Exploring Normalization for High Convergence on Federated Learning for Drones

Abstract

The usage of mobile devices like drones has been increasing in various fields, ranging from package delivery to emergency services and environmental monitoring. Intelligent services increasingly use the processing power of these devices in conjunction with techniques such as Federated Learning (FL), which allows machine learning to be carried out in a decentralized way using data accessed by clients or devices. However, in normal operations, the data accessed by clients is distributed heterogeneously among themselves, negatively impacting learning results. This article discusses the normalization in Federated Learning local training to mitigate results obtained in heterogeneous distributions. In this context, we propose Federated Learning with Weight Standardization on Convolutional Neural Networks (FedWS) and evaluate it with Batch Normalization, Layer Normalization, and Group Normalization in experiments with heterogeneous data distributions. The experiments demonstrated that FedWS achieved higher accuracy results ranging from 3% to 6% and reduced the computational and communication costs between 25% and 40%, being more suitable for use in devices with computational resource limitations.