Personalized Federated Learning through Local Memorization

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

ICML 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We show on a suite of federated datasets that this approach achieves significantly higher accuracy and fairness than state-of-the-art methods.
Researcher Affiliation Collaboration 1Inria, Universit e Cˆote d Azur, Sophia Antipolis, France 2Accenture Labs, Sophia Antipolis, France.
Pseudocode Yes Algorithm 1 k NN-Per (Typical usage) Learn global model using available clients with Fed Avg. for each client m [M] (in parallel) do Build datastore using Sm. At inference on x X, return hm,λm (x) given by (7) end for
Open Source Code Yes Code is available at https: //github.com/omarfoq/knn-per.
Open Datasets Yes We evaluate k NN-Per on four federated datasets spanning a wide range of machine learning tasks: language modeling (Shakespeare (Caldas et al., 2018; Mc Mahan et al., 2017)), image classification (CIFAR-10 and CIFAR100 (Krizhevsky, 2009)), handwritten character recognition (FEMNIST (Caldas et al., 2018)).
Dataset Splits Yes For FEMNIST and Shakespeare, we randomly split each local dataset into training (60%), validation (20%), and test (20%) sets. For CIFAR-10 and CIFAR-100, we maintained the original training/test data split and used 20% of the training dataset as validation dataset.
Hardware Specification No The paper does not provide specific hardware details (e.g., GPU/CPU models, memory) used for running its experiments.
Software Dependencies No The paper mentions using the 'FAISS library' and 'Proto NN' but does not specify their version numbers or any other software dependencies with versions.
Experiment Setup Yes In all experiments with CIFAR-10 and CIFAR-100, training spanned 200 rounds with full clients participation at each round for all methods. The learning rate was reduced by a factor 10 at round 100 and then again at round 150. For Shakespeare, 10% of clients were sampled uniformly at random without replacement at each round, and we trained for 300 rounds with a constant learning rate. For FEMNIST, 5% of the clients participated at each round for a total 1000 rounds, with the learning rate dropping by a factor 10 at round 500 and 750.