LoCoDL: Communication-Efficient Distributed Learning with Local Training and Compression

Authors: Laurent Condat, Artavazd Maranjyan, Peter Richtarik

ICLR 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We evaluate the performance of our proposed method Lo Co DL and compare it with several other methods that also allow for CC and converge linearly to x . We also include Grad Skip (Maranjyan et al., 2022) and Scaffold (Mc Mahan et al., 2017) in our comparisons. We focus on a regularized logistic regression problem, which has the form (1) with ... We show the results with the a5a , diabetes , w1a datasets in Figures 1, 2, 3, respectively.
Researcher Affiliation Academia Laurent Condat, Artavazd Maranjyan & Peter Richtárik Computer Science Program, CEMSE Division King Abdullah University of Science and Technology (KAUST) Thuwal, 23955-6900, Kingdom of Saudi Arabia & SDAIA-KAUST Center of Excellence in Data Science and Artificial Intelligence (SDAIA-KAUST AI) EMAIL
Pseudocode Yes Algorithm 1 Lo Co DL
Open Source Code No The paper does not contain an explicit statement about releasing the source code for the Lo Co DL method, nor does it provide a link to a code repository.
Open Datasets Yes We considered several datasets from the Lib SVM library (Chang & Lin, 2011) (3-clause BSD license). We show the results with the a5a , diabetes , w1a datasets in Figures 1, 2, 3, respectively. ... Finally, we also run experiments on MNIST dataset (Le Cun et al., 1998) in Figure 6.
Dataset Splits No We prepared each dataset by first shuffling it, then distributing it equally among the n clients (since m in (11) is an integer, the remaining datapoints were discarded). This describes how data was distributed among clients but does not specify explicit training, validation, or test splits for evaluating the model.
Hardware Specification No The paper does not provide any specific details about the hardware (e.g., GPU models, CPU types, memory) used for running the experiments.
Software Dependencies No The paper does not provide specific version numbers for any software components or libraries used in the experiments.
Experiment Setup Yes We focus on a regularized logistic regression problem, which has the form (1) with ... µ is the regularization parameter, set so that κ = 104. ... For all algorithms, we used the theoretical parameter values given in their available convergence results (Corollary 3.2 for Lo Co DL). We tried to tune the parameter values, such as k in rand-k and the (average) number of local steps per round, but this only gave minor improvements. For instance, ADIANA in Figure 1 was a bit faster with the best value of k = 20 than with k = 30. Increasing the learning rate γ led to inconsistent results, with sometimes divergence.