<ul><li>Personalized federated learning has become popular for training on devices with different data, but typically requires labeled data for personalization.</li><li>FLowDUP is a new method proposed in this paper that can generate personalized models using only unlabeled data with a forward pass.</li><li>The model parameters generated by FLowDUP are in a low-dimensional subspace, allowing for efficient communication and computation.</li><li>Experimental evaluation of FLowDUP shows strong empirical performance on various datasets with unlabeled clients, supported by theoretical results.</li></ul>

Federated Learning with Unlabeled Clients: Personalization Can Happen in Low Dimensions

Discover more