We distinguish academic applications from those funded by companies. Academic applications are strongly related to our research. Those funded by companies are based on our expertize which spectrum is larger than our research topics.
- Language, vision, dialog systems, deep reinforcement learning:
- recommandation systems
- vision by deep learning with CHRU (hospital) of Lille (starting in 2019)
- reinforcement learning and robotics, with team Defrost (starting in 2019)
- sustainable development, with CIRAD, and CIAT (starting in 2019)
Relations with companies
We have many connections with companies of various sizes, from all over the world.
Some applications we deal with in collaboration with companies, and some we have dealt with, in SequeL.
Please note that confidentiality issues strongly restrict the content of this page.
Recommendation systems is really a generic term, encompassing many different problems actually. Basically, the goal is to be able to provide some content based on a request, which is either explicit (such as a set of keywords given to a search engine), or implicit (such as the mere fact that you visit a certain webpage).
Recommendation systems are a field of applications in which we are working since 2010. We have had many collaborations with private companies on this topic. We also have more academic activities.
These applications are based on our research on bandit problems.
- on e-learning with OtherLang (2017) and Le Livre Scolaire (2017-2020)
- image recommendation with 500px (2016)
- cold-start problem with Nuukik (2014-2016)
- Music recommendation with Deezer (2013)
- We have also participated to the FUI Hermès project (2013-2016).
- From 2009 to 2011, we have participated to the VVU project (Ubiquitous Virtual Seller) aiming at designing and implementing adaptive agents that would advice clients on merchant web sites.
These applications are based on our research on reinforcement learning (deep and shallow), and approximate dynamic programming.
- an on-going collaboration involving a PhD student with Renault (2017-2020)
- an on-going collaboration involving a PhD student with Critéo (2017-2020)
- an on-going collaboration involving a PhD student with 55 (2015-2018)
- an on-going collaboration involving a PhD student with Oranges Labs (2014-2017)
We are investigating deep learning and apply it to vision tasks.
We work with companies:
- Sidexa (2017-2018)
- with What A Nice Place (2015-2016)
- Leroy-Merlin (2016)
We had a very fruitful with the company TBS on web server load prediction (2014-2015).
We had a very fruitful with the company Squoring (2010-2014). The topic was an exploration of how machine learning can help in software quality assessment and in software development.
- Ad selection for web portals
- Game of Go
- Approximation of photometric solids
- Game of Poker
- Control of anaerobic reactor
- Prediction of queues in supermarket
- Anomaly detection in loudspeakers, on the production line
Since 2009, we work on the problem of optimal selection of advertisement banners on web portals with Orange Labs.
To deal with these problems, we are investigating new methods and algorithms mixing statistics and combinatorial optimization.
- An interview of Rémi Coulom in Wired
- Some explanations about software playing Go (in french).
- More on Crazy Stone page.
Since 2009, we have begun a collaboration aiming aim at using machine learning methods to approximate photometric solids, in order to create synthetic images (computer graphics). This is basically a regression problem, but in an interesting setting with certain idiosyncracies; furthermore, the goal is not to maximize the accuracy of the predictor, but to obtain good-looking images, as fast as possible; good-looking images and maximizing an accuracy are not the same things.
Our participation to this project is related to the design of recommendation systems.
Jérémie Mary is working on the game of Poker; see his page for more information.
The research concerns the control of waste-water treatment by an anaerobic digestion process. organic waste-water is treated in a biological reactor by means of an anaerobic digestion process, which produces bio-gas (methane) that can be used to power electricity generators. The digestion of the organic material is done by an appropriate mixing of (mainly) two different species of bacteria, each doing a different job at a different step in the digestion process. Maintaining suitable conditions in the digester is essential for the viability of the bacterial population. The problem of designing a good controller is crucial. A mandatory condition is to stabilize the system and maintain the bacterial populations alive. Under these constraints, an additional objective is to optimize the production of bio-gas.
From the scientific standpoint, this problem is formalized as a partially observable Markov decision problem (POMDP), where the state dynamics is not perfectly known: there are several available models for the dynamics, but most of the parameters of these models are unknown. Besides, the observation process is very poor. We aim at designing an adaptive policy that is both secure and which optimizes the methane production.
This work is being held in collaboration with the Naskeo Environment spin-off (Paris), the INRA laboratory in Narbonne (France), and the COMORE team with INRIA-Sophia-Antipolis. This is supported by the INRIA as an ARC; see this project website.
Auchan is a major international group which operates more that 150 Hypermarkets worldwide. One of its crucial issues is to be able to predict many variables, such as the number of customers reaching the cashiers at a given time of the day, the number of breads to cook, etc. For each day, this is a functional prediction problem, which may be seen as non-stationary, because the customer habits evolve.
SequeL members and Auchan have currently investigated this problem by extensively using past events. State-of-the art inference methods have been implemented with excellent performance (not described here for confidentiality reasons). We have improved the prediction accuracy of about 25% percent.
The goal was to detect automatically loudspeaker flaws on the production line. This successful application involved the proper use of 1-class SVMs.