Registered Data

[00972] Reducing Communication in Federated Learning with Variance Reduction Methods

  • Session Time & Room : 4E (Aug.24, 17:40-19:20) @D408
  • Type : Contributed Talk
  • Abstract : In Federated Learning $\text{(}$FL$\text{)}$, inter-client heterogeneity and partial participation of clients at each communication cause client sampling error. We control this client sampling error by developing a novel single-loop variance reduction algorithm. While sampling a small number of clients, the proposed FL algorithms require provably fewer or at least equivalent communication rounds compared to any existing method, for finding first and even second-order stationarypoints in the general nonconvex setting, and under the PL condition.
  • Classification : 90Cxx, 68Wxx, 68Txx
  • Format : Talk at Waseda University
  • Author(s) :
    • Kazusato Oko (The University of Tokyo, AIP RIKEN)
    • Shunta Akiyama (The University of Tokyo)
    • Tomoya Murata ( The University of Tokyo, NTT DATA Mathematical Systems Inc.)
    • Taiji Suzuki (The University of Tokyo, AIP RIKEN)