COLA: Decentralized Linear Learning

Authors: Lie He, An Bian, Martin Jaggi

NeurIPS 2018 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Here we illustrate the advantages of COLA in three respects: firstly we investigate the application in different network topologies and with varying subproblem quality ; secondly, we compare COLA with state-of-the-art decentralized baselines: 1 , DIGing [Nedic et al., 2017], which generalizes the gradient-tracking technique of the EXTRA algorithm [Shi et al., 2015], and 2 , Decentralized ADMM (aka. consensus ADMM), which extends the classical ADMM (Alternating Direction Method of Multipliers) method [Boyd et al., 2011] to the decentralized setting [Shi et al., 2014, Wei and Ozdaglar, 2013]; Finally, we show that COLA works in the challenging unreliable network environment where each node has a certain chance to drop out of the network.
Researcher Affiliation Academia EPFL lie.he@epfl.ch An Bian ETH Zurich ybian@inf.ethz.ch Martin Jaggi EPFL martin.jaggi@epfl.ch
Pseudocode Yes Algorithm 1: COLA: Communication-Efficient Decentralized Linear Learning
Open Source Code Yes Our implementation is publicly available under github.com/epfml/cola.
Open Datasets Yes Table 1: Datasets Used for Empirical Study URL 2M 3M 3.5e-5 Webspam 350K 16M 2.0e-4 Epsilon 400K 2K 1.0 RCV1 Binary 677K 47K 1.6e-3
Dataset Splits No No specific details about train/validation/test splits are provided in the main text. The order of columns is mentioned to be shuffled before distribution, but no split ratios.
Hardware Specification Yes The decentralized network topology is simulated by running one thread per graph node, on a 2 12 core Intel Xeon CPU E5-2680 v3 server with 256 GB RAM.
Software Dependencies No We implement all algorithms in Py Torch with MPI backend. No specific version numbers for PyTorch or MPI are provided.
Experiment Setup No Due to space limit, details on the experimental configurations are included in Appendix D. These details are not present in the provided main text.