Learning Algebraic Multigrid Using Graph Neural Networks

Authors: Ilay Luz, Meirav Galun, Haggai Maron, Ronen Basri, Irad Yavneh

ICML 2020 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Experiments on a broad class of problems demonstrate improved convergence rates compared to classical AMG, demonstrating the potential utility of neural networks for developing sparse system solvers.
Researcher Affiliation Collaboration 1Weizmann Institute of Science, Rehovot, Israel. 2NVIDIA Research 3Technion, Israel Institute of Technology, Haifa, Israel.
Pseudocode Yes Algorithm 1 Two-Level Algorithm
Open Source Code Yes Code for reproducing experiments is available at https://github.com/ilayluz/learning-amg.
Open Datasets No The training data are comprised of block-circulant graph Laplacian matrices, composed of 4x4 blocks with 64 points in each block, yielding 1024 variables. ... To this end, we sample points uniformly on the unit square, and compute a Delaunay triangulation.
Dataset Splits No The paper describes generating training data and evaluating performance but does not specify explicit train/validation/test dataset splits by percentage or sample count.
Hardware Specification Yes All experiments were conducted using the Tensor Flow framework (Abadi et al., 2016) using NVIDIA V100 GPU.
Software Dependencies No The paper mentions 'Tensor Flow framework' and 'Adam optimizer' but does not provide specific version numbers for these software components.
Experiment Setup Yes We use a batch size of 32 and employ the Adam optimizer (Kingma & Ba, 2014) with a learning rate of 3x10^-3.