Retour à la vue des calendrier
Mardi 5 Juin
Heure: 10:30 - 12:00
Lieu: Salle B107, bâtiment B, Université de Villetaneuse
Résumé: Théorie de la complexité et géométrie des orbites du déterminant et du permanent
Description: Christophe Tollu Après un rappel sur les circuits arithmétiques et le problème de Valiant (VP vs VNP), je présenterai quelques résultats récents sur la "complexité déterminantale" du permanent, puis montrerai comment la version purement algébrique du problème VP vs VNP se prête à une reformulation en termes de géométrie des orbites du déterminant et du permanent (pour l'action d'un groupe algébrique sur les polynômes homogènes). Plusieurs ingrédients de base du programme de théorie géométrique de la complexité de Mulmuley et Sohoni seront présentés au cours de l'exposé bien que celui-ci ne soit pas "A crash course on Geometric Complexity Theory".
Jeudi 7 Juin
Heure: 11:30 - 13:00
Lieu: Salle B107, bâtiment B, Université de Villetaneuse
Résumé: Coends and proof equivalence in MLL2
Description: Paolo Pistone > Proof nets provide permutation-independent representations of proofs and are used to investigate coherence problems for monoidal categories. We investigate a coherence problem concerning Second Order Multiplicative Linear Logic MLL2, that is, the one of characterizing the equivalence over proofs generated by the interpretation of quantifiers by means of ends and coends. This equivalence is naturally induced by the usual second order translation of multiplicative units and connectives and is related to the investigations on the parametric models of System F.
> By adapting the "rewiring approach" used in the proof net characterization of the free *-autonomous category, we provide a compact representation of proof nets for a fragment of MLL2 related to the Yoneda isomorphism. We prove that the equivalence generated by coends over proofs in this fragment is fully characterized by the rewiring equivalence over proof nets.
Mercredi 13 Juin
Heure: 14:00 - 15:00
Lieu: TBA (Institut Galilée)
Résumé: Network Interdiction
Description: Joe Naoum-Sawaya Network interdiction is a class of leader-follower optimization problem that seeks to identify network components to disrupt and inflict a maximum damage to a network. The objective of such models is to study the structural connectivity of the network in order to identify vulnerabilities. The application areas are diverse and include energy, telecommunication, and supply chain networks among others. This talk will review two particular variations of network interdiction: connectivity disruption and flow disruption. The connectivity disruption model identifies the nodes in a network whose disruption minimizes the maximum number of connected node pairs. The flow disruption model identifies the edges whose disruption minimizes the maximum flow between sources and destinations. We will present optimization models as well as solution approaches. We will particularly focus on the cases where uncertainty is present in the edge weights and propose customized solution approaches based on robust optimization, cutting planes, and Benders decomposition. The proposed cutting planes and Benders decomposition exploit the structure of the underlying optimization model and allows the modeling and the solution of general classes of uncertainty sets.
Jeudi 14 Juin
Heure: 12:15 - 13:30
Lieu: Salle B107, bâtiment B, Université de Villetaneuse
Résumé: Deep Cooperative Reconstruction with Privacy Constraints
Description: Denis Maurel Nowadays, we can observe a multiplication of multi-view data in domains such as marketing, bank administration or even survey analysis. In this context, Machine Learning methods are used to analyze data from several heterogeneous sources (here called views) with the following problem: an individual described in some views might be missing in some other ones. This proliferation is accompanied by a global privacy awareness: one should never have access to data from all sources at once. To solve these problems, we propose a method called the Cooperative Reconstruction System (CRS) which aims at reconstructing missing individuals locally using information contained in external views without data transfer from a view to another.
Vendredi 15 Juin
Heure: 10:30 - 12:30
Lieu: Salle B107, bâtiment B, Université de Villetaneuse
Résumé: Definable Ellipsoid Method, Sums-of-Squares Proofs, and the Graph Isomorphism Problem
Description: Joanna Ochremiak The isomorphisms between two graphs can be described by the solutions of a system of polynomial inequalities and equations. We analyse the relative power of different proof systems which can be used to certify that a system corresponding to a pair of non-isomorphic graphs has no solution. Our results complete a full cycle of implications to show that, for the graph isomorphism problem, the Sherali-Adams, Polynomial Calculus and Sums-of-Squares proof systems are equally powerful, up to a constant loss in the degree. We prove this statement purely about the relative strength of proof systems through an excursion into the descriptive complexity of the ellipsoid method and bounded-variable infinitary logics. This is joint work with Albert Atserias.
Jeudi 21 Juin
Heure: 12:15 - 13:30
Lieu: Salle B107, bâtiment B, Université de Villetaneuse
Résumé: Proposition et consommation de contenus sur le web : le problème de la diversité.
Description: Lionel Tabourier La question de la diversité des contenus proposés et consommés sur le web apparaît tôt dans la littérature de recherche d'information. Pourtant, ce n'est que récemment que celle-ci a évolué vers un débat de société, parce qu'un nombre croissant d'utilisateurs des moteurs de recherche ou des plateformes de recommandation sont confontés à leurs effets secondaires néfastes.

Un des plus notables est l'enfermement dans des bulles d'information, c'est-à-dire l'exposition à des contenus de moins en moins divers, correspondant à un environnement culturel restreint. L'objet du projet ANR Algodiv (http://algodiv.huma-num.fr/) est d'étudier comment le concept de diversité est traduit et mis en oeuvre par les algorithmes du web. Dans cet exposé, je présenterai deux aspects de cette question, qui sont des chantiers de travail actifs du projet.

D'abord, nous examinons l'effet des algorithmes de recommandation sur la diversité proposée aux utilisateurs. Moyennant une définition adéquate de la notion de diversité sur une plateforme de recommandation, nous cherchons à "auditer" les archétypes de recommandation, c'est-à-dire à mesurer la tendance d'une méthode à enfermer l'utilisateur à plus ou moins long terme.

D'autre part, nous étudions les caractéristiques des pratiques de navigation des utilisateurs sur l'exemple de la navigation sur le site Melty (https://www.melty.fr/). Nous examinons à quel point ceux-ci tendent à consommer des contenus variés ou non en fonction de ce qui leur est proposé, de la période, du moyen d'accéder à l'information, etc.
Lundi 25 Juin
Heure: 14:00 - 15:00
Lieu: Salle B107, bâtiment B, Université de Villetaneuse
Résumé: The Arabic Ontology and Modernization of Lexical Semantic Resources
Description: Mustafa Jarrar The importance of lexical-semantic resources (linguistic ontologies, wordnets, thesauri, and dictionaries) is increasing in many modern application scenarios, such as multilingual big data, data governance, information retrieval, NLP, social networks, and others. In this talk we will present our experience in digitizing 150 multilingual lexicons and present the Arabic Ontology, which is a formal Arabic Wordnet with ontologically-clean content. We will also discuss how this content is ontologically we ll-founded and benchmarked to scientific advances rather than to speakers’ nai?ve beliefs as wordnets typically do, in addition to top levels and other ontology engineering challenges and mappings to other ontologies.
Mercredi 27 Juin
Heure: 14:00 - 15:00
Lieu: Salle Darwin, institut galilée
Résumé: Research Efforts at the Computational Approaches to Modeling Language Lab.
Description: Nizar Habash TBA
Jeudi 28 Juin
Heure: 14:00 - 15:30
Lieu: Salle B107, bâtiment B, Université de Villetaneuse
Résumé: Hyperparameter Optimization for Neural Networks
Description: Razvan Andonie We introduce a dynamic early stopping condition for Random Search optimization algorithms. We test our algorithm for SVM hyperparameter optimization for classification tasks, on six commonly used datasets. According to the experimental results, we reduce significantly the number of trials used. Since each trial requires a re-training of the SVM model, our method accelerates the RS optimization. The code runs on a multi-core system and we analyze the achieved scalability for an increasing number of cores.
Lundi 2 Juillet
Heure: 14:00 - 15:00
Lieu: Salle B107, bâtiment B, Université de Villetaneuse
Résumé: Multi-Arabic Dialect Applications and Resources
Description: Nizar Habash We present the Multi-Arabic Dialect Applications and Resources (MADAR) Project. MADAR is an effort to build parallel resources for 25 Arab city dialects including lexicons, parallel corpora, and orthographic and morphological annotation guidelines. The created resources have been used to develop dialect identification and machine translation applications. We discuss the challenges facing Arabic dialect modeling, as well as our solutions and results.
Mercredi 4 Juillet
Heure: 14:00 - 15:00
Lieu: Salle B107, bâtiment B, Université de Villetaneuse
Résumé: Cost Efficient Prediction of Wine Quality - A Machine Learning Approach
Description: Razvan Andonie The quality of wines can be assessed both from chemical/biological tests and sensory tests (which rely mainly on human experts). Determining which is the subset of tests to be used is a difficult problem. Each test has its own contribution for predicting the quality of wines and, in addition, its own cost. We use our own database, consisting of 32 wine characteristics applied to 180 wine samples. In addition we use wine quality labels assigned by a wine expert. To the extent of our knowledge, this is the first study of this kind on wines from Washington State, and also the first wine study in general to include cost minimization of the measurements as a goal. Our approach is based on two stages. First, we identify reasonably good classifiers (from a given set of classifiers). Next, we search for the optimal subset of features to maximize the performance of the best classifier and also minimize the overall cost of the measurements. As a result, through our method we can answer queries like ``the best performing subset of tests for a given threshold cost’'.
Jeudi 5 Juillet
Heure: 12:15 - 13:30
Lieu: Amphi Copernic, Institut Galilée, Université de Villetaneuse
Résumé: Towards more Autonomous Robots
Description: Eduardo Morales With the increasing incorporation of robots into daily life
activities, autonomy and interaction with non expert users play a
central role in robotics research. In this talk, I will describe two
developments towards this aim. First I will describe how a robot can
autonomously extract information from Internet to decide where to find
an unknown object and how to learn on-line a recognition model.
In the second part of the talk I will describe how a non-expert user
can train a robot to perform simple tasks combining programming by
demonstration, reinforcement learning and user's feedback.