An Efficient Subspace Algorithm for Federated Learning on Heterogeneous Data

Name: An Efficient Subspace Algorithm for Federated Learning on Heterogeneous Data
Uploaded: 2026-10-04T02:27:58.445305+00:00
Duration: 810 s
Description: FedSub is an efficient federated learning algorithm tackling client drift and high costs in heterogeneous data settings. It uses low-dimensional subspace projections to drastically reduce communication, computation, and memory requirements. Dual variables are added to correct model drift caused by non-IID data distributions, leading to better convergence than standard methods.

FedSub is an efficient federated learning algorithm tackling client drift and high costs in heterogeneous data settings. It uses low-dimensional subspace projections to drastically reduce communication, computation, and memory requirements. Dual variables are added to correct model drift caused by non-IID data distributions, leading to better convergence than standard methods.

Open the full episode