Publications

2024

• Device-independent quantum key distribution with arbitrarily small nonlocality
  
  • Wooltorton Lewis
  • Brown Peter
  • Colbeck Roger
  Physical Review Letters, American Physical Society, 2024, 132 (21), pp.210802. Device-independent quantum key distribution (DIQKD) allows two users to set up shared cryptographic key without the need to trust the quantum devices used. Doing so requires nonlocal correlations between the users. However, in [Phys. Rev. Lett. 127, 050503 (2021)] it was shown that for known protocols nonlocality is not always sufficient, leading to the question of whether there is a fundamental lower bound on the minimum amount of nonlocality needed for any DIQKD implementation. Here we show that no such bound exists, giving schemes that achieve key with correlations arbitrarily close to the local set. Furthermore, some of our constructions achieve the maximum of 1 bit of key per pair of entangled qubits. We achieve this by studying a family of Bell-inequalities that constitute all self-tests of the maximally entangled state with a single linear Bell expression. Within this family there exist non-local correlations with the property that one pair of inputs yield outputs arbitrarily close to perfect key. Such correlations exist for a range of Clauser-Horne-Shimony-Holt (CHSH) values, including those arbitrarily close to the classical bound. Finally, we show the existence of quantum correlations that can generate both perfect key and perfect randomness simultaneously, whilst also displaying arbitrarily small CHSH violation; this opens up the possibility of a new class of cryptographic protocol. (10.1103/PhysRevLett.132.210802)
  DOI : 10.1103/PhysRevLett.132.210802
• Joint Metrics for EMF Exposure and Coverage in Real-World Homogeneous and Inhomogeneous Cellular Networks
  
  • Gontier Quentin
  • Wiame Charles
  • Wang Shanshan
  • Di Renzo Marco
  • Wiart Joe
  • Horlin François
  • Tsigros Christo
  • Oestges Claude
  • de Doncker Philippe
  IEEE Transactions on Wireless Communications, Institute of Electrical and Electronics Engineers, 2024, 23 (10), pp.13267 - 13284. This paper evaluates the downlink performance of cellular networks in terms of coverage and electromagnetic field exposure (EMFE), in the framework of stochastic geometry. The model is constructed based on datasets for sub-6 GHz macro cellular networks but it is general enough to be applicable to millimeter-wave networks as well. On the one hand, performance metrics are calculated for β-Ginibre point processes which are shown to faithfully model a large number of motion-invariant networks. On the other hand, performance metrics are derived for inhomogeneous Poisson point processes with a radial intensity measure, which are shown to be a good approximation for motion-variant networks. For both cases, joint and marginal distributions of the EMFE and the coverage, and the first moments of the EMFE are provided and validated by Monte Carlo simulations using realistic sets of parameters from two sub-6 GHz macro urban cellular networks, i.e., 5G NR 2100 (Paris, France) and LTE 1800 (Brussels, Belgium) datasets. In addition, this paper includes the analysis of the impact of the network parameters and discusses the achievable trade-off between coverage and EMFE. (10.1109/TWC.2024.3400612)
  DOI : 10.1109/TWC.2024.3400612
• Misplaced trust in AI : the explanation paradox and the human-centric path. A characterisation of the cognitive challenges to appropriately trust algorithmic decisions and applications in the financial sector
  
  • Bertrand Astrid
  , 2024. As AI is becoming more widespread in our everyday lives, concerns have been raised about comprehending how these opaque structures operate. In response, the research field of explainability (XAI) has developed considerably in recent years. However, little work has studied regulators' need for explainability or considered effects of explanations on users in light of legal requirements for explanations. This thesis focuses on understanding the role of AI explanations to enable regulatory compliance of AI-enhanced systems in financial applications. The first part reviews the challenge of taking into account human cognitive biases in the explanations of AI systems. The analysis provides several directions to better align explainability solutions with people's cognitive processes, including designing more interactive explanations. It then presents a taxonomy of the different ways to interact with explainability solutions. The second part focuses on specific financial contexts. One study takes place in the domain of online recommender systems for life insurance contracts. The study highlights that feature based explanations do not significantly improve non expert users' understanding of the recommendation, nor lead to more appropriate reliance compared to having no explanation at all. Another study analyzes the needs of regulators for explainability in anti-money laundering and financing of terrorism. It finds that supervisors need explanations to establish the reprehensibility of sampled failure cases, or to verify and challenge banks' correct understanding of the AI. (10.70675/216d3647z78bcz4c68z9ba9z2d0a029cf1b4)
  DOI : 10.70675/216d3647z78bcz4c68z9ba9z2d0a029cf1b4
• Non-invasive performance prediction of high-speed softwarized network services with limited knowledge
  
  • Liu Qiong
  • Zhang Tianzhu
  • Linguaglossa Leonardo
  , 2024. Modern telco networks have experienced a significant paradigm shift in the past decade, thanks to the proliferation of network softwarization. Despite the benefits of softwarized networks, the constituent software data planes cannot always guarantee predictable performance due to resource contentions in the underlying shared infrastructure. Performance predictions are thus paramount for network operators to fulfill ServiceLevel Agreements (SLAs), especially in high-speed regimes (e.g., Gigabit or Terabit Ethernet). Existing solutions heavily rely on in-band feature collection, which imposes non-trivial engineering and data-path overhead. This paper proposes a non-invasive performance prediction approach, which complements state-ofthe-art solutions by measuring and analyzing low-level features ubiquitously available in the network infrastructure. Accessing these features does not hamper the packet data path. Our approach does not rely on prior knowledge of the input traffic, VNFs’ internals, and system details. We show that (i) low-level hardware features exposed by the NFV infrastructure can be collected and interpreted for performance issues, (ii) predictive models can be derived with classical ML algorithms, (iii) and can be used to predict performance impairments in real NFV systems accurately. Our code and datasets are publicly available.
• Spectral Structure Analysis of FFT-based Digital Predistortion for Wideband 5G Applications
  
  • Bouazza Tayeb
  • Pham Dang-Kièn Germain
  • Mohellebi Reda
  • Desgreys Patricia
  , 2024, pp.1-5. In this paper, a spectral structure analysis for the FFT-based subband digital predistortion is performed. This analysis consists in investigating whether the spectrum of the signal at the output of the power amplifier (PA) shows any particular structuring of the information. Various scenarios are investigated by simulation and validated by experiments, on a real PA. It is found that some spectral zones are more critical than others to maintain good linearization, which makes possible to reduce the requirements of the feedback path. (10.1109/ISCAS58744.2024.10558459)
  DOI : 10.1109/ISCAS58744.2024.10558459
• Investigating on mobile phone transmitting power under a non-standalone NR network
  
  • Liu Jiang
  • Zhang Yarui
  • Ben Chikha Wassim
  • Wang Shanshan
  • Samaras Theodoros
  • Jawad Ourouk
  • Ourak Lamine
  • Conil Emmanuelle
  • Wiart Joe
  , 2024, pp.1-3. Since the invention of mobile communication technology, successive generations have consistently superseded their predecessors for the past four decades. Nowadays, with the global deployment of the fifth generation (5G) wireless communication network, there has been a notable surge in awareness regarding the potential risks associated with exposure to radio-frequency electromagnetic fields (RFEMF). To address the public concern, it is crucial to investigate and monitor the RF-EMF exposure. In this paper, we aim to measure and analyze the uplink exposure from the daily usage of mobile phone, especially with data-demanding application running in a 5G network. To achieve this objective, extensive measurements of file transfer protocol (FTP) are conducted in the Greater Paris region, where non standalone (NSA) New Radio (NR) is prevalent across most areas. The statistical distribution of measurement data is given and the comparison between Long Term Evolution (LTE) and NSA-NR is analyzed. (10.46620/URSIATRASC24/SFIF9621)
  DOI : 10.46620/URSIATRASC24/SFIF9621
• The Semantic Web : 21st International Conference, ESWC 2024, Hersonissos, Crete, Greece, May 26–30, 2024, Proceedings, Part II
  
  • Meroño Peñuela Albert
  • Dimou Anastasia
  • Troncy Raphaël
  • Hartig Olaf
  • Acosta Maribel
  • Alam Mehwish
  • Paulheim Heiko
  • Lisena Pasquale
  , 2024, 14665, pp.XXVI, 252. The two-volume set LNCS 14664 and 14665 constitutes the refereed proceedings of the 21st International Conference on The Semantic Web, ESWC 2024, held in Hersonissos, Crete, Greece, during May 26-30, 2024. The 32 full papers presented were carefully reviewed and selected from 138 submissions. They focus on all aspects of theoretical, analytical, and empirical aspects of the semantic web, semantic technologies, knowledge graphs and semantics on the web in general. (10.1007/978-3-031-60635-9)
  DOI : 10.1007/978-3-031-60635-9
• The Semantic Web : 21st International Conference, ESWC 2024, Hersonissos, Crete, Greece, May 26–30, 2024, Proceedings, Part I
  
  • Meroño Peñuela Albert
  • Dimou Anastasia
  • Troncy Raphaël
  • Hartig Olaf
  • Acosta Maribel
  • Alam Mehwish
  • Paulheim Heiko
  • Lisena Pasquale
  , 2024, 14664, pp.XXVI, 344. The two-volume set LNCS 14664 and 14665 constitutes the refereed proceedings of the 21st International Conference on The Semantic Web, ESWC 2024, held in Hersonissos, Crete, Greece, during May 26-30, 2024. The 32 full papers presented were carefully reviewed and selected from 138 submissions. They focus on all aspects of theoretical, analytical, and empirical aspects of the semantic web, semantic technologies, knowledge graphs and semantics on the web in general. (10.1007/978-3-031-60626-7)
  DOI : 10.1007/978-3-031-60626-7
• Momentum‐preserving inversion alleviation for elastic material simulation
  
  • Jeong Heejo
  • Kim Seung‐wook
  • Lee JaeHyun
  • Um Kiwon
  • Kee Min Hyung
  • Han JungHyun
  Computer Animation and Virtual Worlds, Wiley, 2024, 35 (3), pp.e2249. Abstract This paper proposes a novel method that enhances the optimization‐based elastic body solver. The proposed method tackles the element inversion problem, which is prevalent in the prediction‐projection approach for numerical simulation of elastic bodies. At the prediction stage, our method alleviates inversions such that the subsequent projection solver can benefit in stability and efficiency. To prevent excessive suppression of predicted inertial motion when alleviating, we introduce a velocity decomposition method and adapt only the non‐rigid motion while preserving the rigid motion, that is, linear and angular momenta. Thanks to the respected inertial motion in the prediction stage, our method produces lively motions while keeping the entire simulation more stable. The experiments demonstrate that our alleviation method successfully stabilizes the simulation and improves the efficiency particularly when large deformations hamper the solver. (10.1002/cav.2249)
  DOI : 10.1002/cav.2249
• A Parser-Based Data Collector for Intrusion Detection
  
  • Quétel Grégor
  • Alata Eric
  • Gimenez Pierre-François
  • Pautet Laurent
  • Robert Thomas
  , 2024, pp.1-2. Intrusion detection systems often struggle to identify attacks directed at applications. A contributing factor is the various syntactical forms these attacks can take. This paper introduces a methodology to design and adapt applicative data collectors (DCs) to software projects by integrating them into the application's parsers. This data collector aims to enhance applications' security by providing semantic information to intrusion detection mechanisms.
• Minimax optimal seriation in polynomial time
  
  • Issartel Yann
  • Giraud Christophe
  • Verzelen Nicolas
  , 2024. We consider the statistical seriation problem, where the statistician seeks to recover a hidden ordering from a noisy observation of a permuted Robinson matrix. In this paper, we tightly characterize the minimax rate for this problem of matrix reordering when the Robinson matrix is bi-Lipschitz, and we also provide a polynomial time algorithm achieving this rate; thereby answering two open questions of [Giraud et al., 2021]. Our analysis further extends to broader classes of similarity matrices. (10.1214/25-AOS2615)
  DOI : 10.1214/25-AOS2615
• WebGraph: The Next Generation (Is in Rust)
  
  • Fontana Tommaso
  • Vigna Sebastiano
  • Zacchiroli Stefano
  , 2024. We report the results of a yearlong effort to port the WebGraph framework from Java to Rust. For two decades WebGraph has been instrumental in the analysis and distribution of large graphs for the research community of TheWebConf, but the intrinsic limitations of the Java Virtual Machine had become a bottleneck for very large use cases, such as the Software Heritage Merkle graph with its half a trillion arcs. As part of this clean-slate implementation of WebGraph in Rust, we developed a few ancillary projects bringing to the Rust ecosystem some missing features of independent interest, such as easy, consistent and zero-cost memory mapping of data structures. WebGraph in Rust offers impressive performance improvements over the previous implementation, enabling open-source graph analytics on very large datasets on top of a modern system programming language. (10.1145/3589335.3651581)
  DOI : 10.1145/3589335.3651581
• AI is Entering Regulated Territory: Understanding the Supervisors' Perspective for Model Justifiability in Financial Crime Detection
  
  • Bertrand Astrid
  • Eagan James R
  • Maxwell Winston
  • Brand Joshua
  , 2024, pp.Article No.: 480, Pages 1 - 21. Artificial intelligence (AI) has the potential to bring significant benefits to highly regulated industries such as healthcare or banking. Adoption, however, remains low. AI's entry into complex sociotechno-legal systems raises issues of transparency, specifically for regulators. However, the perspective of supervisors, regulators who monitor compliance with applicable financial regulations, has rarely been studied. This paper focuses on understanding the needs of supervisors in anti-money laundering (AML) to better inform the design of AI justifications and explanations in highly regulated fields. Through scenario-based workshops with 13 supervisors and 6 banking professionals, we outline the auditing practices and sociotechnical context of the supervisor. By combining the workshops' insights with an analysis of compliance requirements, we identify the AML obligations that conflict with AI opacity. We then formulate seven needs that supervisors have for model justifiability. We discuss the role of explanations as reliable evidence on which to base justifications. (10.1145/3613904.3642326)
  DOI : 10.1145/3613904.3642326
• Random walks on simplicial complexes
  
  • Bonis Thomas
  • Decreusefond Laurent
  • Chi Tran Viet
  • Zhang Iris Zhihan
  , 2024. The notion of Laplacian of a graph can be generalized to simplicial complexes and hypergraphs, and contains information on the topology of these structures. Even for a graph, the consideration of associated simplicial complexes is interesting to understand its shape. Whereas the Laplacian of a graph has a simple probabilistic interpretation as the generator of a continuous time Markov chain on the graph, things are not so direct when considering simplicial complexes. We define here new Markov chains on simplicial complexes. For a given order~$k$, the state space is the set of $k$-cycles that are chains of $k$-simplexes with null boundary. This new framework is a natural generalization of the canonical Markov chains on graphs. We show that the generator of our Markov chain is the upper Laplacian defined in the context of algebraic topology for discrete structure. We establish several key properties of this new process: in particular, when the number of vertices is finite, the Markov chain is positive recurrent. This result is not trivial, since the cycles can loop over themselves an unbounded number of times. We study the diffusive limits when the simplicial complexes under scrutiny are a sequence of ever refining triangulations of the flat torus. Using the analogy between singular and Hodge homologies, we express this limit as valued in the set of currents. The proof of tightness and the identification of the limiting martingale problem make use of the flat norm and carefully controls of the error terms in the convergence of the generator. Uniqueness of the solution to the martingale problem is left open. An application to hole detection is carried.
• Separating common from salient patterns with Contrastive Representation Learning
  
  • Louiset Robin
  • Duchesnay Edouard
  • Grigis Antoine
  • Gori Pietro
  , 2024. Contrastive Analysis is a sub-field of Representation Learning that aims at separating common factors of variation between two datasets, a background (i.e., healthy subjects) and a target (i.e., diseased subjects), from the salient factors of variation, only present in the target dataset. Despite their relevance, current models based on Variational Auto-Encoders have shown poor performance in learning semantically-expressive representations. On the other hand, Contrastive Representation Learning has shown tremendous performance leaps in various applications (classification, clustering, etc.). In this work, we propose to leverage the ability of Contrastive Learning to learn semantically expressive representations well adapted for Contrastive Analysis. We reformulate it under the lens of the InfoMax Principle and identify two Mutual Information terms to maximize and one to minimize. We decompose the first two terms into an Alignment and a Uniformity term, as commonly done in Contrastive Learning. Then, we motivate a novel Mutual Information minimization strategy to prevent information leakage between common and salient distributions. We validate our method, called SepCLR, on three visual datasets and three medical datasets, specifically conceived to assess the pattern separation capability in Contrastive Analysis.
• Obstruction Alternating-time Temporal Logic: A Strategic Logic to Reason about Dynamic Models
  
  • Catta Davide
  • Leneutre Jean
  • Malvone Vadim
  • Murano Aniello
  , 2024. (10.5555/3635637.3662875)
  DOI : 10.5555/3635637.3662875
• Strategic Reasoning under Capacity-constrained Agents
  
  • Ballot Gabriel
  • Malvone Vadim
  • Leneutre Jean
  • Laarouchi Youssef
  , 2024, pp.123--131. (10.5555/3635637.3662859)
  DOI : 10.5555/3635637.3662859
• Investigation of Intensity Noise in an Interband Cascade Laser Epitaxially Grown on Silicon and Designed for High-speed Applications
  
  • Kim Hyunah
  • Didier Pierre
  • Zaminga Sara
  • Huang Heming
  • Díaz-Thomas Daniel Andrés
  • Baranov A. N.
  • Rodriguez Jean-Baptiste
  • Tournié Eric
  • Knötig Hedwig
  • Schwarz Benedikt
  • Cerutti Laurent
  • Spitz Olivier
  • Grillot Frédéric
  , 2024. The high-speed parameters of an interband cascade laser grown on silicon are analyzed through the prism of relative intensity noise. The evolution of the relaxation frequency allows deriving a modulation bandwidth in the GHz range.
• Double InfoGAN for Contrastive Analysis
  
  • Carton Florence
  • Louiset Robin
  • Gori Pietro
  , 2024. Contrastive Analysis (CA) deals with the discovery of what is common and what is distinctive of a target domain compared to a background one. This is of great interest in many applications, such as medical imaging. Current state-of-the-art (SOTA) methods are latent variable models based on VAE (CA-VAEs). However, they all either ignore important constraints or they don't enforce fundamental assumptions. This may lead to suboptimal solutions where distinctive factors are mistaken for common ones (or viceversa). Furthermore, the generated images have a rather poor quality, typical of VAEs, decreasing their interpretability and usefulness. Here, we propose Double InfoGAN, the first GAN based method for CA that leverages the high-quality synthesis of GAN and the separation power of InfoGAN. Experimental results on four visual datasets, from simple synthetic examples to complex medical images, show that the proposed method outperforms SOTA CA-VAEs in terms of latent separation and image quality. Datasets and code are available online.
• Sketch In, Sketch Out: Accelerating both Learning and Inference for Structured Prediction with Kernels
  
  • Ahmad Tamim El
  • Brogat-Motte Luc
  • Laforgue Pierre
  • d'Alché-Buc Florence
  , 2024, 238. Surrogate kernel-based methods offer a flexible solution to structured output prediction by leveraging the kernel trick in both input and output spaces. In contrast to energy-based models, they avoid to pay the cost of inference during training, while enjoying statistical guarantees. However, without approximation, these approaches are condemned to be used only on a limited amount of training data. In this paper, we propose to equip surrogate kernel methods with approximations based on sketching, seen as low rank projections of feature maps both on input and output feature maps. We showcase the approach on Input Output Kernel ridge Regression (or Kernel Dependency Estimation) and provide excess risk bounds that can be in turn directly plugged on the final predictive model. An analysis of the complexity in time and memory show that sketching the input kernel mostly reduces training time while sketching the output kernel allows to reduce the inference time. Furthermore, we show that Gaussian and sub-Gaussian sketches are admissible sketches in the sense that they induce projection operators ensuring a small excess risk. Experiments on different tasks consolidate our findings.
• AirCon: Over-the-Air Consensus for Wireless Blockchain Networks
  
  • Xie Xin
  • Hua Cunqing
  • Hong Jianan
  • Gu Pengwenlong
  • Xu Wenchao
  IEEE Transactions on Mobile Computing, Institute of Electrical and Electronics Engineers, 2024, 23 (5), pp.4566-4582. (10.1109/TMC.2023.3292898)
  DOI : 10.1109/TMC.2023.3292898
• Airway tree caliber heterogeneity and airflow obstruction among older adults
  
  • Vameghestahbanati Motahareh
  • Kingdom Leina
  • Hoffman Eric
  • Kirby Miranda
  • Allen Norrina
  • Angelini Elsa
  • Bertoni Alain
  • Hamid Qutayba
  • Hogg James
  • Jacobs David
  • Laine Andrew
  • Maltais Francois
  • Michos Erin
  • Sack Coralynn
  • Sin Don
  • Watson Karol
  • Wysoczanksi Artur
  • Couper David
  • Cooper Christopher
  • Han Meilan
  • Woodruff Prescott
  • Tan Wan
  • Bourbeau Jean
  • Barr R. Graham
  • Smith Benjamin
  Journal of Applied Physiology, American Physiological Society, 2024, 136 (5), pp.1144-1156. In this study, by leveraging two community-based samples and a case-control study of heavy smokers, we show that among older adults, airway tree caliber heterogeneity quantified by CT is associated with airflow obstruction and COPD independent of age, sex, height, race-ethnicity, and dysanapsis. These observations suggest that airway tree caliber heterogeneity is a structural trait associated with low baseline lung function and normal decline trajectory that is relevant to COPD. (10.1152/japplphysiol.00694.2022)
  DOI : 10.1152/japplphysiol.00694.2022
• Deliverable D1 - Technical Report NF-PERSEUS 2023
  
  • Abdel Nour Charbel
  • Adjih Cédric
  • Amis Karine
  • Begaud Xavier
  • Crussière Matthieu
  • Durant Antoine
  • Di Renzo Marco
  • Douillard Catherine
  • El Hassani Hajar
  • Farah Joumana
  • Fijalkow Inbar
  • Gaillot Davy
  • Gorce Jean-Marie
  • Goursaud Claire
  • Guillaud M.
  • Le Ruyet Didier
  • Asma Mabrouk
  • Paganini Pascal
  • Pham Dang-Kièn Germain
  • Prabhu Balakrishna
  • Rekaya Ben Othman Ghaya
  • Simon Eric Pierre
  • Zayani Rafik
  , 2024, pp.1-86. This is deliverable D1 "Technical Report NF-PERSEUS 2023". This document provides an overview of the progress of the studies undertaken in NF-PERSEUS project and their progress by the end of 2023. Chapter 1 delves into the advantages of cell-free architectures and their pivotal role in shaping the future of 6G wireless networks. Cell-free architectures, which move away from the traditional cellular network model, offer several compelling benefits that make them a promising solution for the next generation of mobile communication. Chapter 2 describes the main use cases identified for the NF-PERSEUS project and their KPI, providing an inventory of use case specifications, representative deployment scenarios and technical requirements. It presents also the MAMIMOSA sounder which will be employed to perform massive MIMO channel measurments for the considered NF-PERSEUS scenarios. Chapter 3 examines and discusses relevant reference scenarios based on RIS. This analysis sheds light on the technical challenges involved and the potential performance that can be achieved. Chapter 4 provides an in-depth analysis of the performance metrics used to evaluate cell-free network architectures. It presents the studies and research activities undertaken within Work Package 3 (WP3) and summarizes the progress made by the end of 2023. It presents some potential PHY layer solutions studied in the framework of NF-PERSEUS project. These solutions comprise advanced precoding/combing schemes and multi-carrier waveforms which are adequate for cell-free architectures. Chapter 5 provides an overview of the progress of the studies undertaken in WP4 and their progress by the end of 2023. WP4 deals with radio resource management and aims at introducing novel multi-user access schemes and resource allocation algorithms dedicated to distributed antenna systems, with an emphasis on achieving power and spectrum efficient massive access in scalable B5G sub-7GHz networks. Chapter 6 presents an overview of electromagnetically consistent communication models for Reconfigurable Intelligent Surfaces (RISs). These models aim to describe the communication mechanisms and performance characteristics of RISs, which are emerging technologies that can dynamically control the propagation of electromagnetic waves. The chapter then goes on to showcase preliminary results obtained based on a communication model for RISs that is grounded in multiport network theory. Multiport network theory provides a framework for analyzing the electromagnetic behavior of complex systems, which is particularly relevant for understanding the operation and capabilities of RISs. Chapter 7 summarizes the studies undertaken in WP2 related to Radio-Frequency Front-End Modules, Reconfigurable antennas and RIS aspects of the project NF-PERSEUS. Specifically, it provides details about the design and manufacturing of Hybrid PA architecture, design and manufacturing of highly efficient miniature and reconfigurable antennas using in particular agile metamaterials and/or biosourced technologies. Chapter 8 gives the conclusion.
• Generative models for complex visual content
  
  • Careil Marlène
  , 2024. In this thesis, we explore generative image models with a focus on improving the compositionality of objects as well as enabling better user controllability. Semantic image synthesis is a specific type of conditional generative task that enables to tackle both problems. It consists of synthesizing images conditioned on semantic segmentation maps which indicate per-pixel class information. We aim to develop methods that alleviate the need to train on large-scale annotated datasets. Toward this goal, we make the following contributions. Firstly, in OCO-GAN, we design a generative adversarial model that shares backbones for unconditional and semantic image synthesis and is jointly trained for both tasks. When little unconditional or labeled data are available, we show thatwe leverage synergy between the two tasks which benefit from each other. Secondly, in CAT for Class Affinity Transfer, we design a few-shot transfer method for semantic image synthesis applicable to diffusion models and Generative Adversarial Networks. It leverages a pretrained semantic image synthesis model trained on a large source dataset and finetunes it on a small target dataset. This method exploits class affinity between source and target classes to provide prior knowledge to the model when finetuning on the target dataset. Thirdly, we tackle a more challenging scenario that is training-free and is a more flexible version of semantic image synthesis in ZestGuide. We replace the per-pixel class label conditioning with a free-form text conditioning. We build upon large pretrained text-to-image diffusion models to develop a method that adapts the generation process to take into account this spatial conditioning. Finally, in PerCo, for Perceptual Compression, we explore how to use generative models for the task of compression. Similar to semantic image synthesis, we condition a generative model on discrete representation, but in this case, the discrete representation is learned and constrained to have a fixed low bitrate. We target image compression at extremely low bitrates to recover realistic images with satisfying semantic preservation. After introducing the context and related work on generative image models on which this thesis builds upon, we present each of the four contributions in detail, with a chapter dedicated to each of them. Finally, we will conclude this document with a discussion of the limitations of the presented work and perspectives for future research. (10.70675/71c84118z5b99z48a2za601zc5fb1c90b2b9)
  DOI : 10.70675/71c84118z5b99z48a2za601zc5fb1c90b2b9
• Variational Perspective on Fair Edge Prediction
  
  • Gourru Antoine
  • Laclau Charlotte
  • Choudhary Manvi
  • Largeron Christine
  , 2024, 14641, pp.93-104. Algorithmic fairness has been of great interest in the machine learning community and more recently in the graph context. In this paper, we address the problem of dyadic fairness where the task at hand is edge prediction, and the population of interest (nodes) is divided into a protected and a non-protected group, e.g. men and women. The goal is then to ensure that there should be no statistically significant difference in the prediction outcomes between the two groups, after accounting for any relevant factors that may impact the outcome. To proceed, we design a novel loss based on the variational information bottleneck principle to learn individual node representation while controlling a given level of dyadic fairness. The optimization of the loss is done with a Graph Neural Network. Experiments carried out on several real-world datasets confirmed the capacity of the proposed method, to maintain high accuracy on the edge prediction task while significantly reducing potential bias. (10.1007/978-3-031-58547-0_8)
  DOI : 10.1007/978-3-031-58547-0_8