Bayesian Lasso and multinomial logistic regression on GPU

Area of Science:

• Computational Statistics
• Machine Learning

Background:

• Bayesian statistical models like Lasso and multinomial logistic regression are computationally intensive.
• Efficient implementation is crucial for large-scale data analysis.

Purpose of the Study:

• To present an efficient Bayesian parallel GPU implementation of the Lasso and multinomial logistic regression models.
• To demonstrate significant speedups using graphics processing units (GPUs).

Main Methods:

• Parallelization of key computational components: matrix multiplication, matrix inversion, and sampling.
• Implementation using OpenCL for broad GPU compatibility.
• Focus on Bayesian inference for statistical modeling.

Main Results:

• Achieved up to 100-fold speedups for Bayesian Lasso and multinomial logistic regression on mid-to-high-end GPUs.
• Demonstrated substantial 25-fold speedups on older and lower-end GPUs.
• OpenCL implementation offers practical convenience and advantages over existing methods.

Conclusions:

• The proposed GPU implementation offers a highly efficient and practical solution for Bayesian statistical modeling.
• This approach significantly reduces computation time for Lasso and multinomial logistic regression.
• The use of OpenCL ensures broad applicability across diverse computational hardware.