Publications

2023

• FALL: A Modular Adaptive Learning Platform for Streaming Data
  
  • Halstead Ben
  • Koh Yun Sing
  • Riddle Patricia
  • Pechenizkiy Mykola
  • Bifet Albert
  , 2023, pp.3619--3622. A growing number of tasks require adaptive machine learning systems capable of learning continuously from incoming data and adapting to changes in their environment. In order to enable the widespread adoption of machine learning for streaming data, it is crucial that practitioners and researchers have the tools to efficiently build and evaluate adaptive learning systems. In this paper we demonstrate FALL, a Framework for Adaptive Life-long Learning, which we have developed to enable the full adaptive learning pipeline to be built using modular, reusable components, enabling users to easily and efficiently develop, implement, and evaluate state-of-the-art adaptive learning systems. Source code, documentation, and examples may be found at https://benhalstead.dev/FALL/. (10.1109/ICDE55515.2023.00282)
  DOI : 10.1109/ICDE55515.2023.00282
• RF Energy Harvesting and Wireless Power Transfer for Energy Autonomous Wireless Devices and RFIDs
  
  • Niotaki Kyriaki
  • Carvalho Nuno Borges
  • Georgiadis Apostolos
  • Gu Xiaoqiang
  • Hemour Simon
  • Wu Ke
  • Matos Diogo
  • Belo Daniel
  • Pereira Ricardo
  • Figueiredo Ricardo
  • Chaves Henrique
  • Mendes Bernardo
  • Correia Ricardo
  • Oliveira Arnaldo
  • Palazzi Valentina
  • Alimenti Federico
  • Mezzanotte Paolo
  • Roselli Luca
  • Benassi Francesca
  • Costanzo Alessandra
  • Masotti Diego
  • Paolini Giacomo
  • Eid Aline
  • Hester Jimmy
  • Tentzeris Manos M.
  • Shinohara Naoki
  IEEE Journal of Microwaves, IEEE, 2023, 3 (2), pp.63 - 782. Radio frequency (RF) energy harvesting and wireless power transmission (WPT) technologies —both near-field and far-field—have attracted significant interest for wireless applications and RFID systems. We already utilize near-field WPT products in our life and it is expected that RF EH and far-field WPT systems can drive the future low-power wireless systems. In this article, we initially present a brief historical overview of these technologies. The main technical challenges of rectennas and WPT transmitters are discussed. Furthermore, this paper presents the recent advances on the development of these technologies, including the possibility of powering RFID systems through the millimeter wave power from 5G networks, the trends in flexible rectennas design and the technological developments on the simultaneous wireless information and power transfer (SWIPT) (10.1109/JMW.2023.3255581)
  DOI : 10.1109/JMW.2023.3255581
• Comparing Two Samples Through Stochastic Dominance: A Graphical Approach
  
  • Arza Etor
  • Ceberio Josu
  • Irurozki Ekhiñe
  • Pérez Aritz
  Journal of Computational and Graphical Statistics, Taylor & Francis, 2023, 32 (2), pp.551-566. Nondeterministic measurements are common in real-world scenarios: the performance of a stochastic optimization algorithm or the total reward of a reinforcement learning agent in a chaotic environment are just two examples in which unpredictable outcomes are common. These measures can be modeled as random variables and compared among each other via their expected values or more sophisticated tools such as null hypothesis statistical tests. In this article, we propose an alternative framework to visually compare two samples according to their estimated cumulative distribution functions. First, we introduce a dominance measure for two random variables that quantifies the proportion in which the cumulative distribution function of one of the random variables stochastically dominates the other one. Then, we present a graphical method that decomposes in quantiles (i) the proposed dominance measure and (ii) the probability that one of the random variables takes lower values than the other. With illustrative purposes, we reevaluate the experimentation of an already published work with the proposed methodology and we show that additional conclusions—missed by the rest of the methods—can be inferred. Additionally, the software package RVCompare was created as a convenient way of applying and experimenting with the proposed framework. (10.1080/10618600.2022.2084405)
  DOI : 10.1080/10618600.2022.2084405
• Multimodal Hierarchical Attention Neural Network: Looking for Candidates Behaviour Which Impact Recruiter's Decision
  
  • Hemamou Léo
  • Guillon Arthur
  • Martin Jean-Claude
  • Clavel Chloé
  IEEE Transactions on Affective Computing, Institute of Electrical and Electronics Engineers, 2023, 14 (2), pp.969-985. Automatic analysis of job interviews has gained in interest amongst academic and industrial research. The particular case of asynchronous video interviews allows to collect vast corpora of videos where candidates answer standardized questions in monologue videos, enabling the use of deep learning algorithms. On the other hand, state-of-the-art approaches still face some obstacles, among which the fusion of information from multiple modalities and the interpretability of the predictions. We study the task of predicting candidates performance in asynchronous video interviews using three modalities (verbal content, prosody and facial expressions) independently or simultaneously, using data from real interviews which take place in real conditions. We propose a sequential and multimodal deep neural network model, called Multimodal HireNet. We compare this model to state-of-the-art approaches and show a clear improvement of the performance. Moreover, the architecture we propose is based on attention mechanism, which provides interpretability about which questions, moments and modalities contribute the most to the output of the network. While other deep learning systems use attention mechanisms to offer a visualization of moments with attention values, the proposed methodology enables an in-depth interpretation of the predictions by an overall analysis of the features of social signals contained in these moments (10.1109/TAFFC.2021.3113159)
  DOI : 10.1109/TAFFC.2021.3113159
• Detecting DDoS attacks using adversarial neural network
  
  • Mustapha Ali
  • Khatoun Rida
  • Zeadally Sherali
  • Chbib Fadlallah
  • Fadlallah Ahmad
  • Fahs Walid
  • El Attar Ali
  Computers & Security, Elsevier, 2023, 127, pp.103117. In a Distributed Denial of Service (DDoS) attack, a network of compromised devices is used to overwhelm a target with a flood of requests, making it unable to serve legitimate requests. The detection of these attacks is a challenging issue in cybersecurity, which has been addressed using Machine Learning (ML) and Deep Learning (DL) algorithms. Although ML/DL can improve the detection accuracy, but they can still be evaded - ironically - through the use of ML/DL techniques in the generation of the attack traffic. In particular, Generative Adversarial Networks (GAN) have proven their efficiency in mimicking legitimate data. We address the above aspects of ML/DL-based DDoS detection and anti-detection techniques. First, we propose a DDoS detection method based on the Long Short-Term Memory (LSTM) model, which is a type of Recurrent Neural Networks (RNNs) capable of learning long-term dependencies. The detection scheme yields a high accuracy level in detecting DDoS attacks. Second, we tested the same technique against different types of adversarial DDoS attacks generated using GAN. The results show the inefficiency of the LSTM-based detection scheme. Finally, we demonstrate how to enhance this scheme to detect adversarial DDoS attacks. Our experimental results show that our detection model is efficient and accurate in identifying GAN-generated adversarial DDoS traffic with a detection ratio ranging between 91.75% and 100%. (10.1016/j.cose.2023.103117)
  DOI : 10.1016/j.cose.2023.103117
• Les matrices de Passage
  
  • Zayana Karim
  Quadrature, EDP Sciences, 2023. Les matrices de Passage ou les messages de Patrick ? Au même titre que leur nom, les matrices de passage sont source de confusion. Les raisons tiennent à des conventions de langage pas toujours très heureuses autant qu'à des logiques pouvant friser le paradoxe. Voici donc quelque moyens pour s'y retrouver.
• Statistical Characterization and Modeling of Indoor RF-EMF Down-Link Exposure
  
  • Mulugeta Biruk Ashenafi
  • Wang Shanshan
  • Ben Chikha Wassim
  • Liu Jiang
  • Roblin Christophe
  • Wiart Joe
  Sensors, MDPI, 2023, 23 (7), pp.3583. With the increasing use of wireless communication systems, assessment of exposure to radio-frequency electromagnetic field (RF-EMF) has now become very important due to the rise of public risk perception. Since people spend more than 70% of their daily time in indoor environments, including home, office, and car, the efforts devoted to indoor RF-EMF exposure assessment has also increased. However, assessment of indoor exposure to RF-EMF using a deterministic approach is challenging and time consuming task as it is affected by uncertainties due to the complexity of the indoor environment and furniture structure, existence of multiple reflection, refraction, diffraction and scattering, temporal variability of exposure, and existence of many obstructions with unknown dielectric properties. Moreover, it is also affected by the existence of uncontrolled factors that can influence the indoor RF-EMF exposure such as the constant movement of people and random movement of furniture and doors as people are working in the building. In this study, a statistical approach is utilized to characterize and model the total indoor RF-EMF down-link (DL) exposure from all cellular bands on each floor over the length of a wing since the significance of distance is very low between any two points on each floor in a wing and the variation of RF-EMF DL exposure is mainly influenced by the local indoor environment. Measurements were conducted in three buildings that are located within a few hundred meters vicinity of two base station sites supporting several cellular technologies (2G, 3G, 4G, and 5G). We apply the one-sample Kolmogorov–Smirnov test on the measurement data, and we prove that the indoor RF-EMF DL exposure on each floor over the length of a wing is a random process governed by a Gaussian distribution. We validate this proposition using leave-one-out cross validation technique. Consequently, we conclude that the indoor RF-EMF DL exposure on each floor over the length of a wing can be modeled by a Gaussian distribution and, therefore, can be characterized by the mean and the standard deviation parameters. (10.3390/s23073583)
  DOI : 10.3390/s23073583
• Uniform Reliability for Unbounded Homomorphism-Closed Graph Queries
  
  • Amarilli Antoine
  , 2023, 26th International Conference on Database Theory (ICDT 2023). We study the uniform query reliability problem, which asks, for a fixed Boolean query Q, given an instance I, how many subinstances of I satisfy Q. Equivalently, this is a restricted case of Boolean query evaluation on tuple-independent probabilistic databases where all facts must have probability 1/2. We focus on graph signatures, and on queries closed under homomorphisms. We show that for any such query that is unbounded, i.e., not equivalent to a union of conjunctive queries, the uniform reliability problem is #P-hard. This recaptures the hardness, e.g., of s-t connectedness, which counts how many subgraphs of an input graph have a path between a source and a sink. This new hardness result on uniform reliability strengthens our earlier hardness result on probabilistic query evaluation for unbounded homomorphism-closed queries [Amarilli and Ceylan, 2021]. Indeed, our earlier proof crucially used facts with probability 1, so it did not apply to the unweighted case. The new proof presented in this paper avoids this; it uses our recent hardness result on uniform reliability for non-hierarchical conjunctive queries without self-joins [Antoine Amarilli and Benny Kimelfeld, 2022], along with new techniques. (10.4230/LIPIcs.ICDT.2023.14)
  DOI : 10.4230/LIPIcs.ICDT.2023.14
• Aging and rejuvenating strategies for fading windows in multi-label classification on data streams
  
  • Roseberry Martha
  • Dzeroski Saso
  • Bifet Albert
  • Cano Alberto
  , 2023, pp.390--397. Combining the challenges of streaming data and multi-label learning, the task of mining a drifting, multi-label data stream requires methods that can accurately predict labelsets, adapt to various types of concept drift and run fast enough to process each data point before the next arrives. To achieve greater accuracy, many multi-label algorithms use computationally expensive techniques, such as multiple adaptive windows, with little concern for runtime and memory complexity. We present Aging and Rejuvenating kNN (ARkNN) which uses simple resources and efficient strategies to weight instances based on age, predictive performance, and similarity to the incoming data. We break down ARkNN into its component strategies to show the impact of each and experimentally compare ARkNN to seven state-of-the-art methods for learning from multi-label data streams. We demonstrate that it is possible to achieve competitive performance in multi-label classification on streams without sacrificing runtime and memory use, and without using complex and computationally expensive dual memory strategies. (10.1145/3555776.3577625)
  DOI : 10.1145/3555776.3577625
• GEO Feeder uplinks: tip-tilt-focus estimation at PAA aided by on-axis phase and amplitude sensing and LGS off-axis high-order measurements
  
  • Lognoné Perrine
  • Conan Jean-Marc
  • Rekaya Ben Othman Ghaya
  • Paillier Laurie
  • Bonnefois Aurélie Montmerle
  • Vedrenne Nicolas
  , 2023. Reaching very high data rates in GEO Feeder optical uplinks is greatly impaired by the fading nature of the atmospheric channel. One of the current solutions to reduce the atmospheric turbulence impact on the signal, and thus improve the statistics of the received power aboard the satellite, is to apply a pre-compensation by adaptive optics (AO). In such a configuration, the phase is measured on the downlink beam and applied to the uplink. However, due to the point-ahead anisoplanatism inherent to the geometry of the GEO Feeder, uplink correction efficiency is limited. As a result, the received power aboard the satellite still undergoes long and deep fades and the information signal is impaired. We have recently proposed a new method to estimate the phase at PAA using the phase and log-amplitude sensed on the downlink beam through a Minimum Mean Square Error (MMSE) approach that exploits a priori statistical information [1]. This approach allows reducing the estimation phase error variance down to 40% of the anisoplanatic error variance. An alternative solution is to use a laser guide star (LGS) to directly measure the phase at PAA [2]. However, this method does not currently allow to retrieve the Tip, Tilt and Focus modes which are however crucial modes to be corrected in order to improve the coupling statistics. We propose here to take advantage of both techniques and to combine them in order to estimate the Tip, Tilt and Focus at PAA by incorporating the LGS high order measurements in the MMSE estimation. We develop the associated theoretical reconstructor, and evaluate the performance of the phase estimation as well as the gain on the coupled flux statistics aboard the GEO-satellite. The new estimator is shown to reduce the Tip, Tilt and Focus error variance up to 70% of their initial value. Finally, we discuss the impact of an imperfect measurement of the high order modes at PAA, which error is induced by a lack of stabilisation of LGS position in the sky.
• Grote: Group Testing for Privacy-Preserving Face Identification
  
  • Ibarrondo Alberto
  • Chabanne Hervé
  • Despiegel Vincent
  • Önen Melek
  , 2023. (10.1145/3577923.3583656)
  DOI : 10.1145/3577923.3583656
• Joint scheduling-offloading policies in NOMA-based mobile edge computing systems
  
  • Djemai Ibrahim
  • Sarkiss Mireille
  • Ciblat Philippe
  , 2023, pp.1-6. We consider a Non Orthogonal Multiple Access (NOMA)-based wireless network where User Equipments (UEs) are connected to a Base Station (BS) equipped with a Mobile Edge Computing (MEC) server. The UEs can process their buffered data packets with strict delay either locally or by offloading them to the base station's MEC server. In order to minimize the dropped packets due to buffer overflow or delay violation, the scheduling-offloading problem is formulated as a Markov Decision Process (MDP) and solved using various optimal and Reinforcement Learning (RL) algorithms. The output of each policy is, for each user, the number of packets to be processed and the type of processing (locally or remotely). The decisions rely on the channel state information and the buffers states. The numerical results show the great advantage of using NOMA compared to Orthogonal Multiple Access (OMA). We further analyze the scalability capabilities of the used algorithms, which validates the benefits of using Deep Reinforcement Learning (DRL) techniques. (10.1109/WCNC55385.2023.10119046)
  DOI : 10.1109/WCNC55385.2023.10119046
• Statistical Modelling of the Delay Spread of the WBAN channel Considering Room Geometry and Material Characteristics
  
  • Youssef Badre
  • Roblin Christophe
  , 2023, pp.1-5. This communication presents the development of Delay Spread statistical models of the Wireless Body Area Network indoor channel taking into account parametrically the geometry and the material characteristics of the rooms within the framework of a scenario-based approach in the UWB (Ultra Wide Band) context. The study is performed in the 1 st UWB subband B = [3.1, 4.8] GHz and restricted to the stationary (timeinvariant) channel, the subject being motionless. The models are derived from statistical samples obtained thanks to a homemade simplified Ray-Tracing code. The trends of the simulations were compared to the behavior of the Delay Spread of the "classic" indoor channel observed in several experimental studies (for different frequency bands), which confirmed the reliability and consistency of our approach. The models extracted from a set made up of several categories of premises present a very satisfactory goodness of fit for the three considered radio links. (10.23919/EuCAP57121.2023.10133637)
  DOI : 10.23919/EuCAP57121.2023.10133637
• Super-resolution in brain positron emission tomography using a real-time motion capture system
  
  • Chemli Yanis
  • Tétrault Marc-André
  • Marin Thibault
  • Normandin Marc D
  • Bloch Isabelle
  • El Fakhri Georges
  • Ouyang Jinsong
  • Petibon Yoann
  NeuroImage, Elsevier, 2023, 272, pp.120056-01:120056-12. Super-resolution (SR) is a methodology that seeks to improve image resolution by exploiting the increased spatial sampling information obtained from multiple acquisitions of the same target with accurately known subresolution shifts. This work aims to develop and evaluate an SR estimation framework for brain positron emission tomography (PET), taking advantage of a high-resolution infra-red tracking camera to measure shifts precisely and continuously. Moving phantoms and non-human primate (NHP) experiments were performed on a GE Discovery MI PET/CT scanner (GE Healthcare) using an NDI Polaris Vega (Northern Digital Inc), an external optical motion tracking device. To enable SR, a robust temporal and spatial calibration of the two devices was developed as well as a list-mode Ordered Subset Expectation Maximization PET reconstruction algorithm, incorporating the high-resolution tracking data from the Polaris Vega to correct motion for measured line of responses on an event-by-event basis. For both phantoms and NHP studies, the SR reconstruction method yielded PET images with visibly increased spatial resolution compared to standard static acquisitions, allowing improved visualization of small structures. Quantitative analysis in terms of SSIM, CNR and line profiles were conducted and validated our observations. The results demonstrate that SR can be achieved in brain PET by measuring target motion in real-time using a high-resolution infrared tracking camera. (10.1016/j.neuroimage.2023.120056)
  DOI : 10.1016/j.neuroimage.2023.120056
• Intrusion detection with deep learning for in-vehicle networks
  
  • Khatib Natasha Al-
  , 2023. In-vehicle communication which refers to the communication and exchange of data between embedded automotive devices plays a crucial role in the development of intelligent transportation systems (ITS), which aim to improve the efficiency, safety, and sustainability of transportation systems. The proliferation of embedded sensor-centric communication and computing devices connected to the in-vehicle network (IVN) has enabled the development of safety and convenience features including vehicle monitoring, physical wiring reduction, and improved driving experience. However, with the increasing complexity and connectivity of modern vehicles, the expanding threat landscape of the IVN is raising concerns. A range of potential security risks can compromise the safety and functionality of a vehicle putting the life of drivers and passengers in danger. Numerous approaches have thus been proposed and implemented to alleviate this issue including firewalls, encryption, and secure authentication and access controls. As traditional mechanisms fail to fully counterattack intrusion attempts, the need for a complementary defensive countermeasure is necessary. Intrusion Detection Systems (IDS) have been thus considered a fundamental component of every network security infrastructure, including IVN. Intrusion detection can be particularly useful in detecting threats that may not be caught by other security measures, such as zero-day vulnerabilities or insider attacks. It can also provide an early warning of a potential attack, allowing car manufacturers to take preventive measures before significant damage occurs. The main objective of this thesis is to investigate the capability of deep learning techniques in detecting in-vehicle intrusions. Deep learning algorithms have the ability to process large amounts of data and recognize complex patterns that may be difficult for humans to discern, making them well-suited for detecting intrusions in IVN. However, since the E/E architecture of a vehicle is constantly evolving as new technologies and requirements emerge, we propose different deep learning-based solutions for different E/E architectures and for various tasks including anomaly detection and classification.
• Real-time and efficient control for autonomous racing
  
  • Li Nan
  , 2023. Recently, the autonomous driving domain has made tremendous advancements. By investigating the challenge of autonomous racing, a special form of autonomous driving, we seek to better understand how vehicles could be efficiently controlled in real-time settings for handling intricate dynamic situations. We develop an approach based on Nonlinear Model Predictive Control (NMPC), a cutting-edge control technique, that can attain the optimal progress time of the vehicle while accounting for nonlinear system dynamics and obstacle-related time-varying constraints. To deal with the presence of an opponent vehicle, we combine NMPC with Mixed Integer Programming (MIP) for encoding safe and efficient overtaking maneuvers. However, it is challenging to implement NMPC on embedded devices due to its high calculation complexity. One concern is ensuring real-time execution of the controller, which necessitates strict adherence to the time budget restriction and rigorous compliance with deadlines. Another problem is managing to make the control efficient, which calls for the maintenance of an adequate level of system performance. We propose a multi-step recomputation approach for the single-vehicle race mode, which is triggered based on specific events. One of the triggering conditions aims at ensuring that the real-time budget constraints are respected. The other triggering condition serves for reducing computational time while retaining quasi-optimal lap time performance. For head-to-head racing mode, we propose an algorithm as an online feasible alternative to MIP encoding. It efficiently aggregates overtaking decisions and schedules them at a deterministic control frequency to meet real-time requirements. In a generic system architecture, we also take into account other software components besides the controller, such as opponent detection and self-localization algorithms, which collectively constitute a Directed Acyclic Graph (DAG). To assign DAG components to available processors with varying degrees of parallelism, we propose a task execution model which decreases the latency, increases the control update rate, and eventually enhances the system performance. In summary, this thesis provides a set of mechanisms aimed at an efficient implementation of real-time control in autonomous systems.
• Fast and accurate estimation of correctness rate in combinatorial circuits based on clustering
  
  • Goudet Esther
  • Naviner Lirida
  • Daveau Jean Marc
  • Roche Philippe
  , 2023, pp.1-6.
• Contextes de proportionnalités algébriques et géométriques en lien avec les programmes collège et lycée
  
  • Zayana Karim
  , 2023. À l'invitation du séminaire de l'IREM, exposé sur quelques contextes donnant lieu (ou pas) à proportionnalité. Attention, à 24'36, il faut entendre "onde transversale" et non "onde longitudinale" : https://video.irem.univ-paris-diderot.fr/w/fd2b4e97-95d6-48d7-a5d0-748ce7a3cfd7
• Fast analysis of combinatorial netlists correctness rate based on binomial law and partitioning
  
  • Goudet Esther
  • Treviño Luis Peña
  • Naviner Lirida
  • Daveau Jean-Marc
  • Roche Philippe
  , 2023, pp.1-6. This paper targets fault propagation evaluation in combinational circuits and focuses on netlist-based logic masking prediction. We propose an approach that combines clustering and analytical estimation methods to allow to control the trade-off between accuracy and computation time. The experimental results performed on netlists with large number of gates and complex reconvergent structures show that the proposed solution is better adapted to large combinatorial netlists than the analysis found in the literature. (10.1109/lats58125.2023.10154491)
  DOI : 10.1109/lats58125.2023.10154491
• Monte Carlo Methods and Stochastic Approximation : Theory and Applications to Machine Learning
  
  • Leluc Rémi
  , 2023. Across a breadth of research areas, whether in Bayesian inference, reinforcement learning or variational inference, the need for accurate and efficient computation of integrals and parameters minimizing risk functions arises, making stochastic optimization and Monte Carlo methods one of the fundamental problems of statistical and machine learning research. This thesis focuses on Monte Carlo integration and stochastic optimization methods, both from a theoretical and practical perspectives, where the core idea is to use randomness to solve deterministic numerical problems. From a technical standpoint, the study is mainly based on two standard concepts: variance reduction and adaptive sampling techniques. The first part of the thesis focuses on various control variates techniques for Monte Carlo integration. The study is based on mathematical tools coming from probability theory and statistics aiming to understand the behavior of certain existing algorithms and to design new ones with thorough analysis of the integration error. First, we present a LASSO-type procedure to allow the use of high-dimensional control variates. Then, a weighted least-squares estimate, called AISCV, is proposed to incorporate control variates within the adaptive importance sampling framework. Finally, a Monte Carlo method with control variates based on nearest neighbors estimates, called Control Neighbors, is provided. The second part of the thesis deals with stochastic optimization algorithms. First, we investigate a general class of stochastic gradient descent (SGD) algorithms, called conditioned SGD, based on a preconditioning of the gradient direction. Then we present a general framework to perform coordinate sampling for SGD algorithms. While classical forms of SGD algorithms treat the different coordinates in the same way, a framework allowing for adaptive (non uniform) coordinate sampling is developed to leverage structure in data. To emphazise the practical applications of the proposed methods, all algorithms are implemented and tested against state-of-the-art procedures and extensive numerical experiments are provided to allow reproducibility. All algorithms developed in this thesis are open-sourced and available online.
• Stochastic Second Order Methods and Finite Time Analysis of Policy Gradient Methods
  
  • Yuan Rui
  , 2023. To solve large scale machine learning problems, first-order methods such as stochastic gradient descent and ADAM are the methods of choice because of their low cost per iteration. The issue with first order methods is that they can require extensive parameter tuning, and/or knowledge of the parameters of the problem. There is now a concerted effort to develop efficient stochastic second order methods to solve large scale machine learning problems. The motivation is that they require less parameter tuning and converge for wider variety of models and datasets. In the first part of the thesis, we presented a principled approach for designing stochastic Newton methods for solving both nonlinear equations and optimization problems in an efficient manner. Our approach has two steps. First, we can re-write the nonlinear equations or the optimization problem as desired nonlinear equations. Second, we apply new stochastic second order methods to solve this system of nonlinear equations. Through our general approach, we showcase many specific new second-order algorithms that can solve the large machine learning problems efficiently without requiring knowledge of the problem nor parameter tuning. In the second part of the thesis, we then focus on optimization algorithms applied in a specific domain: reinforcement learning (RL). This part is independent to the first part of the thesis. To achieve such high performance of RL problems, policy gradient (PG) and its variant, natural policy gradient (NPG), are the foundations of the several state of the art algorithms (e.g., TRPO and PPO) used in deep RL. In spite of the empirical success of RL and PG methods, a solid theoretical understanding of even the “vanilla” PG has long been elusive. By leveraging the RL structure of the problem together with modern optimization proof techniques, we derive new finite time analysis of both PG and NPG. Through our analysis, we also bring new insights to the methods with better hyperparameter choices.
• Protocol for quantitative evaluation of the impact of paracrine senescence on cellular reprogramming in cultured cells and mouse models
  
  • Chantrel Jérémy
  • Chen Cheng
  • Zhang Jun
  • Li Han
  STAR Protocols, Elsevier, 2023, 4 (1), pp.102106-1:102106-17. We present a protocol to evaluate the impact of senescence secretome on reprogramming to pluripotency using both cellular and mouse models. First, we describe the in vitro reprogramming procedure using conditioned medium derived from senescent cells. Next, to explore the impact of senescence on in vivo reprogramming, we detail the steps to identify senescent and reprogrammed cells in mouse skeletal muscle, followed by semi-automatic quantification. This protocol can be used to study the effect of paracrine senescence on cellular plasticity. For complete details on the use and execution of this protocol, please refer to von Joest et al. (2022). (10.1016/j.xpro.2023.102106)
  DOI : 10.1016/j.xpro.2023.102106
• Synthetic learning for neural image restoration methods
  
  • Achddou Raphaël
  , 2023. Photography has become an important part of our lives. In addition, expectations in terms of image quality are increasing while the size of imaging devices is decreasing. In this context, the improvement of image processing algorithms is essential.In this manuscript, we are particularly interested in image restoration tasks. The goal is to produce a clean image from one or more noisy observations of the same scene. For these problems, deep learning methods have grown dramatically in the last decade, outperforming the state of the art for the vast majority of traditional tests.While these methods produce impressive results, they have a number of drawbacks. First of all, they are difficult to interpret because of their "black box" operation. Moreover, they generalize rather poorly to acquisition or distortion modalities absent from the training database. Finally, they require large databases, which are sometimes difficult to acquire.We propose to attack these different problems by replacing the data acquisition by a simple image generation algorithm, based on the dead leaves model. Although this model is very simple, the generated images have statistical properties close to those of natural images and many invariance properties (scale, translation, rotation, contrast...). Training a restoration network with this kind of image allows us to identify the important properties of the images for the success of the restoration networks. Moreover, this method allows us to get rid of the data acquisition, which can be tedious.After presenting this model, we show that the proposed method allows to obtain restoration performances very close to traditional methods for relatively simple tasks. After some adaptations of the model, synthetic learning also allows us to tackle difficult concrete problems, such as RAW image denoising. We then propose a statistical study of the color distribution of natural images, allowing to elaborate a realistic parametric model of color sampling for our generation algorithm. Finally, we present a new perceptual loss function based on camera evaluation protocols, using the dead leaf images. The training performed with this function shows that we can jointly optimize the evaluation of the cameras, while keeping identical performances on natural images.
• L’expérience de la thèse CIFRE : étudier et rendre visible la recherche salariée en design
  
  • Reunkrilerk Dorian
  • Peneau Justine
  • Zaidi Nawelle
  • Bonnardot Zoé
  • Huron Samuel
  Sciences du Design, Presses Universitaires de France, 2023, n° 18 (2), pp.42-60. Cet article s’inscrit dans l’étude du rôle de la recherche en design comme levier de transformation. Plus particulièrement, il identifie un premier périmètre de travail sur les spécificités de la thèse CIFRE (Convention industrielle de formation par la recherche) dans le champ du design, un format de plus en plus mobilisé pour le financement de la recherche en design. Par ce travail collectif, les auteur·rice·s, ayant fait l’expérience de ce dispositif, souhaitent rendre visible le point de vue de ceux et celles qui en sont le cœur (les doctorant·e·s), afin de penser des voies d’adaptation de la thèse CIFRE en design. En capitalisant sur l’expérience vécue de 19 doctorant·e·s et docteur·e·s, l’article identifie quatre volets de préoccupations, et ouvre ainsi des perspectives pour la communauté des chercheur·e·s en design et les organisations : la place donnée au design en tant que discipline au sein du parcours, l’intérêt de rendre sa thèse plastique, la gestion des temporalités au quotidien et les facteurs de motivations des thèses CIFRE en design. (10.3917/sdd.018.0042)
  DOI : 10.3917/sdd.018.0042
• Neural network-derived perfusion maps: A model-free approach to computed tomography perfusion in patients with acute ischemic stroke
  
  • Gava Umberto
  • D’agata Federico
  • Tartaglione Enzo
  • Renzulli Riccardo
  • Grangetto Marco
  • Bertolino Francesca
  • Santonocito Ambra
  • Bennink Edwin
  • Vaudano Giacomo
  • Boghi Andrea
  • Bergui Mauro
  Frontiers in Neuroinformatics, Frontiers Media, 2023, 17. Objective In this study, we investigate whether a Convolutional Neural Network (CNN) can generate informative parametric maps from the pre-processed CT perfusion data in patients with acute ischemic stroke in a clinical setting. Methods The CNN training was performed on a subset of 100 pre-processed perfusion CT dataset, while 15 samples were kept for testing. All the data used for the training/testing of the network and for generating ground truth (GT) maps, using a state-of-the-art deconvolution algorithm, were previously pre-processed using a pipeline for motion correction and filtering. Threefold cross validation had been used to estimate the performance of the model on unseen data, reporting Mean Squared Error (MSE). Maps accuracy had been checked through manual segmentation of infarct core and total hypo-perfused regions on both CNN-derived and GT maps. Concordance among segmented lesions was assessed using the Dice Similarity Coefficient (DSC). Correlation and agreement among different perfusion analysis methods were evaluated using mean absolute volume differences, Pearson correlation coefficients, Bland-Altman analysis, and coefficient of repeatability across lesion volumes. Results The MSE was very low for two out of three maps, and low in the remaining map, showing good generalizability. Mean Dice scores from two different raters and the GT maps ranged from 0.80 to 0.87. Inter-rater concordance was high, and a strong correlation was found between lesion volumes of CNN maps and GT maps (0.99, 0.98, respectively). Conclusion The agreement between our CNN-based perfusion maps and the state-of-the-art deconvolution-algorithm perfusion analysis maps, highlights the potential of machine learning methods applied to perfusion analysis. CNN approaches can reduce the volume of data required by deconvolution algorithms to estimate the ischemic core, and thus might allow the development of novel perfusion protocols with lower radiation dose deployed to the patient. (10.3389/fninf.2023.852105)
  DOI : 10.3389/fninf.2023.852105