Error Feedback Reloaded: From Quadratic to Arithmetic Mean of Smoothness Constants

Authors: Peter Richtárik, Elnur Gasanov, Konstantin Pavlovich Burlachenko

ICLR 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Finally, we validate our theoretical findings with suitable experiments.
Researcher Affiliation Academia Peter Richt arik AI Initiative KAUST , Saudi Arabia; Elnur Gasanov AI Initiative KAUST, Saudi Arabia; Konstantin Burlachenko AI Initiative KAUST, Saudi Arabia
Pseudocode Yes Algorithm 1 EF21: Error Feedback 2021
Open Source Code No If you are interested in the source code for all experiments, please contact the authors.
Open Datasets Yes We utilized six datasets from LIBSVM (Chang & Lin, 2011).
Dataset Splits Yes The dataset shuffling strategy, detailed in Appendix I.5, was employed to emulate heterogeneous data distribution. Each client was assigned the same number of data points.
Hardware Specification Yes We carried out experiments on a compute node with Ubuntu 18.04 LTS, 256 GBytes of DRAM DDR4 memory at 2.9GHz, and 48 cores (2 sockets with 24 cores per socket) of Intel(R) Xeon(R) Gold 6246 CPU at 3.3GHz. All our computations were carried on CPU.
Software Dependencies No We used the Python software suite FL Py Torch (Burlachenko et al., 2021) to simulate the distributed environment for training.
Experiment Setup Yes The initial gradient estimators were chosen as g0 i = fi(x0) for all i [n]." and "The coefficient λ for (b) (f) is set to 0.001, and for (a) is set to 1, 000 for numerical stability." and "The number of clients n is 1, 000." and "The step size for EF21 is set according to (Richt arik et al., 2021), and the step size for EF21-W is set according to Theorem 3."