Thursday, April 03, 2014

From Direct Imaging to Machine Learning ... a rapid panorama (JIONC 2014)

I have attended, for the past three years, the "Journées d'imagerie optique non-conventionnelle" (or JIONC for short) organized at ESPCI (don't look, there is currently no website for these meetings) and felt that the generic dichotomy of looking at sensors and then reconstruction solvers needed some perspective. So I decided to put in an abstract for a presentation two hours before the deadline. As it turns out, the committee was overwhelmed with abstracts this time (more than double the usual number) and they had to deal with different constraints, one of which was a time constraint. I was accepted for a poster which was not what I had expected. Then again, it was pretty much a blessing in disguise as I somehow honed my story several times with the few unwitting people who stumbled close to the dark lit corner of the room where poster was located. To them, thank you and sorry at the same time :-). 

The whole "poster" is located here: From Direct Imaging to Machine Learning ... a rapid panorama. Any insight on how it could be better presented is more than welcome. Many shortcuts were taken, the reference section is dismal but the point was trying to give a bird's eye view of fields that are currently on a collision course, especially if you look at them with the lens (pun intended) of advanced matrix factorizations.

JIONC was the same day as the IoT presentation so it took me a little while to clean up the original poster.

In the meantime, few readers have pointed me to the TechReview on Patrick Gill's lensless camera, I'll come to that later because it is a nice illustration of some of the comments made during the presentation of this poster. 

Tonight a 7:00pm, Pierre Sermanet is in Paris and will talk to us about OverFeat [1, 3]. Why ? because as mentioned in the poster, there is a convergence between deep learning and generic sensing. In particular, in this case, you rarely see the word "astounding" from a different team working with your approach [2]. Hence the invitation for him to speak, two days ago, to our ML and not so ML people.

As an aside, if Yann LeCun, Yoshua Bengio or Andrew Ng or any of the big thoughts leaders in ML are bored in Paris and want to give a talk to our audience, it looks like that with a 2-day notice, we can get a location and more than 150 people to attend. 

[1] OverFeat: Integrated Recognition, Localization and Detection using Convolutional Networks, Pierre Sermanet, David Eigen, Xiang Zhang, Michael Mathieu, Rob Fergus, Yann LeCun,

[2] CNN Features off-the-shelf: an Astounding Baseline for Recognition, Ali Sharif Razavian, Hossein Azizpour, Josephine Sullivan, Stefan Carlsson,

No comments: