FedImp: New Algorithm Accelerates Federated Learning Convergence
Researchers have introduced a novel algorithm, Federated Impurity Weighting (FedImp), to tackle issues linked to non-IID data in federated learning. This method evaluates the significance of local data on each device to determine their respective contributions, thereby refining aggregation weights for the model. In experiments conducted using the EMNIST and CIFAR-10 datasets, FedImp demonstrated a substantial enhancement in convergence speed, cutting down communication rounds by notable percentages, such as 64.4% on EMNIST. Compared to traditional methods like FedAvg, FedProx, and FedAdp, FedImp proved to be superior, especially in scenarios with unbalanced data distribution. The research is available on arXiv under ID 2608.14654.
Key facts
- FedImp is a novel algorithm for federated learning.
- It addresses non-IID data distributions across devices.
- It quantifies device contributions based on informational content.
- Experiments on EMNIST and CIFAR-10 datasets.
- Reduces communication rounds by up to 64.4% on EMNIST.
- Reduces communication rounds by up to 44.2% on CIFAR-10.
- Outperforms FedAvg, FedProx, and FedAdp.
- Paper available on arXiv (2608.14654).
Entities
Institutions
- arXiv