Nuit Blanche Reader's Mailbag

First Muthu sent me an email and asked me if I would join him in hosting one of his experimental video hangouts on Google + so we can discuss compressed sensing algorithms and applications. I am thinking yes, it's an outstanding idea. Stay tuned for more details.

Silvio Ventres sent me the following:

Hello, Igor.

Found a nice more-or-less intuitive explanation of Belief-Propagation algorithm, which might be interesting for your less math-inclined readers:

http://www.cs.huji.ac.il/~yweiss/1010TPWeiss.pdf

And, following with the Rumelhart's research which triggered BP invention, here is a nice example of a model of visual recognition which shows BP at work with nice pictures, which should help understanding :)

http://jn.physiology.org/content/90/2/798.full

This might also be a good example of "CS in Nature", since effectively it implies that, at least for visual recognition of written text, a very sparse sampling of the stimuli should suffice, since it's afterwards decoded with a greedy BP algorithm, quite similarly to many CS experiments.

Thanks for continuing with high quality reporting! In this day of multitudes of rehashed news sources, your blog is a refreshingly enjoyable read !

Thanks Silvio !

And finally, Marco Duarte sent me the following as a follow-up to our Q&A this week:

Hi Igor, 

We have received some feedback from our discussion featured in your blog. In particular, Phil Schniter pointed out to us that there exist structured sparsity models and corresponding recovery algorithms that exploit correlations between signal coefficients; these are non-i.i.d. models. We think it is appropriate to clarify that our remarks on your blog were not sufficiently precise, and we considered only standard (e.g., non-structured) sparsity models in our comparison to the state of the art.

Additionally, we noticed that in our discussion we incorrectly said "at the same time, our reconstruction performance is as good as (superior) Bayesian reconstruction for any signal despite not knowing its statistics." At this point, we can only conjecture that we can perform nearly as well as Bayesian reconstruction.

Thanks Marco, Dror and Phil  for enriching this discussion.

Image Credit: NASA/JPL/Space Science Institute. N00177354.jpg was taken on October 19, 2011 and received on Earth October 20, 2011. The camera was pointing toward ENCELADUS at approximately 65,634 kilometers away, and the image was taken using the P0 and GRN filters. This image has not been validated or calibrated. A validated/calibrated image will be archived with the NASA Planetary Data System in 2012.

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.