This paper investigates the impact of non-identical data distributions on federated visual classification using the Federated Averaging (FedAvg) algorithm. The authors propose a method to synthesize datasets with varying levels of identicalness and evaluate the performance of FedAvg under these conditions. They find that performance degrades as the distributions become more non-identical, with classification accuracy dropping from 30.1% to 76.9% in the most skewed settings. To mitigate this issue, they introduce server momentum, which improves performance by synchronizing weights more frequently. Experiments on the CIFAR-10 dataset demonstrate that FedAvgM (FedAvg with server momentum) achieves better classification accuracy, often approaching the centralized learning baseline. The study also explores the hyperparameter sensitivity of FedAvgM, showing that careful tuning is necessary to achieve optimal performance, especially when the reporting fraction is low.This paper investigates the impact of non-identical data distributions on federated visual classification using the Federated Averaging (FedAvg) algorithm. The authors propose a method to synthesize datasets with varying levels of identicalness and evaluate the performance of FedAvg under these conditions. They find that performance degrades as the distributions become more non-identical, with classification accuracy dropping from 30.1% to 76.9% in the most skewed settings. To mitigate this issue, they introduce server momentum, which improves performance by synchronizing weights more frequently. Experiments on the CIFAR-10 dataset demonstrate that FedAvgM (FedAvg with server momentum) achieves better classification accuracy, often approaching the centralized learning baseline. The study also explores the hyperparameter sensitivity of FedAvgM, showing that careful tuning is necessary to achieve optimal performance, especially when the reporting fraction is low.