Federated Multi-Task Learning under a Mixture of Distributions

Authors: Othmane Marfoq, Giovanni Neglia, Aurélien Bellet, Laetitia Kameni, Richard Vidal

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

Reproducibility Variable Result LLM Response
Research Type Experimental Experimental results on FL benchmarks show that our approach provides models with higher accuracy and fairness than state-of-the-art methods.
Researcher Affiliation Collaboration 1Inria, Université Côte d Azur, France, {othmane.marfoq, giovanni.neglia}@inria.fr 2Inria, Université de Lille, France, aurelien.bellet@inria.fr 3Accenture Labs, France, {richard.vidal, laetitia.kameni}@accenture.com
Pseudocode Yes Algorithm 1: Fed EM (see also the more detailed Alg. 2 in App. D.1)
Open Source Code Yes Code is available at https://github.com/omarfoq/Fed EM.
Open Datasets Yes We evaluated our method on five federated benchmark datasets spanning a wide range of machine learning tasks: image classification (CIFAR10 and CIFAR100 [33]), handwritten character recognition (EMNIST [8] and FEMNIST [7]),5 and language modeling (Shakespeare [7, 47]).
Dataset Splits Yes For all tasks, we randomly split each local dataset into training (60%), validation (20%) and test (20%) sets.
Hardware Specification No The paper does not provide specific hardware details (e.g., exact GPU/CPU models, memory amounts, or detailed computer specifications) used for running its experiments.
Software Dependencies No The paper mentions 'Mobile Net-v2 [55]' and 'Stacked-LSTM [25]' as models but does not specify software dependencies with version numbers (e.g., Python 3.8, PyTorch 1.9, TensorFlow 2.x).
Experiment Setup Yes For each method and each task, the learning rate and the other hyperparameters were tuned via grid search (details in App. I.2). Fed Avg+ updated the local model through a single pass on the local dataset. Unless otherwise stated, the number of components considered by Fed EM was M = 3, training occurred over 80 communication rounds for Shakespeare and 200 rounds for all other datasets. At each round, clients train for one epoch.