Publications

2024

• Domain Adaptation for Learned Image Compression with Supervised Adapters
  
  • Presta Alberto
  • Spadaro Gabriele
  • Tartaglione Enzo
  • Fiandrotti Attilio
  • Grangetto Marco
  , 2024, pp.33-42. In Learned Image Compression (LIC), a model is trained at encoding and decoding images sampled from a source domain, often outperforming traditional codecs on natural images; yet its performance may be far from optimal on images sampled from different domains. In this work, we tackle the problem of adapting a pre-trained model to multiple target domains by plugging into the decoder an adapter module for each of them, including the source one. Each adapter improves the decoder performance on a specific domain, without the model forgetting about the images seen at training time. A gate network computes the weights to optimally blend the contributions from the adapters when the bitstream is decoded. We experimentally validate our method over two state-of-the-art pre-trained models, observing improved rate-distortion efficiency on the target domains without penalties on the source domain. Furthermore, the gate’s ability to find similarities with the learned target domains enables better encoding efficiency also for images outside them. (10.1109/DCC58796.2024.00011)
  DOI : 10.1109/DCC58796.2024.00011
• Machine-learning-based technique to establish ASE or Kerr impairment dominance in optical transmission
  
  • Andrenacci Isaia
  • Lonardi Matteo
  • Ramantanis Petros
  • Awwad Élie
  • Irurozki Ekhiñe
  • Clémençon Stephan
  • Almonacil Sylvain
  Journal of Optical Communications and Networking, Piscataway, NJ ; Washington, DC : IEEE : Optical Society of America, 2024, 16 (4), pp.481-492. Data extraction from optical networks has increased substantially with the evolution of monitoring and telemetry methods. Using data analysis and machine learning, this paper aims to derive insights from this data, contributing to the development of self-optimized optical networks. More particularly, it focuses on predicting the Kerr and amplified spontaneous emission dominance by examining the fluctuations in the signal-to-noise ratio due to polarization-dependent loss. Building on previous work, which used the SNR statistic as the input feature of machine learning, our primary goal is to enhance prediction precision while concurrently decreasing the computational model’s complexity. After refining the selection parameters of the input features, we observed a 70% reduction in the input feature length with respect to our previous work. The model reached a 98% accuracy rate, and it was able to successfully classify the regimes in a limited set of unseen experimental instances. (10.1364/JOCN.506931)
  DOI : 10.1364/JOCN.506931
• On the Resiliency of Protected Masked S-Boxes Against Template Attack in the Presence of Temperature and Aging Misalignments
  
  • Anik Md Toufiq Hasan
  • Danger Jean-Luc
  • Guilley Sylvain
  • Karimi Naghmeh
  IEEE Transactions on Very Large Scale Integration (VLSI) Systems, IEEE, 2024, pp.1-14. Profiling side-channel analysis (SCA) attacks have received a lot of attention in the recent years. To perpetrate these attacks, the adversary creates a profile of a sensitive device at her disposal, and uses it to model a target device with a similar implementation to extract its key. Template attacks are recognized to be the most powerful profiling attacks when the measurement noise is Gaussian. To tackle SCA attacks, different countermeasures have been proposed in the literature, among which masking schemes have received the utmost attention. By adding randomness to the circuit, masking schemes prevent the adversary from relating the power consumption to the evaluated data, thus making the attack more difficult. In this article, we study the protection provided by several masking schemes against template attacks. More precisely, we investigate how the success of the template attack is changed when there is a misalignment between the target and profiling devices in terms of temperature and process variations. As another innovative analysis angle, we extensively study the impact of device aging on the template attack and demonstrate quantitatively how aging misalignments in side-channel traces, between the profiling and the target devices, do hinder the attack. The main objective of this study is to get accurate and numerous results allowing the designer to compare different implementations of masking and accordingly choose one which corresponds to the best compromise among complexity, security, and sensitivity to temperature and aging. We target the S-Box module of the unprotected PRESENT cipher along with its five masking variants including global lookup table (GLUT), rotating S-Box masking (referred to as RSM-LOG hereafter), RSM with read-only memory (RSM-ROM), Ishai–Sahai–Wagner masking (ISW), and threshold implementation (TI). The unprotected circuit gets impacted by such aging misalignments with $\approx$ 12.5% increase in the number of traces needed to reach 80% success rate (SR) in the course of 20 weeks of aging at 105 $^{\circ}$ C. Such increase is 23.3%, 37.19%, and 38.24% for ISW, GLUT, and RSM-LOG, respectively. For RSM-ROM such increase is 193.37% for ten weeks of aging. Interestingly, TI is not much affected by aging in this regard. (10.1109/TVLSI.2024.3374257)
  DOI : 10.1109/TVLSI.2024.3374257
• Feasibility Study of Joint Modeling of Environmental and Morphological Effects for WBAN
  
  • Youssef Badre
  • Roblin Christophe
  , 2024, pp.1-5. <div><p>In this article, we assess the feasibility of the joint modeling of the effects of the environment and morphology in the context of the Wireless Body Area Networks (WBAN) for the scenario based approach in the 1 st Ultra Wide Band (UWB) sub-band [3.1, 4.8] GHz. Previous works have enabled to obtain important representative statistical samples for these two variables, to our knowledge not reached to date for WBAN communications, with a dedicated methodology combining simulations results and experimental designs supported by measurements. Our approach is to combine these variables in a disjointed manner in order to extract a more complete model and naturally limit the simulations campaigns to a reference subject for all the environments to be considered. Thus, for a given subject we therefore estimate the propagation channel transfer function from its on-body contribution and that of the environment obtained with the reference subject. This feasibility study is the first step in extracting interesting future models and a huge saving of time. The comparison results obtained between the estimated transfer function and the nominally calculated one highlight fairly small differences in the majority of cases, which is encouraging for the use of this simplifying approach.</p></div> (10.23919/EuCAP60739.2024.10500941)
  DOI : 10.23919/EuCAP60739.2024.10500941
• Single-Branch Hybrid Resistance Compression Technique for Enhanced Rectifier Performance
  
  • Yang Furong
  • Georgiadis Apostolos
  • Daskalakis Spyros
  • Niotaki Kyriaki
  • Hu Yichao
  • Yang Jichao
  • Song Chaoyun
  , 2024, pp.1-4. In this paper, we introduce a novel resistance compression technique which is termed as the Single-branch Hybrid Resistance Compression Technique (SHRCT) aimed at enhancing the performance of rectifiers. While the conventional Resistance Compression Network (RCN) mitigates non-linear impedance variation effects in rectifier circuits, ensuring stable operation under varying input power and load resistance conditions, it typically necessitates at least two branches and consequently multiple surface-mount devices (SMD), resulting in substantial power losses. The introduction of SHRCT revolutionizes resistance compression by eliminating the need for multiple branches, thus reducing the requirement for SMDs, and improving overall efficiency. The proposed rectifier exhibits a notably high RF-DC conversion efficiency, reaching up to a maximum of 74.2%. The SHRCT empowers the rectifier with robust performance capabilities even when faced with huge variations in load resistance and input power. Specifically, the rectifier maintains a consistent 50% RF-DC conversion efficiency across load resistance values ranging from 1k to 10k ohms and input power levels spanning from -10 to 0 dBm. (10.23919/EuCAP60739.2024.10500994)
  DOI : 10.23919/EuCAP60739.2024.10500994
• Learning High-Quality and General-Purpose Phrase Representations
  
  • Chen Lihu
  • Varoquaux Gaël
  • Suchanek Fabian M.
  , 2024. Phrase representations play an important role in data science and natural language processing, benefiting various tasks like Entity Alignment, Record Linkage, Fuzzy Joins, and Paraphrase Classification. The current state-of-theart method involves fine-tuning pre-trained language models for phrasal embeddings using contrastive learning. However, we have identified areas for improvement. First, these pretrained models tend to be unnecessarily complex and require to be pre-trained on a corpus with context sentences. Second, leveraging the phrase type and morphology gives phrase representations that are both more precise and more flexible. We propose an improved framework to learn phrase representations in a context-free fashion. The framework employs phrase type classification as an auxiliary task and incorporates character-level information more effectively into the phrase representation. Furthermore, we design three granularities of data augmentation to increase the diversity of training samples. Our experiments across a wide range of tasks show that our approach generates superior phrase embeddings compared to previous methods while requiring a smaller model size. The code is available at https: //github.com/tigerchen52/PEARL
• Understanding Interaction and Breakouts of Safety Boundaries in Virtual Reality Through Mixed-Method Studies
  
  • Tseng Wen-Jie
  • Kontrazis Petros Dimitrios
  • Lecolinet Eric
  • Huron Samuel
  • Gugenheimer Jan
  , 2024, pp.482-492. Virtual Reality (VR) technologies become ubiquitous, allowing people to employ immersive experiences in their homes. Since VR participants are visually disconnected from their real-world environment, commercial products propose safety boundaries to prevent colliding with their surroundings. However, there is a lack of empirical knowledge on how people perceive and interact with safety boundaries in everyday VR usage. This paper investigates this re- search gap with two mixed-method empirical studies. Study 1 reports an online survey (n=48) collecting data about attitudes towards safety boundaries, behavior while interacting with them, and reasons for breakout. Our analysis with open coding reveals that some VR participants ignored safety boundaries intentionally, even breaking out of them and continuing their actions. Study 2 investigates how and why VR participants intentionally break out when interacting close to safety boundaries and obstacles by replicating breakouts in a lab study (n=12). Our interview and breakout data discover three strategies, revealing VR participants sometimes break out of boundaries based on their real-world spatial information. Finally, we discuss improving future VR safety mechanisms by supporting participants’ real-world spatial mental models using landmarks. (10.1109/VR58804.2024.00069)
  DOI : 10.1109/VR58804.2024.00069
• The Social Impact of Extended Reality Spatial Productivity in Constrained, Public and Passenger Spaces
  
  • Medeiros Daniel
  • Wilson Graham
  • Brewster Stephen
  • Mcgill Mark
  , 2024. In this workshop submission, we reflect on the need to balance a breadth of design considerations when supporting mobile, spatial productivity. Whilst performance, ergonomics and usability remain key, there is an increasing realisation that the social impact of our designs must also be considered - from the social comfort and acceptability of a given workspace or interaction technique, to the social collisions they provoke with other passengers, to the environmental and social awareness the design facilitates in allowing the user to focus on their task whilst maintaining awareness of their environment and those around them.
• METHOD FOR CHARACTERIZING AN ORGAN OF A PATIENT IN A MEDICAL IMAGE
  
  • Vétil Rebeca
  • Abi-Nader Clément
  • Bône Alexandre
  • Rohé Marc-Michel
  • Gori Pietro
  • Bloch Isabelle
  , 2024.
• Reinforcement Learning for Uncoordinated Multiple Access
  
  • Robaglia Benoît-Marie
  , 2024. Distributed Medium Access Control (MAC) protocols are fundamental in wireless communication, yet traditional random access-based protocols face significant limitations dealing with the Internet-of-Things (IoT) use cases. Indeed, they struggle with latency guarantees, making them unsuitable for Ultra Reliable Low Latency Communications (URLLC). This thesis addresses these challenges by leveraging the potential of Deep Reinforcement Learning (DRL), a paradigm where decision-makers optimize actions by interacting with an environment.This thesis tackles key challenges in the Medium Access (MA) problem for URLLC networks, including the latency in centralized protocols, the collision and retransmission issues in Grant-Free (GF) protocols, the complexities to handle device heterogeneity and dynamic environments. Furthermore, the thesis explores the integration of new physical layer techniques like Non-Orthogonal Multiple Access (NOMA).Our methodology applies DRL to develop intelligent protocols, which has already shown effectiveness in addressing IoT applications. Initially, we model the URLLC problem within a centralized paradigm, where the Base Station (BS) orchestrates device transmissions. This setup has the benefit to ensure collision-free communication but introduces partial observability as the BS does not have access to the users' buffer and channel state. We tackle this problem by introducing two algorithms: FilteredPPO and NOMA-PPO. While the former outperforms the benchmarks in scenarios with periodic traffic patterns, the latter demonstrates superior performance over the state-of-the-art baselines on scenarios with sporadic traffic. The third and fourth contributions, SeqDQN and MCA-PPO, study the application of Multi-Agent Reinforcement Learning (MARL) for URLLC where each device is equipped by a DRL algorithm. While SeqDQN explores a method to reduce non-stationarity and enhances scalability and training efficiency, MCA-PPO presents a theoretically robust solution for the Dynamic Multi-Channel Access (DMCA) challenge allowing users to optimize bandwidth utilization, and thus enhancing the URLLC performance. (10.70675/08ed61cdzb5a5z4c2az9f76z431c4c92727d)
  DOI : 10.70675/08ed61cdzb5a5z4c2az9f76z431c4c92727d
• Pair-Matching: Link Prediction with Adaptive Queries
  
  • Giraud Christophe
  • Issartel Yann
  • Lehericy Luc
  • Lerasle Matthieu
  Mathematical Statistics and Learning, EMS Publishing House, 2024. The pair-matching problem appears in many applications where one wants to discover good matches between pairs of entities or individuals. Formally, the set of individuals is represented by the nodes of a graph where the edges, unobserved at first, represent the good matches. The algorithm queries pairs of nodes and observes the presence/absence of edges. Its goal is to discover as many edges as possible with a fixed budget of queries. Pair-matching is a particular instance of multi-armed bandit problem in which the arms are pairs of individuals and the rewards are edges linking these pairs. This bandit problem is non-standard though, as each arm can only be played once. Given this last constraint, sub-linear regret can be expected only if the graph presents some underlying structure. This paper shows that sub-linear regret is achievable in the case where the graph is generated according to a Stochastic Block Model (SBM) with two communities. Optimal regret bounds are computed for this pair-matching problem. They exhibit a phase transition related to the Kesten-Stigum threshold for community detection in SBM. The pair-matching problem is considered in the case where each node is constrained to be sampled less than a given amount of times. We show how optimal regret rates depend on this constraint. The paper is concluded by a conjecture regarding the optimal regret when the number of communities is larger than 2. Contrary to the two communities case, we argue that a statistical-computational gap would appear in this problem. (10.4171/msl/46)
  DOI : 10.4171/msl/46
• Développement d'un modèle de «Machine Learning» d'aide à la prescription de posologies individualisées destinées aux patients traités par amoxicilline en perfusion continue
  
  • Guillot Robin
  • El-Helali N.
  • Mory Celine
  • Gutton Johann
  • Hocquet Grégory
  • Le Monnier Alban
  • Buronfosse Anne
  • Billuart Olivier
  • Le Folgoc Loïc
  • Maynadier Xavier
  Journal of Epidemiology and Population Health, Elsevier Masson SAS, 2024, 72 (Supplément 1), pp.202294. Introduction Dans les situations infectieuses impliquant une hospitalisation, l'obtention d'une concentration sanguine efficace d'antibiotique constitue un enjeu qui passe par le choix d'une posologie adaptée. Si un pharmacologue peut être sollicité afin de tenir compte des caractéristiques individuelles, sa démarche d'adaptation posologique repose encore sur des principes très généraux. L'objectif de notre étude était de construire un modèle de Machine Learning de prédiction de la concentration d'amoxicilline à partir de la dose administrée et des caractéristiques du patient impliqué. Méthodes Etude rétrospective sur données qui porte sur les patients hospitalisés au sein du Groupe hospitalier Paris Saint-Joseph entre 2018 et 2023, traités par amoxicilline en perfusion continue et pour lesquels un dosage de la concentration sérique de cet antibiotique a été réalisé. Les caractéristiques démographiques (âge, sexe, IMC) et biologiques (fonctions rénale, hépatique, cardiaque) ont été extraites du dossier médical informatisé. Plusieurs modèles de prédiction de la concentration sanguine ont été construits. Le plus simple s'est fondé sur une régression linéaire utilisant la seule dose administrée. Trois modèles plus avancés - régression linéaire multivariée, forêt aléatoire, XGBoost – se sont appuyés sur un sous-ensemble de caractéristiques patient déterminé par la méthode de sélection de variables LASSO. La prédiction a été considérée comme performante lorsqu'elle correspondait à la concentration observée avec une marge d'erreur de 20 %. Résultats L'entrainement des modèles a porté sur 237 dosages sériques, l’évaluation de leur performance sur 57 dosages; 19 % des prédictions de la concentration sanguine à partir de la seule dose administrée étaient adaptées. Les modèles prenant aussi en compte les caractéristiques liées au poids et à la fonction rénale du patient amélioraient les performances. La régression linéaire multivariée, la forêt aléatoire et le XGBoost atteignaient respectivement 47 % [35 %-60 %] (IC95%), 47 % [35 %-61 %] et 51 % [37 %-63 %] de bonnes prédictions. Conclusion Les résultats suggèrent qu'un ajustement des posologies en fonction de caractéristiques individuelles est bénéfique. Une évaluation de la performance de ces modèles en prospectif est en cours afin de déterminer leur plus-value par rapport aux pratiques actuelles des pharmacologues. (10.1016/j.jeph.2024.202294)
  DOI : 10.1016/j.jeph.2024.202294
• Estimation of a causal directed acyclic graph process using non-gaussianity
  
  • Einizade Aref
  • Giraldo Jhony H.
  • Malliaros Fragkiskos D.
  • Hajipour Sardouie Sepideh
  Digital Signal Processing, Elsevier, 2024, 146, pp.104400. In machine learning and data mining, causal relationship discovery is a critical task. While the state-of-the-art Vector Auto-Regressive Linear Non-Gaussian Acyclic Model (VAR-LiNGAM) method excels in uncovering both instantaneous and time-lagged connections, it entails analyzing multiple VAR matrices, leading to heightened parameter complexity. To address this challenge, we introduce the Causal Graph Process-LiNGAM (CGP-LiNGAM), a novel approach that significantly reduces parameter load by focusing on a single causal graph, a Directed Acyclic Graph (DAG). Leveraging Graph Signal Processing (GSP) techniques, our method interprets causal relations with graph shift invariance and uniqueness. Our experimental results demonstrate the superiority and robustness of CGP-LiNGAM, particularly in high-noise environments. Moreover, we showcase its real-world applicability in studying brain connectivity during sleep, underlining its compatibility with previous sleep-related neuroscientific research. (10.1016/j.dsp.2024.104400)
  DOI : 10.1016/j.dsp.2024.104400
• Degradation-Invariant Music Indexing
  
  • Mignot Rémi
  • Peeters Geoffroy
  , 2024. For music indexing robust to sound degradations and scalable for big music catalogs, this scientific report presents an approach based on audio descriptors relevant to the music content and invariant to sound transformations (noise addition, distortion, lossy coding, pitch/time transformations, or filtering e.g.). To achieve this task, one of the key point of the proposed method is the definition of high-dimensional audio prints, which are intrinsically (by design) robust to some sound degradations. The high dimensionality of this first representation is then used to learn a linear projection to a sub-space significantly smaller, which reduces again the sensibility to sound degradations using a series of discriminant analyses. Finally, anchoring the analysis times on local maxima of a selected onset function, an approximative hashing is done to provide a better tolerance to bit corruptions, and in the same time to make easier the scaling of the method.
• Shaping Strategies to Embrace Nonlinear Effects in Optical Fiber Communications
  
  • Liu Jingtian
  , 2024. The main impediment in long-distance communications is nonlinear interference (NLI), stemming from nonlinear effects in optical fibers. While Digital signal processing algorithms offer partial mitigation, the inherent nonlinear nature of the fiber, coupled with predominant dispersion effects, continues to challenge the increase of transmission throughputs. Addressing nonlinearity at the information source through signal modulation technology is at the heart of our research. Traditional modulation schemes, as spectral efficiency climbs, such as QAM, become increasingly susceptible to NLI while their Mean Squared Euclidean Distance (MSED) diminishes. While multi-dimensional (MD) modulation yields improved linear and partial nonlinear gains, it has not yet demonstrated tangible benefits. On the other hand, the emergence of probabilistic constellation shaping (PCS), preferred for its enhanced linear gain and compatibility with conventional modulation hardware and software, introduces additional NLI. Consequently, the design of nonlinear-tolerant PCS is emerging as a pivotal research direction. Our thesis begins with a novel MD modulation for uniformly distributed signals. Then, we propose a novel approach combining MD with PCS to examine performance variations. Delving into PCS, we investigate the enumerative sphere shaping distribution matcher (DM), initially from an MD stance, and design a DM optimized for nonlinear tolerance over shorter distances. Subsequently, we introduce a new NLI measurement technique, accounting for dispersion effects. Integrating this with the sequence selection framework of PCS, we achieve successful long-distance transmission with notable nonlinear gains. (10.70675/5feb3c63z1895z4083zbf5ezd2a46463b6b2)
  DOI : 10.70675/5feb3c63z1895z4083zbf5ezd2a46463b6b2
• Improving the automatic diagnosis of hepatocellular carcinoma with contrastive learning
  
  • Sarfati Emma
  , 2024. Several deep learning methods have been proposed to automatically classify liver lesions in MRI or CT, with good performance [1,2,3]. They all share a classical training procedure. Contrastive learning (CL) is a novel deep learning paradigm where pairs of cases, instead of cases taken in isolation, are leveraged to train the model. CL methods can be either trained without labels, in this case it learns global mathematical representations of the input images [4], or it can use labels during training in order to help to discriminate input images based on a specific characteristic [5]. In this study, we evaluate the potential of CL to improve the automatic classification of hepatocellular carcinoma (HCC) in CT-scans. We formulate it through a binary classification problem (HCC versus no HCC).
• Exploring nonlinear dynamics and amplitude squeezing of quantum dot lasers
  
  • Ding Shihao
  , 2024. Photonic integrated circuits (PICs) utilizing silicon photonics technology show significant potential in high-speed communication systems, optical computing, and quantum technology. Quantum dot (QD) lasers, particularly those epitaxially grown on silicon, exhibit notable characteristics such as strong defect tolerance, low threshold current, and good temperature stability. As a result, they are gradually emerging as promising on-chip laser sources for PICs. This thesis aims to explore the nonlinear dynamics and quantum state properties of QD lasers, laying the foundation for various potential applications. The first section of the thesis delves into the role of the linewidth enhancement factor (αH-factor) in QD lasers. I employed an optical phase modulation technique to extract the above-threshold αH-factor of QD lasers. The small αH-factor of QD lasers strongly improved optical feedback tolerance. This facilitates the development of high-speed optical transmission on chips without an optical isolator and increases integration density. The second section explores the dynamic characteristics of QD lasers under optoelectronic feedback (OEF). Unlike optical feedback, I demonstrated that QD lasers exhibit enhanced sensitivity to OEF, leading to various complex dynamics. This extreme sensitivity is crucial for the advancement of on-chip silicon-based optical computing and Ising machine applications, a phenomenon not observed in quantum well (QW) lasers. In the third section, I analyzed the dependency of frequency noise on external carrier noise. The research revealed that quiet pumping is highly advantageous for minimizing the αH-factor, reducing frequency noise, and consequently narrowing the optical linewidth of QD lasers. The last section highlights the generation of amplitude-squeezed light directly from a constant-current-driven semiconductor QD laser. With the quiet pump, I achieved a substantial gigahertz squeezing bandwidth at room temperature with a squeezing ratio of 3.5 dB. The extreme reflection insensitivity of the squeezed QD generator under optical feedback, in contrast to a reference laser using standard QW technology, is also demonstrated with a squeezing ratio further improved to 5.7 dB. Three distinct measurements, including the sub-shot-noise radiofrequency spectrum, sub-Poissonian photon statistics, and second-order correlation function at zero delay, validate my findings. This research establishes a foundational framework for compact and highly efficient photonic quantum integrated circuits, showcasing the immense application potential of QD lasers in both classical and quantum photonics fields. (10.70675/5650c2b5z0b02z430fza0e9ze1bd36b1a6d9)
  DOI : 10.70675/5650c2b5z0b02z430fza0e9ze1bd36b1a6d9
• Photonic quantum information processing using the frequency continuous-variable of single photons
  
  • Fabre Nicolas
  • Chabaud Ulysse
  , 2024.
• A Formal Verification Approach to Handle Attack Graphs
  
  • Catta Davide
  • Leneutre Jean
  • Mijatovic Antonina
  • Ulin Johanna
  • Malvone Vadim
  , 2024, 3, pp.125-132. We propose a formalization of attack graphs through a multi-agent approach. Specifically, we focus on dynamic scenarios that capture the interaction between an attacker and defenders during a cyberattack. We introduce a formal definition of an attack graph using interpreted systems, demonstrating how this formalization enables us to express interesting security properties. Finally, we present a tool AG2IS, which we have developed as an implementation of our formal definitions, to perform the formal verification of attack graphs. (10.5220/0012310000003636)
  DOI : 10.5220/0012310000003636
• Cycle-Accurate Virtual Prototyping with Multiplicity
  
  • Genius Daniela
  • Apvrille Ludovic
  , 2024, 1, pp.187-194. Model-based design for large applications, especially the mapping of applications' tasks to execution nodes, remains a challenge. In this paper, we explore applications comprising multiple identical software tasks intended for deployment across diverse execution nodes. While these tasks are expected to have a unified representation in their SysML-like block diagrams, each must be specifically mapped to individual processor cores to achieve granular performance optimization. Additionally, inter-task communications should be allocated across multiple channels. We further demonstrate a method for automatically generating parallel POSIX C code suitable for a multiprocessor-on-chip. Our approach has proven especially effective for high-performance streaming applications, notably when such applications have a master-worker task structure. (10.5220/0012386100003645)
  DOI : 10.5220/0012386100003645
• System Architects Are not Alone Anymore: Automatic System Modeling with AI
  
  • Apvrille Ludovic
  • Sultan Bastien
  , 2024, 1, pp.27-38. System development cycles typically follow a V-cycle, where modelers first analyze a system specification before proposing its design. When utilizing SysML, this process predominantly involves transforming natural language (the system specification) into various structural and behavioral views employing SysML diagrams. With their proficiency in interpreting natural text and generating results in predetermined formats, Large Language Models (LLMs) could assist such development cycles. This paper introduces a framework where LLMs can be leveraged to automatically construct both structural and behavioral SysML diagrams from system specifications. Through multiple examples, the paper underscores the potential utility of LLMs in this context, highlighting the necessity for feeding these models with a well-defined knowledge base and an automated feedback loop for better outcomes. (10.5220/0012320100003645)
  DOI : 10.5220/0012320100003645
• Find the lady: permutation and re-synchronization of deep neural networks
  
  • de Sousa Trias Carl
  • Mitrea Mihai Petru
  • Fiandrotti Attilio
  • Cagnazzo Marco
  • Chaudhuri Sumanta
  • Tartaglione Enzo
  , 2024, 38 (19), pp.21001-21009. Deep neural networks are characterized by multiple symmetrical, equi-loss solutions that are redundant. Thus, the order of neurons in a layer and feature maps can be given arbitrary permutations, without affecting (or minimally affecting) their output. If we shuffle these neurons, or if we apply to them some perturbations (like fine-tuning) can we put them back in the original order i.e. re-synchronize? Is there a possible corruption threat? Answering these questions is important for applications like neural network white-box watermarking for ownership tracking and integrity verification. We advance a method to re-synchronize the order of permuted neurons. Our method is also effective if neurons are further altered by parameter pruning, quantization, and fine-tuning, showing robustness to integrity attacks. Additionally, we provide theoretical and practical evidence for the usual means to corrupt the integrity of the model, resulting in a solution to counter it. We test our approach on popular computer vision datasets and models, and we illustrate the threat and our countermeasure on a popular white-box watermarking method. (10.1609/aaai.v38i19.30091)
  DOI : 10.1609/aaai.v38i19.30091
• HKD-SHO: A hybrid smart home system based on knowledge-based and data-driven services
  
  • Qiu Mingming
  • Najm Elie
  • Sharrock Rémi
  • Traverson Bruno
  , 2024. (10.48550/arXiv.2402.15521)
  DOI : 10.48550/arXiv.2402.15521
• A Statistical analysis of algorithms dedicated for rare events
  
  • Aghbalou Anass
  , 2024. This thesis focuses on establishing statistical guarantees for the efficiency of machine learning algorithms in data-scarce environments, particularly within the contexts of extreme value analysis, transfer learning, and imbalanced classification. We develop probability upper bounds that serve as theoretical assurances for the effectiveness of algorithms tailored to these specific scenarios. Our approach begins with a critique of current statistical methods in data-limited set- tings. We identify limitations in existing frameworks and introduce new probability bounds that are specifically designed to provide guarantees for algorithm performance under data scarcity. These bounds are not just theoretically rigorous but are also di- rectly applicable to practical machine learning challenges. We validate our theoretical findings with empirical studies in each of the three focused areas. The results confirm that our derived bounds are effective in certifying the effi- ciency of algorithms in handling extreme values, transferring knowledge in sparse data domains, and classifying imbalanced datasets. Conclusively, the thesis advances the field of statistical learning by providing precise theoretical guarantees for the performance of algorithms in data-scarce situations. This work is particularly relevant for applications where making accurate inferences with limited data is critical. (10.70675/847de9d7za95ez4a74za609za60cd426d222)
  DOI : 10.70675/847de9d7za95ez4a74za609za60cd426d222
• Finding Subgraphs with Maximum Total Density and Limited Overlap in Weighted Hypergraphs
  
  • Balalau Oana
  • Bonchi Francesco
  • Chan T-H. Hubert
  • Gullo Francesco
  • Sozio Mauro
  • Xie Hao
  ACM Transactions on Knowledge Discovery from Data (TKDD), ACM, 2024, 18 (4), pp.1-21. Finding dense subgraphs in large (hyper)graphs is a key primitive in a variety of real-world application domains, encompassing social network analytics, event detection, biology, and finance. In most such applications, one typically aims at finding several (possibly overlapping) dense subgraphs, which might correspond to communities in social networks or interesting events. While a large amount of work is devoted to finding a single densest subgraph, perhaps surprisingly, the problem of finding several dense subgraphs in weighted hypergraphs with limited overlap has not been studied in a principled way, to the best of our knowledge. In this work, we define and study a natural generalization of the densest subgraph problem in weighted hypergraphs, where the main goal is to find at most k subgraphs with maximum total aggregate density, while satisfying an upper bound on the pairwise weighted Jaccard coefficient, i.e., the ratio of weights of intersection divided by weights of union on two nodes sets of the subgraphs. After showing that such a problem is NP-Hard, we devise an efficient algorithm that comes with provable guarantees in some cases of interest, as well as, an efficient practical heuristic. Our extensive evaluation on large real-world hypergraphs confirms the efficiency and effectiveness of our algorithms. (10.1145/3639410)
  DOI : 10.1145/3639410