Making Batch Normalization Great in Federated Deep Learning

Batch Normalization (BN) is commonly used in modern deep neural networks (DNNs) to improve stability and speed up convergence during centralized training. In federated learning (FL) with non-IID decentralized data, previous works observed that training with BN could hinder performance due to the mismatch of the BN statistics between training and testing. Group Normalization (GN) is thus more often used in FL as an alternative to BN. However, from our empirical study across various FL settings, we see no consistent winner between BN and GN. This leads us to revisit the use of normalization layers in FL. We find that with proper treatments, BN can be highly competitive across a wide range of FL settings, and this requires no additional training or communication costs. We hope that our study could serve as a valuable reference for future practical usage and theoretical analysis in FL.