During the 1st Italian Conference on Big Data and Data Science (ITADATA 2022) in Milan, our collaborators Bruno Casella, Roberto Esposito, Carlo Cavazzoni, and Marco Aldinucci from CINI and Leonardo presented their paper titled “Benchmarking FedAvg and FedCurv for Image Classification Tasks” during the session Big Data and AI II held on 21st of September 2022 from 11:00-12:00 CEST.
This paper provides an empirical assessment of the behaviour of FedAvg and FedCurv in common non-IID scenarios. Results show that the number of epochs per round is an essential hyper-parameter that, when tuned appropriately, can lead to significant performance gains while reducing communication costs.
With this contribution, FedAvg produced better models in most non-IID settings despite competing with an explicitly developed algorithm to improve in this scenario. Interestingly, both algorithms perform better when the number of epochs per round increases (reducing communication costs).
Even though this is a new observation, the researchers aim to investigate it shortly. Among the tested datasets, those implementing the quantity and pathological label skew pose the most complex challenges to the algorithms. Also, as expected, the quantity skew appears to be less challenging.
Numerous experiments are currently underway to extend this work. In particular, The EUPILOT experts study the impact of different normalisation layers in Federated Learning and further investigate the relationship between epochs and rounds.
These results are significant for The EUPILOT project as they will help contribute to the project’s goals to build a sustainable exascale HPC supply ecosystem in Europe and ensure European technological autonomy in this field.
Please read the complete publication here.