Publications

2018

• Préface
  
  • Rioul Olivier
  , 2018, pp.1-5.
• SemBib, un dépôt local d’articles scientifiques sémantiquement décrits
  
  • Moissinac Jean-Claude Jc
  , 2018. Le projet SemBib est une initiative au sein de Telecom ParisTech pour constituer et exploiter une base de connaissances sur nos publications scientifiques. Face à de grands entrepôts de références bibliographiques, nous considérons qu’une fédération de projets analogues à SemBib a du sens. Nous présentons ici les avancées actuelles du projet SemBib et ses relations avec d’autres projets.
• View-dependent compression of computer-generated hologram based on matching pursuit
  
  • El Rhammad Anas
  • Gioia Patrick
  • Gilles Antonin
  • Cagnazzo Marco
  • Pesquet-Popescu Beatrice
  , 2018.
• Home energy management system considering modular and deferrable electric loads under time-variable pricing scheme
  
  • Al Zahr Sawsan
  , 2018.
• Big Data, le défi de la formation
  
  • Clémençon Stéphan
  TEC Mobilité intelligente [Revue TEC : Transport Environnement Circulation], ATEC ITS France [Association pour le développement des Transports, de l’Environnement, et de la Circulation] [1973-....], 2018 (237).
• Antenna Aperture Impact on Channel Delay Spread in an Urban Outdoor Scenario at 17 and 60 GHz
  
  • Diakhate Cheikh
  • Conrat Jean-Marc
  • Cousin Jean-Christophe
  • Sibille Alain
  , 2018.
• Flexibility and dynamicity for open network-as-a-service: From VNF and architecture modeling to deployment
  
  • Boubendir Amina
  • Bertin Emmanuel
  • Simoni Noémie
  , 2018.
• Lightweight and Wide-Angle Metamaterial Absorbing Material Concept
  
  • Lepage A. C.
  • Pinto Yenny
  • Rance Olivier
  • Begaud Xavier
  • Capet Nicolas
  , 2018.
• Technical Report: Adding Missing Words to Regular Expressions
  
  • Rebele Thomas
  • Tzompanaki Katerina
  • Suchanek Fabian
  , 2018. Regular expressions (regexes) are patterns that are used in many applications to extract words or tokens from text. However, even hand-crafted regexes may fail to match all the intended words. In this paper, we propose a novel way to generalize a given regex so that it matches also a set of missing (previously non-matched) words. Our method finds an approximate match between the missing words and the regex, and adds disjunctions for the unmatched parts appropriately. We show that this method can not just improve the precision and recall of the regex, but also that it generates much shorter regexes than baselines and competitors on various datasets. This report complements our paper at the PAKDD 2018 conference. [18] Rapport technique: Ajout de mots manquants aux expressions régulières Résumé Les expressions régulières (regex) sont des modèles utilisés dans de nombreuses applications pour extraire des mots ou des parties du texte. Cependant, même les regex faites à la main ne correspondent pas toujours à l'ensemble des mots prévus. Dans cet article, nous proposons une nouvelle façon de généraliser une expression régulière donnée afin qu'elle corresponde également à un ensemble de mots manquants (précédemment non reconnus). Notre méthode trouve une correspondance approximative entre les mots manquants et l'expression regulière, et ajoute des disjonctions pour les parties non recon-nues de façon appropriée. Nous montrons que cette méthode améliore la précision et le rappel de la regex, et aussi qu'elle génère des expressions re-gulières beaucoup plus courtes que l'approche naïve et que les algorithmes concurrents sur différents jeux de données. Ce rapport complète notre article soumis à la conférence PAKDD 2018. [18]
• Narrowing the gap between QoS metrics and Web QoE using Above-the-fold metrics
  
  • da Hora Diego
  • Asrese Alemnew Sheferaw
  • Christophides Vassilis
  • Teixeira Renata
  • Rossi D.
  , 2018, pp.1-13. Page load time (PLT) is still the most common application Quality of Service (QoS) metric to estimate the Quality of Experience (QoE) of Web users. Yet, recent literature abounds with proposals for alternative metrics (e.g., Above The Fold, SpeedIndex and variants) that aim at better estimating user QoE. The main purpose of this work is thus to thoroughly investigate a mapping between established and recently proposed objective metrics and user QoE. We obtain ground truth QoE via user experiments where we collect and analyze 3,400 Web accesses annotated with QoS metrics and explicit user ratings in a scale of 1 to 5, which we make available to the community. In particular, we contrast domain expert models (such as ITU-T and IQX) fed with a single QoS metric, to models trained using our ground-truth dataset over multiple QoS metrics as features. Results of our experiments show that, albeit very simple, expert models have a comparable accuracy to machine learning approaches. Furthermore, the model accuracy improves considerably when building per-page QoE models, which may raise scalability concerns as we discuss.
• Connecting Width and Structure in Knowledge Compilation
  
  • Amarilli Antoine
  • Monet Mikaël
  • Senellart Pierre
  , 2018, 98, pp.1-17. Several query evaluation tasks can be done via knowledge compilation: the query result is compiled as a lineage circuit from which the answer can be determined. For such tasks, it is important to leverage some width parameters of the circuit, such as bounded treewidth or pathwidth, to convert the circuit to structured classes, e.g., deterministic structured NNFs (d-SDNNFs) or OBDDs. In this work, we show how to connect the width of circuits to the size of their structured representation, through upper and lower bounds. For the upper bound, we show how bounded-treewidth circuits can be converted to a d-SDNNF, in time linear in the circuit size. Our bound, unlike existing results, is constructive and only singly exponential in the treewidth. We show a related lower bound on monotone DNF or CNF formulas, assuming a constant bound on the arity (size of clauses) and degree (number of occurrences of each variable). Specifically, any d-SDNNF (resp., SDNNF) for such a DNF (resp., CNF) must be of exponential size in its treewidth; and the same holds for pathwidth when compiling to OBDDs. Our lower bounds, in contrast with most previous work, apply to any formula of this class, not just a well-chosen family. Hence, for our language of DNF and CNF, pathwidth and treewidth respectively characterize the efficiency of compiling to OBDDs and (d-)SDNNFs, that is, compilation is singly exponential in the width parameter. We conclude by applying our lower bound results to the task of query evaluation. (10.4230/LIPIcs.ICDT.2018.6)
  DOI : 10.4230/LIPIcs.ICDT.2018.6
• Enumeration on Trees under Relabelings
  
  • Amarilli Antoine
  • Bourhis Pierre
  • Mengel Stefan
  , 2018. (10.4230/LIPIcs.ICDT.2018.5)
  DOI : 10.4230/LIPIcs.ICDT.2018.5
• Modeling Heterogeneous Embedded Systems with TTool
  
  • Genius Daniela
  • Louërat Marie-Minerve
  • Pêcheux François
  • Apvrille Ludovic
  • Stratigopoulos Haralampos-G.
  , 2018. Embedded systems are increasingly heterogeneous , comprising digital and analog integrated circuits, sensors, and actuators. This paper presents a first step towards an integrated modeling and simulation tool for verification and virtual prototyping of heterogeneous embedded systems on different abstraction levels.
• Contributions to sparse source localization for MEG/EEG brain imaging
  
  • Bekhti Yousra
  , 2018. Understanding the full complexity of the brain has been a challenging research project for decades, yet there are many mysteries that remain unsolved. Being able to model how the brain represents, analyzes, processes, and transforms information of millions of different tasks in a record time is primordial for both cognitive and clinical studies. These tasks can go from language, perception, memory, attention, emotion, to reasoning and creativity. Magnetoencephalography (MEG) and Electroencephalography (EEG) allow us to non-invasively measure the brain activity with high temporal and good spatial resolution using sensors positioned all over the head, in order to be analyzed. For a given magnetic-electric field outside the head, there are an infinite number of electrical current source distributed inside of the brain that could have created it. This means that the M/EEG inverse problem is ill-posed, having many solutions to the single problem. This constrains us to make assumptions about how the brain might work. This thesis investigated the assumption of having sparse source estimate, i.e. only few sources are activated for each specific task. This is modeled as a penalized regression with a spatio-temporal regularization term. The aim of this thesis was to use outstanding methodologies from machine learning field to solve the three steps of the M/EEG inverse problem. The first step is to model the problem in the time frequency domain with a multi-scale dictionary to take into account the mixture of non-stationary brain sources, i.e. brain regions share information resulting in brain activity alternating from a source to another. This is done by formulating the problem as a penalized regression with a data fit term and a spatio-temporal regularization term, which has an extra hyperparameter. This hyperparameter is mostly tuned by hand, which makes the analysis of source brain activity not objective, but also hard to generalize on big studies. The second contribution is to automatically estimate this hyperparameter under some conditions, which increase the objectivity of the solvers. However, these state-of-the-art solvers have a main problem that their source localization solver gives one solution, and does not allow for any uncertainty quantification. We investigated this question by studying new techniques as done by a Bayesian community involving Markov Chain Monte Carlo (MCMC) methods. It allows us to obtain uncertainty maps over source localization estimation, which is primordial for a clinical study, e.g. epileptic activity. The last main contribution is to have a complete comparison of state-of-the-art solvers on phantom dataset. Phantom is an artificial object that mimics the brain activity based on theoretical description and produces realistic data corresponding to complex spatio-temporal current sources. In other words, all solvers have been tested on an almost real dataset with a known ground truth for a real validation.
• The MISO Free-Space Optical Channel at Low and Moderate SNR
  
  • Li Longguang
  • Moser Stefan M
  • Wang Ligong
  • Wigger Michèle
  , 2018, pp.1-6. The capacity of the multiple-input single-output (MISO) free-space optical channel with a per-antenna peak-power constraint and a sum (over all antennas) average-power constraint is studied. The asymptotic low-signal-to-noise-ratio (low-SNR) capacity is determined exactly and close upper and lower bounds are presented in the low-and moderate-SNR regimes. The asymptotic low-SNR limit is achieved by having each transmit antenna signal either with zero or with the maximally allowed peak power, and the latter only if all stronger antennas also send at maximum peak power. In particular, for almost all channel gains, the input to achieve the asymptotic low-SNR capacity is such that its projection on the channel-gain vector has only two or three positive probability point masses, one of them being at zero. The lower bounds at finite SNR are numerical and are obtained using input distributions whose projection on the channel-gain vector has either two, three, or four positive probability masses. Finally, the paper presents two analytic upper bounds on the capacity of the MISO channel: the first one closely follows the proposed numerical lower bounds in the low-SNR regime, and the second one can improve on previous bounds in the moderate-SNR regime. (10.1109/ciss.2018.8362301)
  DOI : 10.1109/ciss.2018.8362301
• On the two-filter approximations of marginal smoothing distributions in general state space models
  
  • Nguyen Thi Ngoc Minh
  • Le Corff Sylvain
  • Moulines Éric
  Advances in Applied Probability, Applied Probability Trust, 2018, 50 (1), pp.154-177. A prevalent problem in general state space models is the approximation of the smoothing distribution of a state conditional on the observations from the past, the present, and the future. The aim of this paper is to provide a rigorous analysis of such approximations of smoothed distributions provided by the two-filter algorithms. We extend the results available for the approximation of smoothing distributions to these two-filter approaches which combine a forward filter approximating the filtering distributions with a backward information filter approximating a quantity proportional to the posterior distribution of the state given future observations. (10.1017/apr.2018.8)
  DOI : 10.1017/apr.2018.8
• Narrow spectral linewidth in InAs/InP quantum dot distributed feedback lasers
  
  • Duan J.
  • Huang H.
  • Lu Z. G
  • Poole P. J
  • Wang C.
  • Grillot F.
  Applied Physics Letters, American Institute of Physics, 2018, 112 (12), pp.121102. (10.1063/1.5022480)
  DOI : 10.1063/1.5022480
• Availability enhancement and analysis for mixed-criticality systems on multi-core
  
  • Medina Roberto
  • Borde Etienne
  • Pautet Laurent
  , 2018, pp.1271-1276.
• Apprentissage de représentations pour l'analyse de scènes sonores
  
  • Bisot Victor
  , 2018. Ce travail de thèse s’intéresse au problème de l’analyse des sons environnementaux avec pour objectif d’extraire automatiquement de l’information sur le contexte dans lequel un son a été enregistré. Ce domaine de recherche a connu un succès grandissant ces dernières années entraînant une rapide évolution du nombre de travaux et des méthodes employées. Nos travaux explorent et contribuent à plusieurs grandes familles d’approches pour l’analyse de scènes et événements sonores allant de l’ingénierie de descripteurs jusqu’aux réseaux de neurones profonds. Notre travail se focalise sur les techniques d’apprentissage de représentations par factorisation en matrices positives (NMF), qui sont particulièrement adaptées à l’analyse d’environnements multi-sources tels que les scènes sonores. Nous commençons par montrer que les spectrogrammes contiennent suffisamment d’information pour discriminer les scènes sonores en proposant une combinaison de descripteurs d’images extraits à partir des images temps-fréquence. Nous quittons ensuite le monde de l’ingénierie de descripteurs pour aller vers un apprentissage automatique des représentations. Nous entamons cette partie du travail en nous intéressant aux approches non-supervisées, en particulier à l’apprentissage de descripteurs par différentes variantes de la NMF. Plusieurs des approches proposées confirment l’intérêt de l’apprentissage de caractéristiques par NMF en obtenant des performances supérieures aux meilleures approches par extraction de descripteurs. Nous proposons ensuite d’améliorer les représentations apprises en introduisant le modèle TNMF, une variante supervisée de la NMF. Les modèles et algorithmes TNMF proposés se basent sur un apprentissage conjoint du classifieur et du dictionnaire de sorte à minimiser un coût de classification. Dans une dernière partie, nous discutons des liens de compatibilité entre la NMF et certaines approches par réseaux de neurones profonds. Nous proposons et adaptons des architectures de réseaux de neurones à l’utilisation de la NMF. Les modèles introduits nous permettent d’atteindre des performances état de l’art sur des tâches de classification de scènes et de détection d’événements sonores. Enfin nous explorons la possibilité d’entraîner conjointement la NMF et les paramètres du réseau, regroupant ainsi les différentes étapes de nos systèmes en un seul problème d’optimisation.
• High Dynamic Range (HDR) image analysis
  
  • Rana Aakanksha
  , 2018. High Dynamic Range (HDR) imaging enables to capture a wider dynamic range and color gamut, thus enabling us to draw on subtle, yet discriminating details present both in the extremely dark and bright areas of a scene. Such property is of potential interest for computer vision algorithms where performance degrades substantially when the scenes are captured using traditional low dynamic range (LDR) imagery. While such algorithms have been exhaustively designed using traditional LDR images, little work has been done so far in contex of HDR content. In this thesis, we present the quantitative and qualitative analysis of HDR imagery for such task-specific algorithms. This thesis begins by identifying the most natural and important questions of using HDR content for low-level feature extraction task, which is of fundamental importance for many high-level applications such as stereo vision, localization, matching and retrieval. By conducting a performance evaluation study, we demonstrate how different HDR-based modalities enhance algorithms performance with respect to LDR on a proposed dataset. However, we observe that none of them can optimally to do so across all the scenes. To examine this sub-optimality, we investigate the importance of task-specific objectives for designing optimal modalities through an experimental study. Based on the insights, we attempt to surpass this sub-optimality by designing task-specific HDR tone-mapping operators (TMOs). In this thesis, we propose three learning based methodologies aimed at optimal mapping of HDR content to enhance the efficiency of local features extraction at each stage namely, detection, description and final matching.
• Experimental detection of steerability in Bell local states with two measurement settings
  
  • Orieux Adeline
  • Kaplan Marc
  • Venuti Vivien
  • Pramanik Tanumoy
  • Zaquine Isabelle
  • Diamanti Eleni
  Journal of Optics, Institute of Physics (IOP), 2018, 20 (4). Steering, a quantum property stronger than entanglement but weaker than non-locality in the quantum correlation hierarchy, is a key resource for one-sided device-independent quantum key distribution applications, in which only one of the communicating parties is trusted. A fine-grained steering inequality was introduced in [PRA 90 050305(R) (2014)], enabling for the first time the detection of steering in all steerable two-qubit Werner states using only two measurement settings. Here we numerically and experimentally investigate this inequality for generalized Werner states and successfully detect steerability in a wide range of two-photon polarization-entangled Bell local states generated by a parametric down-conversion source. (10.1088/2040-8986/aab031)
  DOI : 10.1088/2040-8986/aab031
• Procédé de classification et de localisation d'événements audiovisuels et appareil correspondant, produit-programme lisible par ordinateur et support d'informations lisible par ordinateur
  
  • Duong Quang-Khanh-Ngoc
  • Ozerov Alexey
  • Parekh Sanjeel
  • Essid Slim
  • Richard Gael
  • Pérez Patrick
  , 2018.
• A Reliable Method to Predict Parkinson’s Disease Stage and Progression based on Handwriting and Re-sampling Approaches
  
  • Taleb Catherine
  • Khachab Maha
  • Mokbel Chafic
  • Likforman-Sulem Laurence
  , 2018, pp.7-12. A reliable system depending on algorithms that assist in the decision-making process to diagnose Parkinson's disease (PD) at an early stage and to predict the Hoehn & Yahr (H&Y) stage and the unified Parkinson's disease rating scale (UPDRS) score is developed. In a previous work [3], we used features extracted from Arabic handwriting for diagnosing PD as binary decision. In this work, we use these features for constructing a prediction model that evaluates the H&Y stage and the UPDRS scores. A multi-class support vector machine (SVM) classifier is trained using re-sampling approaches such as adaptive synthetic sampling approach (ADASYN). The classifier is evaluated with 4-fold cross validation. The experiments show that H&Y stage, UPDRS scores, and total UPDRS can be predicted with accuracies of 94%, 92%, and 88% respectively. The proposed method can be implemented as an efficient clinical decision support system for early detection and monitoring the progression of PD. (10.1109/ASAR.2018.8480209)
  DOI : 10.1109/ASAR.2018.8480209
• Convergence and efficiency of adaptive importance sampling techniques with partial biasing
  
  • Fort Gersende
  • Jourdain Benjamin
  • Lelièvre Tony
  • Stoltz Gabriel
  Journal of Statistical Physics, Springer Verlag, 2018, 171 (2), pp.220–268. We consider a generalization of the discrete-time Self Healing Umbrella Sampling method, which is an adaptive importance technique useful to sample multimodal target distributions. The importance function is based on the weights of disjoint sets which form a partition of the space. In the context of computational statistical physics, the logarithm of these weights is, up to a multiplicative constant, the free energy, and the discrete valued function defining the partition is called the reaction coordinate. The algorithm is a generalization of the original Self Healing Umbrella Sampling method in two ways: (i) the updating strategy leads to a larger penalization strength of already visited sets and (ii) the target distribution is biased using only a fraction of the free energy, in order to increase the effective sample size and reduce the variance of importance sampling estimators. The algorithm can also be seen as a generalization of well-tempered metadynamics. We prove the convergence of the algorithm and analyze numerically its efficiency on a toy example. (10.1007/s10955-018-1992-2)
  DOI : 10.1007/s10955-018-1992-2
• A deep learning architecture for temporal sleep stage classification using multivariate and multimodal time series
  
  • Chambon Stanislas
  • Galtier Mathieu
  • Arnal Pierrick J
  • Wainrib Gilles
  • Gramfort Alexandre
  IEEE Transactions on Neural Systems and Rehabilitation Engineering, IEEE Institute of Electrical and Electronics Engineers, 2018, 26 (4), pp.17683810. Sleep stage classification constitutes an important preliminary exam in the diagnosis of sleep disorders. It is traditionally performed by a sleep expert who assigns to each 30 s of signal a sleep stage, based on the visual inspection of signals such as electroencephalograms (EEG), electrooculograms (EOG), electrocardiograms (ECG) and electromyograms (EMG). We introduce here the first deep learning approach for sleep stage classification that learns end-to-end without computing spectrograms or extracting hand-crafted features, that exploits all multivariate and multimodal Polysomnography (PSG) signals (EEG, EMG and EOG), and that can exploit the temporal context of each 30 s window of data. For each modality the first layer learns linear spatial filters that exploit the array of sensors to increase the signal-to-noise ratio, and the last layer feeds the learnt representation to a softmax classifier. Our model is compared to alternative automatic approaches based on convolutional networks or decisions trees. Results obtained on 61 publicly available PSG records with up to 20 EEG channels demonstrate that our network architecture yields state-of-the-art performance. Our study reveals a number of insights on the spatio-temporal distribution of the signal of interest: a good trade-off for optimal classification performance measured with balanced accuracy is to use 6 EEG with 2 EOG (left and right) and 3 EMG chin channels. Also exploiting one minute of data before and after each data segment offers the strongest improvement when a limited number of channels is available. As sleep experts, our system exploits the multivariate and multimodal nature of PSG signals in order to deliver state-of-the-art classification performance with a small computational cost. (10.1109/TNSRE.2018.2813138)
  DOI : 10.1109/TNSRE.2018.2813138