Federated Meta-Learning refers to a learning paradigm that integrates federated learning with meta-learning to develop models that can quickly adapt to new, client-specific tasks while keeping the data decentralized.
Formally, one considers a set of clients \( \{C_i\}_{i=1}^N \), each possessing its own local data distributions \( \mathcal{D}_i \). The goal is to find a meta-initialization \( \theta \) such that after a few steps of local adaptation \( \theta_i = \theta - \alpha \nabla_\theta \mathcal{L}_{\mathcal{D}_i}(\theta) \), the performance on the respective local task is maximized.
The learning process involves iterative aggregation of gradients or parameters across the clients, with privacy ensured by avoiding data transfer and, if applicable, additional techniques such as differential privacy or secure aggregation.
Federated Meta-Learning addresses challenges such as heterogeneous data distributions, communication constraints, and the balance between global generalization and local specialization.
Definition:
“Federated Meta-Learning is an approach that combines federated learning and meta-learning to train models that quickly adapt to new, distributed tasks without sharing local data.”
Source:
Jiang, Y., Huang, J., & Wang, Y. (2020). Federated Meta-Learning with Fast Convergence and Efficient Communication. Advances in Neural Information Processing Systems (NeurIPS).