<ul><li>Federated learning (FL) aims to enhance privacy and scalability by not sharing local data with a central server.</li><li>In FL, dataset imbalance can occur due to unequal label representation across network agents, impacting global model aggregation and local model quality.</li><li>An Optimal Transport-based preprocessing algorithm is introduced to align datasets by minimizing distributional discrepancies, leveraging Wasserstein barycenters for channel-wise averages.</li><li>The proposed approach demonstrates improved generalization capabilities over the CIFAR-10 dataset by reducing variance and achieving higher levels of generalization in fewer communication rounds.</li></ul>

Optimal Transport-based Domain Alignment as a Preprocessing Step for Federated Learning

Discover more