Efficient low-rank federated learning based on singular value decomposition
Jungmin Kwon, Hyunggon Park · 2022
In this paper, we propose a low-rank federated learning (FL) algorithm based on singular value decomposition (SVD). The SVD factorizes the global parameters that need to be exchanged between a global server and clients for distributed model training, significantly reducing the associated communication cost. Experiment results confirm that the number of transmissions is significantly reduced while maintaining the accuracy performance of the local model using the approximately recovered parameters.