agnostic Meta-Learning (MAML) in a federated setting.