In this project we analyze in depth different federated learning algorithms. We first start selecting a convolutional neural network and compute a baseline for reference; we then proceed to evaluate FedAVG, the first and most adopted solution in the federated scenario. Moreover, other federated algorithms are considered, in order to address different issues, for instance stagglers, client model complexity and classifier calibration. For each of the aforementioned algorithms we consider one of the most challenging problems in federated learning, which is dealing with non-IID data distribution, and compare the results.
[File with compiled tables] https://github.com/dadodimauro/MLDL-Federated-Learning/blob/b775decafa6aa228f8d0734a8408af7aff296d0e/Project6_Group1_TablesResults.pdf
[Excel file with all results] https://github.com/dadodimauro/MLDL-Federated-Learning/blob/b775decafa6aa228f8d0734a8408af7aff296d0e/RESULT_SUMMARY.xlsx
[Report Presentation] https://github.com/dadodimauro/MLDL-Federated-Learning/blob/021673edf0c42c97820c31ec9fda8d79452af9c2/Project6_Group1_ReportPresentation.pdf