The meetup will streamed online at this address or you can watch it right below. Please read below for the context of these presentations (in English and French)
First, thank you to our sponsors for this 10th Paris Machine Learning meetup. They are numerous, these events just wouldn't happen without them. First is our historic partner DojoCrea, that allowed us to grow the meetup to what it currently is. We will be hosted at TheFamily this time in their new location. Hopwork has kindly offered to cover the food and refreshments. The meetup will start at 6:45PM with the beginning of the talks at 7:30PM (all Paris time). Besides the first presentation and Q&A, it is likely that the other presentations will be in French.
Let's give some context:
There are many important problems in Machine Learning besides strictly Learning the Identity (exploitation), there is also the exploration part. We will cover the multi-Armed bandit issues at the next meetup but when it comes to pure exploration, unsupervised learning is one of the hardest problem when trying to make sense of something that nobody has really understood before: Animal Communication.
In order to get a sense of what we know and what we don't know, what sort of datasets are available, etc... on these truly hard problems, we will play the video that Brenda McCowan did at the Analyzing Animal Vocal Sequences Investigative Workshop at Nimbios this past fall, and then we will have her in a remote Q&A session live from California. Here is the video for those of you who are thinking about specific questions (you can add those in the comment section if you want me to ask them)
As isomorphismes said in a comment of a blog post featuring the videos of that workshop,
Fantastic. Secondary school teachers should point their kids here to prove maths isn't boring.
Yes, I agree, plus any kid I know has always wanted to try to talk to Flipper or Skippy, a far cry from spending your next working life on worrying about click-thrus :-)
We will then have Philippe Nieuwbourg, live from Montreal, who will give us a bird's eyeview of how traditional businesses are changing thanks to the use of their own data.
At Meetup#8, Lenka Zdeborova ( How hard is it to find a needle in a haystack? ) showed us that certain problems in Machine Learning (clustering or other matrix factorization) can only be explored through heuristics (i.e. easier problems to solve ) but that the downside meant the appearance of sharp or not so sharp phase transitions (sometimes it works, sometimes it never works). The only way to avoid these problems is to directly solve the hard problems that are combinatorial in nature even if that means that one has to develop new technologies such as Quantum Computers. Guillaume Palacios, formerly a physicist at CEA will tell us what he thinks of the D-Wave Quantum computers. To give more context and show that this is not a simple thought experiument, Google has a Quantum AI lab team looking into what these computers can do.
Finally, Gabriel Synnaeve will introduce the subject of ConvNets. This presentation is a follow-up of last week's presentation by Pierre Sermanet ( Deep ConvNets; "Astounding" baseline for vision, ) that we co-organized at Normale Sup at the Paris Machine Learning Specialist Talk. If you recall we had Pierre also on an earlier meetup when he talked to us how he used this type of algorithm to Win Kaggle's Dog's vs Cats contest. Because the subject is getting considerable notice, Gabriel will go over the results mentioned by Pierre last week on these ConvNets that are revolutionizing Computer Vision in nearly all known benchmarks (and allow people to win Kaggle competitions). Gabriel will also present to us the structure of these algorithms. Thank you to him to do this on such short notice.
Out group now has more than 751 members with about 168 who RSVP'ed that they would come.
Now the same announcement in French:
Pour cette 10eme edition nous aurons quatre presentations : Philippe Nieuwbourg, Guillaume Palacios, Brenda McCowan et Gabriel Synnaeve
Nous serons dans les nouveaux locaux de TheFamily.
Au dela du networking habituel qui se passera en premiere partie et apres les presentations, les themes du meetup vont etre varies. Nous aurons deux presentations remote une en anglais, l'autre en francais. Les deux autres presentations seront en Francais. Les slides (en anglais) se trouveront sur les archives prochainement.
Nous montrerons une presentation de Brenda McCowan qui nous parlera de la communication entre/avec les dauphins. Cette video est deja sur YouTube et se trouve ici (http://www.youtube.com/watch?v=X-Nr41KLGaY ). A la suite de la projection (comme nous l'avions fait pour le meetup 8 avec les projets de crowdfunding), nous aurons Brenda en direct de Californie pour repondre a nos questions. L'idee de cette presentation est de montrer qu'au dela des applications traditionelles du Machine Learning, ces techniques peuvent aussi nous permettre de faire des choses tres differentes de leur utilisation habituelle. Franck et moi pensons qu'il est important que vous regardiez la video auparavant de sorte a ce que nous puissions poser de bonnes questions a Brenda. Si vous avez des questions en avance, envoyez les nous et nous les ferons passer a Brenda avant le Q&A
Philippe Nieuwbourg nous parlera en direct de Montreal pour nous donnez un apercu plus general de l'utilisation des donnees dans des domaines divers.
Nous aurons aussi deux presentations de membre de notre groupe de meetup.
Au meetup #8, nous avions vu, grace a Lenka Zdeborova, que certains problemes de ML comme le clustering ou d'autres problemes de factorization de Matrices (meetup #9) ne pouvaient etre aborde qu'a travers des relaxations (c'est a dire des problemes voisins plus facile a resoudre) et que cela avait, des fois, l'incovenient d'engendrer des problemes de transition de phase (des fois ca marche, des fois cela ne marche absolument jamais). Le seul moyen d'eviter ces problematiques etant de resoudre directement des problemes combinatoriels quitte a developer de nouvelles technologies telle que l'ordinateur quantique. Guillaume Palacios, ancien physicien au CEA, nous fera une introduction sur le sujet et nous decrira ce qu'il pense du projet d'ordinateur quantique de DWave. L'un des ces ordinateurs ayant ete acquis par Google.
Finalement, Gabriel Synnaeve nous parlera de ConvNets. Cette presentation fait suite a celle de Pierre Sermanet (meetup #8 ) qui etait a Paris la semaine derniere et que Gabriel et nous avons accueilli a Normale Sup (Paris Machine Learning Specialist Talk). Parceque le sujet est d'une actualite brulante, Gabriel nous reparlera des resultats montres par Pierre sur ces fameux ConvNets qui revolutionnent depuis quelques mois tout les benchmarks existants (et fait gagner dans les competitions Kaggle utilisant des images). Gabriel nous decrira brievement la structure de ces algorithmes et ce qu'il y a sur le net pour les implementations. Merci a lui de faire cela au pied levé.
Les abstracts:
+ Unraveling dolphin communication complexity: Past approaches and next steps Brenda McCowan
La presentation sur YouTube:http://www.youtube.com/watch?v=X-Nr41KLGaY
+ Transformez vos données en dollars, par Philippe Nieuwbourg, analyste indépendant www.decideo.fr
Comment l'analyse prédictive et les données volumineuses bouleversent certains secteurs traditionnels : études de cas dans le domaine de la mode, des études comportementales, de la distribution et de la culture.
+ "The D-Wave quantum computer: Myths and Realities"
La societe Canadienne D-Wave produit le 1er ordinateur quantique a vocation commerciale. Des geants comme Google, la NASA ou encore Amazon comptent parmi les investisseurs de D-Wave et sont aussi les tout premiers utilisateurs de leur machine. Mais qu'y a-t-il dans cette machine etrange quand on souleve le capot ? Le but de ce talk est de donner des elements de reponse et de reflexion. Apres une courte introduction au calcul quantique, nous introduirons le concept de quantum optimization, au coeur du fonctionnement de la machine D-Wave. Nous verrons comment cette technique s'applique a la resolution de problemes combinatoires dit NP-hard et mentionnerons quelques applications aux algorithmes de Machine Learning.
+ ConvNets:
Toutes les presentations seront sur les archives: http://nuit-blanche.blogspot.com/p/paris-based-meetups-on-machine-learning.html
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.
No comments:
Post a Comment