Bayesian Lasso and multinomial logistic regression on GPU

Rok Češnovar; Erik Štrumbelj

doi:10.1371/journal.pone.0180343

Bayesian Lasso and multinomial logistic regression on GPU

PLoS One. 2017 Jun 28;12(6):e0180343. doi: 10.1371/journal.pone.0180343. eCollection 2017.

Authors

Rok Češnovar¹, Erik Štrumbelj¹

Affiliation

¹ Faculty of computer and information science, University of Ljubljana, Večna pot 113, 1000, Ljubljana, Slovenia.

Abstract

We describe an efficient Bayesian parallel GPU implementation of two classic statistical models-the Lasso and multinomial logistic regression. We focus on parallelizing the key components: matrix multiplication, matrix inversion, and sampling from the full conditionals. Our GPU implementations of Bayesian Lasso and multinomial logistic regression achieve 100-fold speedups on mid-level and high-end GPUs. Substantial speedups of 25 fold can also be achieved on older and lower end GPUs. Samplers are implemented in OpenCL and can be used on any type of GPU and other types of computational units, thereby being convenient and advantageous in practice compared to related work.

MeSH terms

Bayes Theorem*
Computer Graphics*
Logistic Models*
Models, Statistical

Grants and funding

The work was supported by the Slovenian Research Agency (ARRS, https://www.arrs.gov.si/en/) applied project grant L1-7542. The funders had no role in study design, data collection and analysis, decision to publish, or preparation of the manuscript.