Sites Inria

Version française

ICML Conference 2015

24/06/2015

Machine learning is meeting up in Lille

For the first time in its history, the International Conference on Machine Learning (ICML 2015) will be taking place in France. Inria Lille - Nord Europe is taking part in the organization of this unmissable event for the machine learning community.

With almost 1,500 participants expected in Lille between July 6-11, this 32nd edition of the ICML conference has once and for all entered a new era. "Up until 2010, there were only a few hundred participants, recalls Philippe Preux, SequeL project team manager (a joint venture between Lille 3 and Lille 1 universities*) and co-organizer of the event.The dynamism of this field of research is reflected in the constant increase in the number of participants: 750 in Edinburgh in 2012, 850 in Atlanta in 2013 and 1,250 in Beijing in 2014!"

Created in 1980 in Pittsburgh at Carnegie-Mellon University (CMU) - particularly renowned for its laboratories for artificial intelligence, machine and robot learning - the conference became an annual meeting from 1988 onwards. "At that time, and even though the idea of machine learning dates back to at least the origins of computer science (and notably cybernetics) during the 1940s and '50s, this was a very new field of study, Philippe Preux explains. As was often the case with computer science, the North American community grew rapidly; it then gradually spread across the globe. Moreover, France was involved in the development of this discipline from a very early stage."Inria hosts a large proportion of the French machine learning research community, in particular through the SequeL team at the Lille - Nord Europe center and Sierra at the Paris- Rocquencourt center. For this 32nd edition held in Lille, the event is supported by many partners, including the Métropole Européenne de Lille (MEL), the Nord-Pas de Calais region and Lille 1 University who are our closest partners.

Machine learning is capable of solving problems that remained unsolvable ten years ago. Great progress lies ahead, especially in areas of treatment linked to natural language thanks to deep learning. very exciting, because we still can't see any limits to what we can do!

An age of democratization

"The conference was originally quite basic, explains Philippe Preux.But over the last 15 years, machine learning has started to become more democratized with the emergence of the first applications for the general public, such as the Google search engine in 1997 and Amazon's recommendation system. These are two uses of machine learning that revolutionized the Internet and the world of business despite - or perhaps due to - their simplicity."

Since then, conceptual advances and their applications have multiplied. For example, this field of research makes it possible to detect undesirable emails, recognize objects in images...and even automate a car."Machine learning is capable of solving problems that remained unsolvable ten years ago, Philippe Preux analyses.However there is still a gap between what is actually done and what could potentially be done. Great progress lies ahead, especially in areas of treatment linked to natural language (recognition and understanding of written texts, speech, dialogue systems with humans) thanks to deep learning - a branch of machine learning. Some barriers still remain to be overcome in order to achieve intercommunication between learning software. It's very exciting, because we still can't see any limits to what we can do!"

The rendezvous of the web giants

The boom currently being experienced by machine learning can be explained by a combination of several factors. "Part of this is due to the power of IT tools and network speeds, says Philippe Preux. In addition, machine learning is closely linked to data science, which consists in extracting and exploiting "hidden" information within these data, whatever their source, format or volume."  With huge collections of data at their disposal, Internet giants such as Facebook or Google are of course very interested in the best way to exploit this digital wealth. The "sponsors" page of the conference's website - which includes Intel, Disney, Baidu and Microsoft, for example - is an illustration of this widespread interest."Numerous international companies are participating in or attending ICML because they are on the look-out for the latest ideas that will enable them to propose new services, solve new problems, and get ahead of their competitors, Philippe Preux reveals.They also come to ICML in order to recruit doctors and researchers. It has become a true source of competition for us, especially since they are equipped with extremely high-level laboratories."Moreover, around 20 companies will have stands at ICML.

Numerous international companies are participating in or attending ICML because they are on the look-out for the latest ideas that will enable them to propose new services, solve new problems, and get ahead of their competitors.

A busy program

For six days, the greatest specialists in the field will be presenting their work. The first day of ICML will be devoted to tutorials - high-level lectures - on, for example, the relationship between machine learning and the social sciences. The next three days will be dedicated to conferences and presentations of scientific publications. Finally, ICML will end with two days of workshops on various themes, from musical recommendation to deep learning."The open-source issue will also be broached, because most machine learning software is traditionally open-source", Philippe Preux adds.

Another of the conference's strong points is its support of young researchers. Over 100 doctoral students receive a travel grant from the NSF or ICML's company sponsors, and around another 100 receive free registration in exchange for a day of volunteer work to ensure the smooth running of the conference."The organization of an event like this requires a lot of support, energy and time, confirms Philippe Preux.But I don't regret putting myself forward to bring ICML to France. Considering the very high level of the French teams, it really wasn't normal that it had never taken place in our country."

* within UMR 9189 CNRS-Centrale Lille-University Lille1, CRIStAL

Keywords: Open source Sequel team Deep learning Machine learning ICML

Top