Eugéne Suter,

Universidade de Évora

Abstract:

This talk presents the result of a one year BII devoted to the implementation of natural gradient algorithms of probability distributions, and more specifically to the Kullback-Leibler optimisation.
A software library was developed, in C language, taking advantage of multicore CPUs and NVidia GPUs (with CUDA).
Problems tackled include relation between memory hierarchy and data organisation, loop optimisation and parallel processing. Numerical issues arising in large problems and some performance tweaks are also considered.
Benchmarks comparing several approaches are shown.

 

Date: 2009-Dec-11     Time: 14:00:00     Room: 04


For more information: