Agha Ali Raza

Publications/Efficient Machine Learning: Data Selection, Pruning, Active & Federated Learning

Conference · 2023

Learning Fast and Slow: Towards Inclusive Federated Learning

Muhammad Tahir Munir, Muhammad Mustansar Saeed, Mahad Ali, Zafar Ayyub Qazi, Ihsan Ayyub Qazi, Agha Ali Raza

ECML PKDD 2023, Turin, Italy · CORE A

Summary

Federated learning trains a shared model across many people's devices without collecting their private data. In practice, slow or low-end devices, common in developing countries, are dropped from training, which excludes their owners and can bias the model. Our system, Hasaas, keeps them in by giving slower devices a smaller version of the model sized to their hardware, and uses insights from the Central Limit Theorem to handle differences in the data across devices. In large-scale simulations and on a small real testbed, it performs robustly on accuracy, fairness, and convergence time compared with state-of-the-art schemes.

Cite this paper

Muhammad Tahir Munir, Muhammad Mustansar Saeed, Mahad Ali, Zafar Ayyub Qazi, Ihsan Ayyub Qazi, Agha Ali Raza. Learning Fast and Slow: Towards Inclusive Federated Learning. ECML PKDD 2023, Turin, Italy.

@inproceedings{munir2023learning,
  author = {Muhammad Tahir Munir and Muhammad Mustansar Saeed and Mahad Ali and Zafar Ayyub Qazi and Ihsan Ayyub Qazi and Agha Ali Raza},
  title = {{Learning Fast and Slow: Towards Inclusive Federated Learning}},
  booktitle = {ECML PKDD 2023, Turin, Italy},
  year = {2023},
  url = {https://aghaaliraza.com/publications/2023--ECML--Learning_Fast_and_Slow-Towards_Inclusive_Federated_Learning.pdf}
}

More on this theme

All 8 works on this theme →