Pages

Wednesday, February 12, 2014

Ce soir/Tonight: Paris Machine Learning Meetup #8: Finding a needle in a Haystack, Beyond SGD, Analyzing Wikipedia, Kolibree, Winning Kaggle "Dogs vs Cats"


Using Advanced Matrix Factorization to perform not so trivial signal/image/video processing tasks ( for more see the videos at  It's CAI, Cable And Igor's Adventures in Matrix Factorization )


The streaming of tonight's meetup on Google Hangout is here and it should start at around 7h30pm Paris Time. The talks will be in French this time with slides (see below in English).

The meetup itiself should start at around 6:45pm.

The idea for this series of meetup is to bring together curious people, researchers, professionals in the same room. The attendant meetup group now has more than 570 members. Let us note that there is now a Berlin ML meetup group and a Machine Learning group in Zurich led by Martin Jaggi whom we have mentioned here before.

In particular, our focus is not on the actual technology (mostly web tech) but on the underlying algorithms that are becoming central to a large section of the economy. What is fascinating are really two aspects of this phenomena. On the one hand, we are producing more data than there are stars in the universe (and this is increasing at a higher rate with the Internet of Things), on the other hand, the field of machine learning is really maturing as a consequence of this new influx of data. Tonight,  Lenka Zdeborova will talk to us about some of the deep limits associated with some of the problems found in Machine Learning such as clustering and other advanced Matrix Factorizations. She and her collaborators are making significant contributions to our knowledge on what computations are feasible and those that are not feasible. In a different direction Francis Bach will show us some of the latest development on getting algorithms to perform online gradient descent, the central battery of many ML algorithms. Central to this theme is the issue of having access to the data only a few times because they are so large. In future meetups, we will have people talking to us about streaming/sketching and related issues. Pierre Sermanet will show us the strategy he used to win First Prize on the Kaggle contest on Dogs vs Cats. Here again, the deep neural networks used did not exist in this form until very recently. The level of accuracy of some of these contests make some people say that some of these algorithms are getting better than Humans. We are still at an early stage here because these deep networks require much tuning, but quite clearly, the trend is becoming clear for certain tasks. At the very least, there is this nagging question on how some of the Matrix Factorization techniques can help better understand how to tune these deep network algorithmsLoïc Cessot will talk to us about Kolibree , the world's first connected toothbrush as he is looking for Machine Learners in order to enrich the data produced by this hardware. Guillaume Pitel, will talk about how he explores Wikipedia with a specific Matrix Factorization technique of his own. 

The program and attendant speaker's slides:
Other references:

1 comment: