New Federated Learning Algorithm FedOBP Introduces Element-Wise Personalization Method

Researchers have developed FedOBP (Federated Optimal Brain Personalization), a new algorithm addressing challenges in Federated Learning. The approach tackles data heterogeneity across client devices and limited mobile resources that can compromise model accuracy. FedOBP employs a quantile-based thresholding mechanism and introduces an element-wise importance score for parameter evaluation. This score extends Optimal Brain Damage pruning theory by incorporating a federated approximation of first-order derivatives from Taylor expansions. The algorithm focuses on model decoupling, separating global and personalized parameters to balance knowledge sharing with local adaptation. Personalized Federated Learning adapts shared global knowledge to specific local data distributions. The paper was announced on arXiv with identifier 2604.16574v1. FedOBP determines which parameters should be personalized to optimize performance across diverse client environments.