Fast Composite Optimization and Statistical Recovery in Federated Learning

Authors: Yajie Bao, Michael Crawshaw, Shan Luo, Mingrui Liu

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

Reproducibility Variable Result LLM Response
Research Type Experimental Numerical experiments in both synthetic and real data demonstrate that our methods perform better than other baselines.
Researcher Affiliation Academia Yajie Bao 1 Michael Crawshaw 2 Shan Luo 1 Mingrui Liu 2 1School of Mathematical Sciences, Shanghai Jiao Tong University, Shanghai, China 2Department of Computer Science, George Mason University, Fairfax, VA 22030, USA. Correspondence to: Mingrui Liu <mingruil@gmu.edu>.
Pseudocode Yes Algorithm 1 Fast-Fed DA(w0, R, E, µ, L), Algorithm 2 C-Fed DA(w0, R, E, ϵ0, µ, L, λ), Algorithm 3 Multi-stage C-Fed DA
Open Source Code No The paper mentions computing resources (URL: https://orc.gmu.edu) but does not provide any explicit statement about releasing its own source code or a link to a code repository for the methodology described.
Open Datasets Yes Federated EMNIST (Caldas et al., 2019) dataset of handwritten letters and digits. We train a multi-class logisitc regression model on two versions of this dataset: EMNIST-10 (digits only, 10 classes), and EMNIST-62 (all alphanumeric characters, 62 classes). Following (Yuan et al., 2021), we use only 10% of the samples, which is sufficient to train a logistic regression model. Our subsampled EMNIST-10 dataset consists of 367 clients with an average of 99 examples each, while EMNIST-62 consists of 379 clients with an average of 194 examples each.
Dataset Splits No The paper mentions training on EMNIST datasets and evaluating 'Train Accuracy' and 'Test Accuracy' in its figures, but it does not specify a 'validation' dataset split or provide details on how the data was partitioned for training, validation, and testing.
Hardware Specification No Computations were run on ARGO, a research computing cluster provided by the Office of Research Computing at George Mason University (URL: https://orc.gmu.edu).
Software Dependencies No The paper describes the algorithms and experiments but does not list specific software dependencies with version numbers.
Experiment Setup Yes For federated sparse linear regression... the batch size is 10 and the regularization parameter is λ = 0.55. The hyperparameters for Fast-Fed DA are µ = 0.1 and L = 550. For C-Fed DA and MC-Fed DA, we choose µ = 0.1 and L = 600. For Fed DA and Fed Mi D, we set the server learning rate ηs = 1.0 and tuned the client learning rate ηc by selecting the best performing value over the set {0.0001, 0.001, 0.01, 0.1}, which was 0.001 for both baselines. ...For both experiments, we use a batch size of 25, a regularization parameter λ = 10 4, and we sample 36 clients to perform local updates at each communication round. For EMNIST-10, each sampled client performs K = 40 updates per communication round for R = 15000 rounds. For EMNIST-62, K = 10 and R = 75000.