Federated-EM with heterogeneity mitigation and variance reduction

Authors: Aymeric Dieuleveut, Gersende Fort, Eric Moulines, Geneviève Robin

NeurIPS 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Numerical results are presented to support our theoretical findings, as well as an application to federated missing values imputation for biodiversity monitoring.
Researcher Affiliation Academia Aymeric Dieuleveut Centre de Math ematiques Appliqu ees Ecole Polytechnique, France Institut Polytechnique de Paris aymeric.dieuleveut@polytechnique.edu Gersende Fort Institut de Math ematiques de Toulouse Universit e de Toulouse; CNRS gersende.fort@math.univ-toulouse.fr Eric Moulines Centre de Math ematiques Appliqu ees Ecole Polytechnique, France CS Dpt, HSE University, Russian Federation eric.moulines@polytechnique.edu Genevi eve Robin Laboratoire de Math ematiques et Mod elisation d Evry Universit e d Evry Val d Essonne; CNRS Evry-Courcouronnes, France genevieve.robin@cnrs.fr
Pseudocode Yes Algorithm 1: Fed EM with partial participation and Algorithm 2: VR-Fed EM
Open Source Code No The paper does not provide any statement or link indicating that the source code for the methodology is openly available.
Open Datasets Yes MNIST Data set. We perform a similar experiment on the MNIST dataset to illustrate the behaviour of Fed EM and VR-Fed EM on a GMM inference problem with real data. We apply Fed Miss EM to the analysis of part of the e Bird data base [34, 1] of field observations reported in France by I = 2,465 observers, across J = 9,721 sites and at L = 525 monthly time points.
Dataset Splits No The paper describes data distribution across servers and minibatch usage but does not explicitly provide training/validation/test dataset splits or mention a 'validation' set.
Hardware Specification No The paper mentions 'edge devices' and 'central server' but does not provide specific hardware details (e.g., CPU/GPU models, memory) used for running the experiments.
Software Dependencies No The paper does not provide specific software dependencies (e.g., library or solver names with version numbers) needed to replicate the experiments.
Experiment Setup Yes For Fed EM, we consider the finite-sum setting when si = m^-1 Pm j=1 sij with m = 10^2; the oracle Sk+1,i is obtained by a sum over a minibatch of b = 20 examples. For VR-Fed EM, we set b = 5 and kin = 20. We run the two algorithms for 500 epochs (one epoch corresponds to N conditional expectation evaluations sij). We choose b = 20 and the step size is constant and set to γ = 10^-3. We run Fed Miss EM for 150 epochs; with γ = 10^-4, = 10^-3, b = 10^2, a rank r = 2 and λ = 0; for the distribution of the variables Xi j , we use a Gaussian distribution with unknown expectation j and variance 1.