Informations pratiques

  • Journée de lancement du Colloquium IA/ML/DS
  • Date : 22 février 2018
  • Lieu : grand Amphi, site Saint-Charles Université d’Aix-Marseille
    • 3 place Victor Hugo, 13003 Marseille (lien Google map)
    • Bâtiment n°2 sur le plan
  • Inscription
    • gratuite mais obligatoire (capacité amphi, plan vigipirate)
    • clôture le 20/02/2018
    • formulaire d’inscription : ici

Planning

  • 09h00 – 09h15 L. Ralaivola & P. Pudlo, ouverture de la journée
  • 09h20 – 10h15 V. Ventos, LRI, INRIA TAO. « Le bridge, nouveau défi de l’intelligence artificielle ? »
  • 10h20 – 10h45 Pause
  • 10h50 – 11h45 J.-J. Schneider, La Provence Innovation. « La Provence, un acteur de la Presse Quotidienne Régionale dans le monde de l’IA »
  • 13h45 – 14h40 S. Gelly, Google. « Deep Generative Models »
  • 14h45 – 15h40 J. Mary, Criteo. « Fighting boredom in advertising through linear reinforcement learning »
  • 15h45 – 16h40 T. Peel, Euranova. « Machine Learning à l’ère du RGPD »
  • 16h45 – 17h00 L. Ralaivola & P. Pudlo, clôture de la journée

Résumés

V. Ventos. « Le bridge, nouveau défi de l’intelligence artificielle ? »

Games have always been an excellent field of experimentation for the nascent techniques in compu- ter science and in different areas of Artificial Intelligence (AI) including Machine Learning (ML). Despite their complexity, game problems are much easier to understand and to model than real life problems. Systems initially designed for games are then used in the context of real applications. In the last decades, designs of champion-level systems dedicated to a game (game AI) were considered as milestones of computer science and AI. Go and Poker are the two most recent successes. In May 2017, AlphaGo (DeepMind) defeated by 3 to 0 the Go world champion Ke Jie. In January 2017, the Poker AI Libratus (Carnegie Mellon University) won a heads-up no-limit Texas hold’em poker event against four of the best professional players. This success has not yet happened with regard to another incomplete information cards game, namely Bridge, which then provides a challenging problem for AI. We think that Deep Learning (DL) cannot be the only AI future. There are many Machine Learning and more generally AI fields which can interact with DL. Bridge is a great example of an application needing more than black box approaches. The AlphaBridge project is dedicated to the design of a Bridge AI taking up this challenge by using hybrid framework in the field of Artificial Intelligence.

The first part of the talk is devoted to the presentation of the different aspects of bridge and of various challenges inherent to it. In a second part, we will present our work concerning the 2optimization of the AI Wbridge5 developped by Yves Costel. This work is based on a recent seed methodology (T. Cazenave, J. Liu and O. Teytaud 2015, 2016) which optimizes the quality of Monte-Carlo simulations and which has been defined and validated in other games. The Wbridge5 version boosted with this method won the World Computer-Bridge Championship twice, in Sep- tember 2016 and in August 2017. Finally, the last part is about various ongoing works related to the design of a hybrid architecture entirely dedicated to bridge using recent numeric and symbolic Machine Learning module.

J.-J. Schneider. « La Provence, un acteur de la PQR dans le monde de l’IA »

Dans la première partie de notre exposé, nous présenterons La Provence acteur majeur de la PQR et la transition opérée en 2017 par la direction avec la création de La Provence Innovation, centre de recherche et de développement en IA pour la valorisation de l’information. Nous présenterons ensuite les programmes déployés depuis septembre 2017 :

  • programme research network
  • programme académie network
  • partenariats technologiques

ainsi que les programmes à venir au printemps 2018.

Nous terminerons notre exposé par une présentation des projets de R&D actuellement en cours de développement ainsi que des méthodes et technologies envisagées.

S. Gelly. « Deep Generative Models »

Deep Generative Models, like Generative Adversarial Networks (GAN), or Variational AutoEnco- ders (VAE) are effective methods to model distributions in high dimension, like images. However, both model classes suffer from limitations. GANs are notoriously hard to train, instable, and suffer from the problem of missing modes where the model is not able to produce examples in certain regions of the space. VAEs are much more stable, cover all the modes, but they produce blurry images. In this talk, we will present some of the known issues and existing workarounds. We will also give more details on some works done in our research group to understand the under- lying issues, and propose some improvements. In particular, we conducted a large scale study of different GAN models, comparing them on equal grounds. We will also present the Wassertein Autoencoders, which can be seen as an alternative to VAEs.

J. Mary. « Fighting boredom in advertising through linear reinforcement learning »

This talk will present the core business as Criteo and the most important current challenges we currently face in terms of research. One of this challenges is the handling of the dynamics of the user in a recommender systems which occurs both at recommendation and creative optimization. In this talk, we first cast this problem as a Markov decision process, where the rewards are a linear function of the recent history of actions, and we show that a policy considering the long-term influence of the recommendations may outperform both fixed-action and contextual greedy policies. We then introduce an extension of the ucrl algorithm to effectively balance exploration and exploitation in an unknown environment, and we derive a regret bound that is independent of the number of states. Finally, we empirically validate the model assumptions and the algorithm in a number of realistic scenarios.

T. Peel. « Machine Learning à l’ère du RGPD »

Le 25 mai prochain le Règlement Général sur la Protection des Données entrera en application. A cette date, les entreprises traitant des données à caractère personnel dans l’Union Européenne devront se conformer à une réglementation stricte sous peine d’importantes sanctions financières. Dans la première partie de cette présentation, nous traverserons les grandes étapes d’un projet de Machine Learning en nous attardant sur les points qui seront les plus impactés par le RGPD. Nous verrons ainsi que le respect de “bonnes pratiques” dans la réalisation de tels projets permet dans la plupart des cas d’être conforme avec les préconisations du règlement.

Dans un second temps, nous mettrons en lumière les récents travaux de la communauté scientifique visant à rendre les modèles de Machine Learning moins “obscures” avant de mettre en avant les atouts, notamment vis à vis du RGPD, de la mise en place de tels algorithmes.