A0644
Title: Accelerating dependency modeling with graphics processing units
Authors: Po-Hsien Huang - National Chengchi University (Taiwan) [presenting]
Abstract: Dependency modeling among response variables is a crucial task in multivariate analysis. A general strategy for this task is to introduce latent variables (or random effects) to capture the common part of the variables. By integrating out the latent variables, the so-called marginal maximum likelihood (MML) can be conducted for parameter estimation. However, when the number of latent factors is large, the MML generally becomes infeasible. We demonstrate how to use graphics processing units (GPU) computing and vectorization to greatly speed up the training process of MML. In particular, the MML for item factor analysis (IFA) is considered. IFA could be understood as a generalization of factor analysis for handling polytomous data. A python package called xifa was developed. Our numerical experiments show that xifa could be 33 times faster than its CPU counterpart. Furthermore, when the number of latent factors is equal to or larger than 5, xifa is much faster than the competing implementations, including the Bock-Aitkin expectation-maximization and the MHRM implemented by mirt (on CPU), and the importance-weighted autoencoder (on GPU). We believe GPU computing would play a central role in large-scale statistical modeling in the near future.
