FedPara: Low-rank Hadamard Product Parameterization for Efficient Federated Learning
share
To overcome the burdens on frequent model uploads and downloads during federated learning (FL), we propose a communication-efficient re-parameterization, FedPara. Our method re-parameterizes the model's layers using low-rank matrices or tensors followed by the Hadamard product. Different from the conventional low-rank parameterization, our method is not limited to low-rank constraints. Thereby, our FedPara has a larger capacity than the low-rank one, even with the same number of parameters. It can achieve comparable performance to the original models while requiring 2.8 to 10.1 times lower communication costs than the original models, which is not achievable by the traditional low-rank parameterization. Moreover, the efficiency can be further improved by combining our method and other efficient FL techniques because our method is compatible with others. We also extend our method to a personalized FL application, pFedPara, which separates parameters into global and local ones. We show that pFedPara outperforms competing personalized FL methods with more than three times fewer parameters.
READ FULL TEXT