Monday, July 13, 2015

Video: Size- and level-adaptive Markov chain Monte Carlo

Dror let me know that his group is in the video production mode as a way to illustrate clearly some of the subject areas he and his students are working on. This tutorial video was made by Junan Zhu.


Join the CompressiveSensing subreddit or the Google+ Community or the Facebook page and post there !
Liked this entry ? subscribe to Nuit Blanche's feed, there's more where that came from. You can also subscribe to Nuit Blanche by Email, explore the Big Picture in Compressive Sensing or the Matrix Factorization Jungle and join the conversations on compressive sensing, advanced matrix factorization and calibration issues on Linkedin.

1 comment:

Dror Baron said...

Igor, thanks for sharing with the community without my even pointing it out. Your ability to retrieve information is always impressive.

I'd like to mention to the audience that our other video on universal denoising and approximate message passing (AMP-UD algorithm) may be the one you prefer to watch. The reason is that the Monte Carlo (MCMC) approach is slower, and can't reach the minimum mean square error (MMSE), which is the best-possible reconstruction performance that one might hope for (I'm assuming that you're interested in square error; for unconventional types of error, see Jin Tan's video from 2013).

In contrast, AMP-UD is reasonably fast (signals of length 10,000 are processed in under 5 minutes), it gets really close to the MMSE for i.i.d. inputs that we've looked at, and on non-i.i.d. inputs - including real data - it usually reconstructs better than the MCMC approach shown in this video.

Feel free to watch both. If you like them, we'll keep making them!