Distribution Alignment for One-Shot Federated Learning via Optimal Transport
Researchers have introduced SLOT-Align, a novel framework designed to harmonize feature representations in One-Shot Federated Learning (OSFL). This method addresses challenges posed by heterogeneous client data distributions, specifically domain and label shifts, which existing OSFL techniques struggle to correct. SLOT-Align employs a shared frozen encoder and optimal transport maps to align local representations efficiently, demonstrating consistent improvements in accuracy and robustness across various benchmarks. AI
IMPACT Enhances robustness and accuracy in federated learning scenarios with extreme communication constraints.