Optimization for Classical Machine Learning Problems on the GPU

Authors: Sören Laue, Mark Blacher, Joachim Giesen7300-7308

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

Reproducibility Variable Result LLM Response
Research Type Experimental Experiments on several classical machine learning problems show that our implementation outperforms state-of-the-art approaches like the combination of CVXPY with GPU-accelerated solvers such as cu OSQP or SCS by several orders of magnitude.
Researcher Affiliation Collaboration 1 Friedrich-Schiller-University Jena, Germany 2 Data Assessment Solutions Gmb H
Pseudocode Yes Algorithm 1: GPU-efficient L-BFGS-B Method
Open Source Code Yes The code for the solver is available the github repository https://www.github.com/slaue/genosolver.
Open Datasets Yes We used the adult data set (48, 842 data points with 120 features) and the census-income data set (299, 285 data points with 400 features) each with female and male as the two subgroups. We used all data sets from the Lib SVM data sets website (Lin and Fan 2021) that had more than 8000 data points with fewer than 1000 features such that a kernel approach is reasonable.
Dataset Splits No The paper states 'For each experiment, we sampled m data points from the full data set' but does not provide explicit training, validation, or testing dataset splits, percentages, or absolute sample counts for data partitioning.
Hardware Specification Yes All experiments were run on a machine equipped with an Intel i9-10980XE 18-core processor running Ubuntu 20.04.1 LTS with 128 GB of RAM, and a Quadro RTX 4000 GPU that has 8 GB of GDDR6 SDRAM and 2304 CUDA cores.
Software Dependencies No The paper mentions software like Python, NumPy, and CuPy but does not provide specific version numbers for these dependencies (e.g., 'Cu Py library (Okuta et al. 2017)' refers to a paper, not a version).
Experiment Setup Yes We applied a standard Gaussian kernel with bandwidth parameter γ = 1 and regularization parameter c = 1. The regularization parameter was set as λ = 1/2.