Federated Learning with Communication Constraints: Challenges and Recent Results

Explore federated learning optimization with compression in this 59-minute presentation from Institut des Hautes Etudes Scientifiques (IHES). Dive into recent findings on convergence and compression rate tradeoffs, as well as user-heterogeneity impacts. Examine two key phenomena: the effect of user-heterogeneity on federated optimization methods under communication constraints, and the robustness of distributed stochastic algorithms to iterate perturbation. Learn about a novel compression scheme utilizing random codebooks and unitary invariant distributions. Gain insights from Aymeric Dieuleveut of CMAP/Ecole polytechnique on addressing challenges in federated learning with communication constraints.