3 thoughts on “NUTS discussed on Xi’an’s Og

  1. It went up last week – http://arxiv.org/abs/1304.1920 – just in time for my move across the pond. Marco’s presentation was solid, but we really need to move past focusing on choosing epsilon and L. epsilon and the total integration time are really the natural parameters of interest.

  2. Yes, it would have been nice if the presentation linked had provided better references. Or any reference, in the case of the split Hamiltonian method that is described on slides 70 and 71, with algorithm descriptions taken from page 9 of this paper. Reference to my blog posts on the No-U-Turn sampler found here and here would also have been a good idea in connection with the example given on slides 58 and 59.

  3. Just noticed this post!
    I agree (and I am most sorry to Mr Neal) that the reference list was far from being complete but I’d like to say in my defense that I did not expect the presentation to leave the room where it was held, I could not imagine it cross the ocean!
    As stated in the very first slide it was intended to present NUTS during the informal bayesian reading seminar that take here in Paris, in front of (mostly) fellow PhD students and relatives advisors. I sure said where all the material came from, including the blog entries, but didn’t put everything on latex (because of laziness mostly, I admit it, but also to keep the references to a number the audience could actually read..)
    To be fair, there’s also a quote from M.Hoffmann on the same incriminated slides.

    I’ll take that (not to underestimate the internet effect) as an advice for future work.

    M. Betancourt paper was found the day before the presentation (as you can imagine from being put on the very last slide with just a few comments!) on the daily arxiv check. I have indeed high expectation for stan/NUTS making use of it and I’ll sure continue to check Michael work on the subject especially since (if I understand correctly) he’s going to work with M.Girolami for a while.
    I also agree that L is just a “tuning gear” rather than a real quantity of interest, but since the correspondence is obvious the leapfrog is still easier to express with it..I understand however how starting from NUTS this is going to fade!

    Thanks for the post anyway, let’s hope that it helps in raising NUTS/HMC popularity..

Comments are closed.