Advances in scalable Bayesian computation [day #3]
We have now gone over the midpoint of our workshop Advances in Scalable Bayesian Computation with three talks in the morning and an open research or open air afternoon. (Maybe surprisingly I chose to stay indoors and work on a new research topic rather than trying cross-country skiing!) If I must give a theme for the day, it would be (jokingly) corporate Big data, as the three speakers spoke of problems and solutions connected with Google, Facebook and similar companies. First, Russ Salakhutdinov presented some hierarchical structures on multimedia data, like connecting images and text, with obvious applications on Google. The first part described Boltzman machines with impressive posterior simulations of characters and images. (Check the video at 45:00.) Then Steve Scott gave us a Google motivated entry to embarrassingly parallel algorithms, along the lines of papers recently discussed on the ‘Og. (Too bad we forgot to start the video at the very beginning!) One of the novel things in the talk (for me) was the inclusion of BART in this framework, with the interesting feature that using the whole prior on each machine was way better than using a fraction of the prior, as predicted by the theory! And Joaquin Quinonero Candela provided examples of machine learning techniques used by Facebook to suggest friends and ads in a most efficient way (techniques remaining hidden!).
Even though the rest of the day was free, the two hours of exercising between the pool in the early morning and the climbing wall in the late afternoon left me with no energy to experiment curling with a large subsample of the conference attendees, much to my sorrow!