Sorry, you need to enable JavaScript to visit this website.
Partager

Publications

 

Les publications de nos enseignants-chercheurs sont sur la plateforme HAL :

 

Les publications des thèses des docteurs du LTCI sont sur la plateforme HAL :

 

Retrouver les publications figurant dans l'archive ouverte HAL par année :

2021

  • 10 Gbit/s free space data transmission at 9 µm wavelength with unipolar quantum optoelectronics
    • Dely Hamza
    • Bonazzi Thomas
    • Spitz Olivier
    • Rodriguez Etienne
    • Gacemi Djamal
    • Todorov Yanko
    • Pantzas Konstantinos
    • Beaudoin Grégoire
    • Sagnes Isabelle
    • Li Lianhe
    • Davies Alexander Giles
    • Linfield Edmund H
    • Grillot Frédéric
    • Vasanelli Angela
    • Sirtori Carlo
    , 2021. The realization of high-frequency unipolar quantum optoelectronic devices enables the demonstration of high bitrate free space data transmission in the second atmospheric window. Data-bits are written onto the laser emission using a large bandwidth amplitude modulator that operates by shifting the absorption of an optical transition in and out of the laser frequency.
  • Binding data mining and expert knowledge for one-day-ahead prediction of hourly global solar radiation
    • Campo-Ávila José Del
    • Takilalte Abdelatif
    • Bifet Albert
    • López Llanos Mora
    Expert Systems with Applications, Elsevier, 2021, 167, pp.114147. AbstractA new methodology to predict one-day-ahead hourly solar global radiation is proposed in this paper. This information is very useful to address many real problems; for instance, energy-market decision making is one of the contexts where that information is essential to ensure the correct integration of grid-connected photovoltaic solar systems. The developed methodology is based on the contribution of different experts to obtain improved data-driven models when included in the data mining process. The modelling phase, when models are induced and new patterns can be identified, is the one that most benefits from that expert knowledge. In this case, it is achieved by combining clustering, regression and classification methods that exploit meteorological data (directly measured or predicted by weather services). The developed models have been embedded in a prediction system that offers reliable forecasts on next-day hourly global solar radiation. As a result of the automatic learning process including the knowledge of different experts, 14 different types of day were identified based on the shape of hourly solar radiation throughout a day. The conventional definitions of types of days, that usually consider 4 options, are updated with this new proposal. The next-day prediction of hourly global radiation is obtained in two phases: in the first one, the next-day type is obtained from among the 14 possible types of day; in the second one, values of hourly global radiation are obtained using the centroid of the predicted type of day and extraterrestrial solar radiation. The relative root mean square error of the prediction model is less than 20%, meaning a significant reduction compared to previous models. Moreover, the proposed models can be recognized in the context of eXplainable Artificial Intelligence. (10.1016/J.ESWA.2020.114147)
    DOI : 10.1016/J.ESWA.2020.114147
  • Transfer learning of convolutional neural networks for texture synthesis and visual recognition in artistic images
    • Gonthier Nicolas
    , 2021. In this thesis, we study the transfer of Convolutional Neural Networks (CNN) trained on natural images to related tasks. We follow two axes: texture synthesis and visual recognition in artworks. The first one consists in synthesizing a new image given a reference sample. Most methods are based on enforcing the Gram matrices of ImageNet-trained CNN features. We develop a multi-resolution strategy to take into account large scale structures. This strategy can be coupled with long-range constraints either through a Fourier frequency constraint, or the use of feature maps autocorrelation. This scheme allows excellent high-resolution synthesis especially for regular textures. We compare our methods to alternatives ones with quantitative and perceptual evaluations. In a second axis, we focus on transfer learning of CNN for artistic image classification. CNNs can be used as off-the-shelf feature extractors or fine-tuned. We illustrate the advantage of the last solution. Second, we use feature visualization techniques, CNNs similarity indexes and quantitative metrics to highlight some characteristics of the fine-tuning process. Another possibility is to transfer a CNN trained for object detection. We propose a simple multiple instance method using off-the-shelf deep features and box proposals, for weakly supervised object detection. At training time, only image-level annotations are needed. We experimentally show the interest of our models on six non-photorealistic.
  • Transfer learning of convolutional neural networks for texture synthesis and visual recognition in artistic images
    • Gonthier Nicolas
    , 2021. In this thesis, we study the transfer of Convolutional Neural Networks (CNN) trained on natural images to related tasks. We follow two axes: texture synthesis and visual recognition in artworks. The first one consists in synthesizing a new image given a reference sample. Most methods are based on enforcing the Gram matrices of ImageNet-trained CNN features. We develop a multi-resolution strategy to take into account large scale structures. This strategy can be coupled with long-range constraints either through a Fourier frequency constraint, or the use of feature maps autocorrelation. This scheme allows excellent high-resolution synthesis especially for regular textures. We compare our methods to alternatives ones with quantitative and perceptual evaluations. In a second axis, we focus on transfer learning of CNN for artistic image classification. CNNs can be used as off-the-shelf feature extractors or fine-tuned. We illustrate the advantage of the last solution. Second, we use feature visualization techniques, CNNs similarity indexes and quantitative metrics to highlight some characteristics of the fine-tuning process. Another possibility is to transfer a CNN trained for object detection. We propose a simple multiple instance method using off-the-shelf deep features and box proposals, for weakly supervised object detection. At training time, only image-level annotations are needed. We experimentally show the interest of our models on six non-photorealistic. This file is the high definition version of the manuscript.
  • Amélioration de l'intelligibilité de signaux audio de parole en contexte bruité automobile
    • Gentet Enguerrand
    , 2021. La quantité de diffusion de signaux de parole dans les habitacles automobiles est de plus en plus importante : télécommunications, radio, système de navigation... Cependant, malgré les efforts et les avancées mécaniques, beaucoup de bruits persistent au sein de l'habitacle dégradant fortement l'intelligibilité de ces signaux de parole. L'objectif de cette thèse est alors de développer des outils de renforcement de la parole visant à traiter les signaux avant leur dégradation afin d'assurer une bonne intelligibilité dans le bruit des habitacles automobiles. Une approche de renforcement de la parole très performante consiste à utiliser un égaliseur fréquentiel afin d’optimiser un critère d’intelligibilité : le Speech Intelligibility Index (SII). Pour faciliter l'optimisation, les méthodes actuelles se basent sur des approximations du critère. De plus, en concentrant l'énergie spectrale du signal dans des zones où l'oreille est plus sensible, ces méthodes augmentent le volume perçu ce qui peut détériorer l'expérience utilisateur. Ainsi, en plus de proposer une méthode de résolution exacte du problème de maximisation du SII, nos travaux proposent d’introduire et étudier l'influence d'une nouvelle contrainte perceptive maintenant les signaux à leur niveau perçu. La popularisation des approches d’apprentissage automatique pousse à apprendre les traitements de renforcement de la parole à partir d’exemples naturellement produits dans le bruit (parole Lombard), ou en sur-articulant (parole claire). Les travaux actuels ne parviennent pas à obtenir des gains d’intelligibilité aussi significatifs qu’avec les modifications naturelles et nous pensons que la négligence de nombreux aspects temporels pourrait en être partiellement responsable. Nos travaux proposent donc d’approfondir ces approches en exploitant des modèles d’apprentissage et des pré-traitements adaptés aux séquences temporelles longues. Nous proposons aussi une nouvelle modélisation des modifications du débit de la parole directement intégrable dans l’apprentissage machine ce qui n'avait jamais été fait auparavant.
  • 3D Buildings Reconstruction with SAR Tomography Guided by Partial Footprints Information
    • Rambour Clément
    • Denis Loïc
    • Tupin Florence
    , 2021. SAR tomography is a powerful 3D reconstruction method. Nevertheless, in urban areas, results are usually rather sparse and some post-processing is needed to obtain 3D buildings. In this paper, external information about building footprints is introduced in the 3D reconstruction process. This is done by extending the recent method REDRESS based on a graphcut method (Rambour et al., Computer Vision and Image Understanding 2019). The graph construction is modified to take into account spatially variable weights depending on the footprints location. Experimental results show a clear improvement of the retrieved shapes.
  • Generalized Likelihood Ratio Tests for Linear Structure Detection in SAR Images
    • Gasnier Nicolas
    • Denis Loïc
    • Tupin Florence
    , 2021.
  • How to handle spatial correlations in SAR despeckling? Resampling strategies and deep learning approaches
    • Dalsasso Emanuele
    • Denis Loïc
    • Tupin Florence
    , 2021, pp.1-6. Speckle noise strongly affects Synthetic Aperture Radar (SAR) images, causing strong intensity fluctuations that make them difficult to analyze. Although many speckle reduction algorithms have been proposed, how to effectively deal with the spatial correlations of speckle remains an open question, especially in the most recent deep learning approaches. This paper tries to address this problem. Existing approaches to tackle the speckle correlations are described. Then, a standard training strategy for deep learning is proposed. Two models are trained and the increased robustness brought by including a Total Variation (TV) term in the loss function is analyzed on Sentinel-1 images.
  • Algebraic geometry codes and some applications
    • Couvreur Alain
    • Randriambololona Hugues
    , 2021, pp.998. This article surveys the development of the theory of algebraic geometry codes since their discovery in the late 70's. We summarize the major results on various problems such as: asymptotic parameters, improved estimates on the minimum distance, and decoding algorithms. In addition, we present various modern applications of these codes such as public-key cryptography, algebraic complexity theory, multiparty computation or distributed storage.
  • Linear codes from functions
    • Mesnager Sihem
    , 2021, pp.463-526. (10.1201/9781315147901-22)
    DOI : 10.1201/9781315147901-22
  • Approximate computing for embedded machine learning
    • Yang Xuecan
    , 2021. Convolutional Neural Networks (CNNs) have been extensively used in many fields such as image recognition, video processing, and naturallanguage processing. However, CNNs are still computational-intensive and resource-consuming. They are often constrained by the limit performanceand memory when deployed on embedded systems. This PhD research project aims at proposing CNNs which are more suitable for embedded systems withlow computing resources and memory requirements. Based on literature review, we propose three methods to accelerate the operation of neural networks : Selective Binarization, Quad-Approx Network and Min- ConvNets. Selective Binarization combines layers with different precisions in CNNs to achieve an acceptable speed and accuracy. As well an FPGA based hardware accelerator is proposed for these optimized structures. With the proposed signed PArameterized Clipping acTivation Function (signed PACT), the CNNs are quantized into 3 bits, and then a loss-less network is established by using approximate multiplier, which is named Quad-Approx Network. In addition to acceleration, what is more valuable is that Quad-Approx shows that CNNs are certain fault tolerance systems, which leads us to propose the MinConvNets. MinConvNet is a set of multiplication-less CNNs whose multiplications are replaced by approximate operations. MinConvNet can achieve negligible loss of prediction compared to exact image classification networks through transfer learning, meanwhile the multiplication which is more resource consuming to implement is replaced by easier implemented operations. Human is ushering the era of the artificial intelligence. In the meantime, the Internet of Things (IoT) makes our lives more convenient. These works bring more complex intelligent algorithms into the edge devices and helps us to create the era of Artificial intelligent Internet of Things (AIoT).
  • Ranked enumeration of MSO logic on words
    • Bourhis Pierre
    • Grez Alejandro
    • Jachiet Louis
    • Riveros Cristian
    , 2021, 186, pp.20:1--20:19. In the last years, enumeration algorithms with bounded delay have attracted a lot of attention for several data management tasks. Given a query and the data, the task is to preprocess the data and then enumerate all the answers to the query one by one and without repetitions. This enumeration scheme is typically useful when the solutions are treated on the fly or when we want to stop the enumeration once the pertinent solutions have been found. However, with the current schemes, there is no restriction on the order how the solutions are given and this order usually depends on the techniques used and not on the relevance for the user. In this paper we study the enumeration of monadic second order logic (MSO) over words when the solutions are ranked. We present a framework based on MSO cost functions that allows to express MSO formulae on words with a cost associated with each solution. We then demonstrate the generality of our framework which subsumes, for instance, document spanners and regular complex event processing queries and adds ranking to them. The main technical result of the paper is an algorithm for enumerating all the solutions of formulae in increasing order of cost efficiently, namely, with a linear preprocessing phase and logarithmic delay between solutions. The novelty of this algorithm is based on using functional data structures, in particular, by extending functional Brodal queues to suit with the ranked enumeration of MSO on words. (10.4230/LIPIcs.ICDT.2021.20)
    DOI : 10.4230/LIPIcs.ICDT.2021.20
  • Uniform Reliability of Self-Join-Free Conjunctive Queries
    • Amarilli Antoine
    • Kimelfeld Benny
    , 2021. The reliability of a Boolean Conjunctive Query (CQ) over a tuple-independent probabilistic database is the probability that the CQ is satisfied when the tuples of the database are sampled one by one, independently, with their associated probability. For queries without self-joins (repeated relation symbols), the data complexity of this problem is fully characterized in a known dichotomy: reliability can be computed in polynomial time for hierarchical queries, and is #P-hard for non-hierarchical queries. Hierarchical queries also characterize the tractability of queries for other tasks: having read-once lineage formulas, supporting insertion/deletion updates to the database in constant time, and having a tractable computation of tuples' Shapley and Banzhaf values. In this work, we investigate a fundamental counting problem for CQs without self-joins: how many sets of facts from the input database satisfy the query? This is equivalent to the uniform case of the query reliability problem, where the probability of every tuple is required to be 1/2. Of course, for hierarchical queries, uniform reliability is in polynomial time, like the reliability problem. However, it is an open question whether being hierarchical is necessary for the uniform reliability problem to be in polynomial time. In fact, the complexity of the problem has been unknown even for the simplest non-hierarchical CQs without self-joins. We solve this open question by showing that uniform reliability is #P-complete for every nonhierarchical CQ without self-joins. Hence, we establish that being hierarchical also characterizes the tractability of unweighted counting of the satisfying tuple subsets. We also consider the generalization to query reliability where all tuples of the same relation have the same probability, and give preliminary results on the complexity of this problem. (10.4230/LIPIcs.ICDT.2021.17)
    DOI : 10.4230/LIPIcs.ICDT.2021.17
  • MIMO links mediated by Reconfigurable Intelligent Surfaces
    • Sibille Alain
    , 2021, pp.1-5. (10.23919/EuCAP51087.2021.9411034)
    DOI : 10.23919/EuCAP51087.2021.9411034
  • Statistical Modeling of WBAN channels in Indoor Environments Based on Measurements and Ray Tracing
    • Youssef Badre
    • Roblin Christophe
    , 2021, pp.1-5. (10.23919/EuCAP51087.2021.9411002)
    DOI : 10.23919/EuCAP51087.2021.9411002
  • Liveness analysis techniques and run-time environment for memory management of dataflow applications
    • Dauphin Benjamin
    , 2021. This thesis has been realized at Télécom Paris and it has been financed by Nokia Bell Labs France. It studies different techniques to handle the issue of deadlocks and memory shortages in computing systems. Its work is motivated by the rise over the past decades of heterogeneous and Non-Uniform Memory Access (NUMA) architectures in all varieties computing systems, from embedded systems running on Multi-Processor Systems on a Chip (MPSoCs) to distributed High-Performance Computing (HPC) systems. We focus more specifically on the issue of memory shortages in embedded systems used for Digital Signal Processing, but our contributions could be applied to different applications and platforms.The contributions of this thesis are threefold:(1) we present a deadlock prevention technique based on the analysis of cliques in Memory Exclusion Graphs, which are graphs representing buffers allocated in memory and whether they might get simultaneously allocated;(2) we present an optimization on the conventional liveness analysis for memory shortages, allowing to execute the liveness analysis in reasonable time for larger systems than previously supported;(3) we developed a deadlock avoidance strategy using results from the liveness analysis, and integrated it into an experimental run-time environment.We evaluate our first and second contributions in comparison to an existing state-of-the-art tool.Finally we propose multiple leads to improve on the contributions of the thesis.
  • A primer on Shannon's information theory
    • Rioul Olivier
    , 2021.
  • Intensity noise and modulation dynamics of epitaxial quantum dot semiconductor lasers on silicon
    • Grillot Frédéric
    , 2021.
  • Temperature tolerance of a hybrid III-V/Si distributed feedback semiconductor laser with a large quality factor
    • Gomez Sandra
    • Huang Heming
    • Grillot Frédéric
    , 2021, pp.21. (10.1117/12.2578663)
    DOI : 10.1117/12.2578663
  • Towards a turnkey private communication system using a quantum cascade laser emitting at 4 microns
    • Spitz Olivier
    • Durupt Lauréline
    • Grillot Frédéric
    , 2021, pp.18. (10.1117/12.2578238)
    DOI : 10.1117/12.2578238
  • Control plane in dynamic software networks
    • Wion Adrien
    , 2021. During the last years, network infrastructure has moved from dedicated-hardware solutions implementing fixed functions to more flexible software based ones. On one hand, SDN (Software Defined Network) can flexibly control forwarding operations, while on the other, NFV (Network Function Virtualization) creates elastic functions that can scale with the user demands. So far, these solutions have been used to simplify network management and operations, but they let envision a network that can automatically react to network events. In this thesis, we explore to what extent these new software networks can be used to react and adapt finely to the network dynamics.Our first contribution focuses on service chaining: the ability to steer flows through a set of waypoints hosting functions before they reach their destinations. We show that a distributed control plane that relies on existing routing protocols and is constituted by autonomous nodes can dynamically steer traffic through chains of services. Our solution finely adapts its decision to the network traffic and automatically balances the induced load on the functions present in the network. Moreover, our proposal, contrary to existing solutions, can be incrementally deployed in today's network.In our second contribution, we compare two types of chaining decisions: a centralized one with an end-to-end view of the chain and a distributed approach that solely routes flow from a function to another. We show that the two decisions are close in realistic topologies. Thus, hop-by-hop chaining could be used without affecting chaining performance. Finally, we explore how software networks can react to network dynamics in datacenters. So far, load balancers use static policies to spread incoming traffic on servers, which leads to imbalance and overprovisioning. We propose to close the loop and dynamically adapt the policy to the server load variation. Our MPC (Model Predictive Control) approach proved to be efficient to reduce load imbalance at a slow pace, thus improving the number of requests a cluster can process.
  • Analysis of stepwise charging limits and its implementation for efficiency improvement in switched capacitor DC–DC converters
    • Veirano Francisco
    • Lisboa Pablo Castro
    • Pérez-Nicoli Pablo
    • Naviner Lirida
    • Silveira Fernando
    Analog Integrated Circuits and Signal Processing, Springer Verlag, 2021, 109 (2), pp.271-282. In this work we analysed the stepwise charging technique to find the limits from which it is beneficial in terms of load capacitance and charge–discharge frequency. We included in the analysis practical limitations such as the consumption of auxiliary logic needed to implement the technique and the minimum size of auxiliary switches imposed by the technology. We proposed an ultra-low-power logic block to push these limits and to obtain benefits from this technique in small capacitances. Finally, we proposed to use a stepwise driver in the driving of the gate capacitance of power switches in switched-capacitor (SC) DC–DC converters. We designed and manufactured, in a 130 nm process, a SC DC–DC converter and measured a 29% energy reduction in the gate-drive losses of the converter. This accounts for an improvement of 4% (from 69 to 73%) in the overall converter efficiency. (10.1007/s10470-021-01810-5)
    DOI : 10.1007/s10470-021-01810-5
  • Data stream analysis: Foundations, major tasks and tools
    • Bahri Maroua
    • Bifet Albert
    • Gama João
    • Gomes Heitor Murilo
    • Maniu Silviu
    Wiley Interdisciplinary Reviews: Data Mining and Knowledge Discovery, Wiley, 2021, 11 (3), pp.e1405. The significant growth of interconnected Internet-of-Things (IoT) devices, the use of social networks, along with the evolution of technology in different domains, lead to a rise in the volume of data generated continuously from multiple systems. Valuable information can be derived from these evolving data streams by applying machine learning. In practice, several critical issues emerge when extracting useful knowledge from these potentially infinite data, mainly because of their evolving nature and high arrival rate which implies an inability to store them entirely. In this work, we provide a comprehensive survey that discusses the research constraints and the current state-of-the-art in this vibrant framework. Moreover, we present an updated overview of the latest contributions proposed in different stream mining tasks, particularly classification, regression, clustering, and frequent patterns. (10.1002/widm.1405)
    DOI : 10.1002/widm.1405
  • Side channel attacks for architecture extraction of neural networks
    • Chabanne Hervé
    • Danger Jean‐luc
    • Guiga Linda
    • Kühne Ulrich
    CAAI Transactions on Intelligent Technologies, Institution of Engineering and Technology, 2021, 6 (1), pp.3-16. (10.1049/cit2.12026)
    DOI : 10.1049/cit2.12026
  • Ultra-Low Power system for atrioventricular synchronization using leadless pacemakers
    • Maldari Mirko
    • Jabbour Chadi
    • Haddab Youcef
    • Desgreys Patricia
    Bulletin of the International Union of Radio Science, 2021 (376).