1.

Parallel draws from the Polya-Gamma distribution for faster Bayesian multinomial and count model inference

We proposed a GPU-parallelized algorithm for sampling from the Polya-Gamma distribution, which is key in recently-developed Gibbs-sampling schemes for mutlinomial and count models. With this parallelization, we can already achieve 20-fold speedups, but the final goal is to parallelize the entire models, which would lead to speedups of order 100.

COBISS.SI-ID: 1537224387