BiCompFL: Stochastic Federated Learning with Bi-Directional Compression
Keywords: Communication-efficiency, Importance sampling, Stochastic federated learning
Abstract: Communication is a prominent bottleneck in federated learning (FL). State-of-the-art accuracy performance under limited uplink communications from the clients to the federator is achieved by stochastic FL approaches. It has been recently shown that leveraging side information in the form of a prior distribution at the federator can drastically reduce the uplink communication cost in stochastic FL. Here, the latest global model distribution serves as a natural prior since it can be shared with the clients under ideal downlink communication from the federator to the clients. Nevertheless, downlink communication is often limited in practical settings, and bi-directional compression must be considered to reduce the overall communication cost. The extension of existing stochastic FL solutions to bi-directional compression is non-trivial due to the lack of a globally shared common prior distribution at each iteration. In this paper, we propose BiCompFL, which employs importance sampling to send samples from the updated local models in the uplink, and the aggregated global model in the downlink by carefully choosing common prior distributions as side-information. We theoretically study the communication cost by a new analysis of importance sampling that refines known results, and exposes the interplay between uplink and downlink communication costs. We also show through numerical experiments that BiCompFL enables multi-fold savings in communication cost compared to the state-of-the-art.
Primary Area: probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)
Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.
Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.
No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.
Submission Number: 3833
Loading