On his blog, Thomas Mailund describes his attempts to speed up hidden Markov models. To do this he is taking advantage of the parallelization of single instruction, multiple data operations, and multiple cores -- though he is running into difficulties because there are many different ways to handle the data in parallel and if that's not done correctly, it leads to a large number of classes. He says that he is now thinking about using a template meta-programming approach that would "weave together the different data representations with the optimal algorithms for working on them."
Markov Model, Maserati Speed
Mar 12, 2009