README.md

(𝙰𝙰𝚐𝚐𝙵𝙵) Pursuing Overall Welfare in Federated Learning
through Sequential Decision Making

ICML`24 | Poster | Paper | Video | Thread (X)

TL; DR

The server-side sequential decision-making framework for the performance uniformity in practical federated settings by constructing adaptive mixing coefficients used to aggregate local updates.

• See also Agnostic Federated Learning (AFL – ICML`19) (i.e., min-max optimization; distributionally-robust optimization) — 𝙰𝙰𝚐𝚐𝙵𝙵 can be regarded as a rectified version of AFL.
• See also Tiltied Empirical Risk minimization (TERM – ICLR`21) — 𝙰𝙰𝚐𝚐𝙵𝙵 can be viewed as an online version of TERM.

Reproduce Experiments

Setup

pip install -r requirements.txt

Implemented with: ( Ubuntu 20.04 LTS | Python 3.10.3 | CUDA 11.4 | cuDNN 8.3.2 )

Run

# to see all arguments
python main.py -h

# e.g., cross-silo setting
sh commands/cross_silo/berka/main_berka.sh
...

# e.g., cross-device setting
sh commands/cross_device/celeba/main_celeba.sh
...

Don't worry, all the data will be downloaded automatically in the specifid path.

Essence

• See ./server/aaggffserver.py with comments.
• You may change the decision loss, different OCO framework (or bandits, or even reinforcement learning setting).

If you find any interesting directions, please drop me a line for collaboration!

License

• For non-commercial use, this code is released under the MIT LICENSE.
• For commercial use, please contact Seok-Ju (Adam) Hahn (sjhahn11512@gmail.com).

Citation

@article{hahn24aaggff,
  title={Pursuing Overall Welfare in Federated Learning through Sequential Decision Making},
  author={Hahn, Seok-Ju and Kim, Gi-Soo and Lee, Junghye},
  booktitle={Proceedings of the 41st International Conference on Machine Learning},
  pages={17246--17278},
  year={2024},
  publisher={Proceedings of Machine Learning Resaerch, PMLR}
}