Publications

2024

• Towards Efficient Exploitation of Large Knowledge Bases by Context Graphs
  
  • Mimouni Nada
  • Moissinac Jean-Claude
  , 2024. One problem related to the exploitation of knowledge graphs, in particular when processing with machine learning methods, is the scaling up problem. We propose here a method to significantly reduce the size of the used graphs to focus on a useful part in a given usage context. We define the notion of context graph as an extract from one or more general knowledge bases (such as DBpedia,Wikidata, Yago) that contains the set of information relevant to a specific domain while preserving the properties of the original graph.We validate the approach on a DBpedia excerpt for entities related to the Data&Musée project and the KORE reference set according to two aspects: the coverage of the context graph and the preservation of the similarity between its entities. The results show that the use of context graphs makes the exploitation of large knowledge bases more manageable and efficient while preserving the properties of the initial graph.
• Socio-Emotional Response Generation: A Human Evaluation Protocol for LLM-Based Conversational Systems
  
  • Vanel Lorraine
  • Ramos Vela Ariel R.
  • Yacoubi Alya
  • Clavel Chloé
  , 2024. Conversational systems are now capable of producing impressive and generally relevant responses. However, we have no visibility nor control of the socio-emotional strategies behind state-of-the-art Large Language Models (LLMs), which poses a problem in terms of their transparency and thus their trustworthiness for critical applications. Another issue is that current automated metrics are not able to properly evaluate the quality of generated responses beyond the dataset's ground truth. In this paper, we propose a neural architecture that includes an intermediate step in planning socio-emotional strategies before response generation. We compare the performance of open-source baseline LLMs to the outputs of these same models augmented with our planning module. We also contrast the outputs obtained from automated metrics and evaluation results provided by human annotators. We describe a novel evaluation protocol that includes a coarse-grained consistency evaluation, as well as a finer-grained annotation of the responses on various social and emotional criteria. Our study shows that predicting a sequence of expected strategy labels and using this sequence to generate a response yields better results than a direct end-to-end generation scheme. It also highlights the divergences and the limits of current evaluation metrics for generated content. The code for the annotation platform and the annotated data are made publicly available for the evaluation of future models.
• Coq Platform docs: A Compilation of Short Interactive Tutorials and How-To Guides for Coq
  
  • Lamiaux Thomas
  • Rousselin Pierre
  • Zimmermann Théo
  , 2024. We present the Coq Platform Docs project to document with interactive tutorials and how-to guides key features of Coq and its ecosystem.
• How Do Robots Become “Social Robots”? An Empirical Specification of the (Non-)Emergence of Robots as Social Agents
  
  • Rudaz Damien
  , 2024. As opposed to viewing the inner workings of human-robot interactions (HRI) as black boxes, this work investigates the finely tuned micro-interactional practices through which a robot emerges as a “social agent” in different settings. Using the micro-analytic approach of Ethnomethodological Conversation Analysis (EMCA), it examines several large corpora of encounters between humans and the humanoid robot Pepper. Their exploration allows us to broaden the list of documented interactional processes occurring during human-robot encounters (indexed to specific settings, sequential contexts, spatial configurations, etc.) by which a robot can be said to be, momentarily and locally, treated as an “agent” in a social interaction. Attending to the moment-by-moment production of the robot’s status as a practical accomplishmentleads our inquiry to a respecification of the interactional work commonly glossed by the lay use of the term “social robot”.However, rather than merely a quietist attempt at clarifying conceptual mix-ups, our approach responds to design, ergonomic, and user experience (UX) concerns regarding “social” robots. That is, by attending to the locally organized practices taking place in human-robot encounters, we attempt to provide a different type of explanation as to “what went wrong” or “what went right” in an interaction with a robot: explanations based on the features made relevant by the participants themselves as they are practically immersed within the urgency of these ongoing human-robot interactions.
• Knowledge Graphs in the Age of Language Models and Neuro-Symbolic AI : Proceedings of the 20th International Conference on Semantic Systems, 17–19 September 2024, Amsterdam, The Netherlands
  
  • Salatino Angelo
  • Alam Mehwish
  • Ongenae Femke
  • Vahdati Sahar
  • Gentile Anna-Lisa
  • Pellegrini Tassilo
  • Jiang Shufan
  , 2024, 60, pp.V-437. Semantic systems encompass a variety of technologies that play a fundamental role in our daily lives, such as artificial intelligence, machine learning, knowledge graphs and ontology engineering, and enterprise vocabulary management. This book presents the proceedings of SEMANTiCS 2024, the 20th International Conference on Semantic Systems, held from 17 to 19 September 2024 in Amsterdam, the Netherlands. The conference has become recognized as important international event providing a regular opportunity for professionals and researchers actively engaged in harnessing the power of semantic computing and artificial intelligence to gather and discuss the possibilities and practical limitations of various transformative technologies. These include semantic web and artificial intelligence, as well as areas such as data science, machine learning, logic programming, content engineering, social computing, natural language processing, digital humanities, and many more. A total of 95 submissions were received for the conference. These were subjected to a double-anonymous peer review process, with a minimum of 3 independent reviews for each submission, after which 26 papers were accepted for presentation and publication, representing an acceptance rate of 27%. The papers are divided into 6 sections: knowledge engineering with large language models; embeddings and machine learning on knowledge graphs; ontologies and knowledge graphs; linked data management; question answering and querying systems; and digital humanities and cultural heritage. Providing an overview of emerging trends and themes within the vast field of semantic computing and artificial intelligence, the book will be of interest to all those working in the field. (10.3233/SSW60)
  DOI : 10.3233/SSW60
• Security and Cryptography for Networks, Part I
  
  • Galdi Clemente
  • Phan Duong Hieu
  , 2024, 14973, pp.XX, 390. The two-volume set LNCS 14973 and 14974 constitutes the proceedings of the 14th International Conference on Security and Cryptography for Networks, SCN 2024, which took place in Amalfi, Italy, during September 11-13, 2024. The 33 full papers included in the proceedings were carefully reviewed and selected from 90 submissions. They were organized in topical sections as follows: Part I: Zero Knowledge; foundations; protocols; voting systems; Part II: Homomorphic encryption; symmetric key encryption; cryptanalysis; key management; blockchains. (10.1007/978-3-031-71070-4)
  DOI : 10.1007/978-3-031-71070-4
• Security and Cryptography for Networks, Part II
  
  • Galdi Clemente
  • Phan Duong Hieu
  , 2024, 14974, pp.XX, 362. The two-volume set LNCS 14973 and 14974 constitutes the proceedings of the 14th International Conference on Security and Cryptography for Networks, SCN 2024, which took place in Amalfi, Italy, during September 11-13, 2024. The 33 full papers included in the proceedings were carefully reviewed and selected from 90 submissions. They were organized in topical sections as follows: Part I: Zero Knowledge; foundations; protocols; voting systems; Part II: Homomorphic encryption; symmetric key encryption; cryptanalysis; key management; blockchains. (10.1007/978-3-031-71073-5)
  DOI : 10.1007/978-3-031-71073-5
• PropColor: Interactive Color Propagation for 2D Animations
  
  • Gowtham Hari Hara
  • Parakkat Amal Dev
  • Cani Marie-Paule
  , 2024. Coloring is a fundamental yet time-consuming task in the 2D animation production pipeline. Traditional methods typically rely on frame-by-frame user interaction, leading to high user time and production costs. In this paper, we introduce PropColor, an interactive yet simple tool to propagate colors between adjacent frames of a hand-drawn animation. Starting with an initial frame colored by the user, our method propagates colors to neighboring frames based on the Delaunay triangulations computed from the sketch contours and the color hints. In addition to propagating colouring between frames, our method also associates a confidence score with each of them. This enables to identify the frames where user intervention is needed the most, either to validate the result or to provide additional color hints. Experiments show that our lightweight tool gives real-time feedback and significantly cuts down the animator's time. (10.2312/vmv.20241210)
  DOI : 10.2312/vmv.20241210
• Information-theoretic generalization bounds for learning from quantum data
  
  • Caro Matthias
  • Gur Tom
  • Rouzé Cambyse
  • Subramanian Sathyawageeswar
  • Daniel Stilck França
  , 2023. Learning tasks play an increasingly prominent role in quantum information and computation. They range from fundamental problems such as state discrimination and metrology over the framework of quantum probably approximately correct (PAC) learning, to the recently proposed shadow variants of state tomography. However, the many directions of quantum learning theory have so far evolved separately. We propose a general mathematical formalism for describing quantum learning by training on classical-quantum data and then testing how well the learned hypothesis generalizes to new data. In this framework, we prove bounds on the expected generalization error of a quantum learner in terms of classical and quantum information-theoretic quantities measuring how strongly the learner's hypothesis depends on the specific data seen during training. To achieve this, we use tools from quantum optimal transport and quantum concentration inequalities to establish non-commutative versions of decoupling lemmas that underlie recent information-theoretic generalization bounds for classical machine learning. Our framework encompasses and gives intuitively accessible generalization bounds for a variety of quantum learning scenarios such as quantum state discrimination, PAC learning quantum states, quantum parameter estimation, and quantumly PAC learning classical functions. Thereby, our work lays a foundation for a unifying quantum information-theoretic perspective on quantum learning.
• Making both ends meet: from efficient simulation to universal quantum computing with quantum Gibbs sampling
  
  • Stilck Franca Daniel
  • Rouzé Cambyse
  • Alhambra Alvaro
  , 2024. The preparation of thermal states of matter is a crucial task in quantum simulation. In this work, we prove that an efficiently implementable dissipative evolution recently introduced by Chen et al. thermalizes into its equilibrium Gibbs state in time scaling polynomially with system size at high enough temperatures for any Hamiltonian that satisfies a Lieb-Robinson bound, such as local Hamiltonians on a lattice. Furthermore, we show the efficient adiabatic preparation of the associated purifications or ``thermofield double" states. To the best of our knowledge, these are the first results rigorously establishing the efficient preparation of high temperature Gibbs states and their purifications. In the low-temperature regime, we show that implementing this family of Lindbladians for inverse temperatures logarithmic in the system's size is polynomially equivalent to standard quantum computation. On a technical level, for high temperatures, our proof makes use of the mapping of the generator of the evolution into a Hamiltonian and the analysis of the stability of its gap. For low temperature, we instead perform a perturbation at zero temperature of the Laplace transform of the energy observable at fixed runtime, and resort to circuit-to-Hamiltonian mappings akin to the proof of universality of quantum adiabatic computing. Taken together, our results show that the family of Lindbladians of Chen et al. efficiently prepares a large class of quantum many-body states of interest, and have the potential to mirror the success of classical Monte Carlo methods for quantum many-body systems.
• ASML: A Scalable and Efficient AutoML Solution for Data Streams
  
  • Verma Nilesh
  • Bifet Albert
  • Pfahringer Bernhard
  • Bahri Maroua
  , 2024. Online learning poses a significant challenge to AutoML, as the best model and configuration may change depending on the data distribution. To address this challenge, we propose Automated Streaming Machine Learning (ASML), an online learning framework that automatically finds the best machine learning models and their configurations for changing data streams. It adapts to the online learning scenario by continuously exploring a large and diverse pipeline configuration space. It uses an adaptive optimisation technique that utilizes the current best design, adaptive random directed nearby search, and an ensemble of best performing pipelines. We experimented with real and synthetic drifting data streams and showed that ASML can build accurate and adaptive pipelines by constantly exploring and responding to changes. In several datasets, it outperforms existing online AutoML and state-of-the-art online learning algorithms.
• Reinforcement Learning and Sequential QAP-Based Graph Matching for Semantic Segmentation of Images
  
  • Chopin Jérémy
  • Fasquel Jean-Baptiste
  • Mouchère Harold
  • Dahyot Rozenn
  • Bloch Isabelle
  , 2024, 09, pp.259-294. This chapter addresses the fundamental task of semantic image analysis by exploiting structural information (spatial relationships between image regions). We propose to combine a deep neural network (DNN) with graph matching (formulated as a quadratic assignment problem (QAP)) where graphs encode efficiently structural information related to regions segmented by the DNN. Our novel approach solves the QAP sequentially for matching graphs, in the context of image semantic segmentation, where the optimal sequence for graph matching is conveniently defined using reinforcement learning (RL) based on the region membership probabilities produced by the DNN and their structural relationships. Our RL-based strategy for solving QAP sequentially allows us to significantly reduce the combinatorial complexity for graph matching. Two experiments are performed on two public datasets dedicated respectively to the semantic segmentation of face images and sub-cortical region of the brain. Results show that the proposed RL-based ordering performs better than using a random ordering, especially when using DNNs that have been trained on a limited number of samples. The open-source code and data are shared with the community. (10.1142/9789811289125_0011)
  DOI : 10.1142/9789811289125_0011
• The Simpler The Better: An Entropy-Based Importance Metric to Reduce Neural Networks’ Depth
  
  • Quétu Victor
  • Liao Zhu
  • Tartaglione Enzo
  , 2024, 14946, pp.92-108. While deep neural networks are highly effective at solving complex tasks, large pre-trained models are commonly employed even to solve consistently simpler downstream tasks, which do not necessarily require a large model’s complexity. Motivated by the awareness of the ever-growing AI environmental impact, we propose an efficiency strategy that leverages prior knowledge transferred by large models. Simple but effective, we propose a method relying on an Entropy-bASed Importance mEtRic (EASIER) to reduce the depth of over-parametrized deep neural networks, which alleviates their computational burden. We assess the effectiveness of our method on traditional image classification setups. Our code is available at https://github.com/VGCQ/EASIER. (10.1007/978-3-031-70365-2_6)
  DOI : 10.1007/978-3-031-70365-2_6
• Reconstructing the Unseen: GRIOT for Attributed Graph Imputation with Optimal Transport
  
  • Serrano Richard
  • Laclau Charlotte
  • Jeudy Baptiste
  • Largeron Christine
  , 2024, 14946, pp.269-286. In recent years, there has been a significant surge in machine learning techniques, particularly in the domain of deep learning, tailored for handling attributed graphs. Nevertheless, to work, these methods assume that the attribute values are fully known, which is not realistic in numerous real-world applications. This paper explores the potential of Optimal Transport (OT) to impute missing attribute values on graphs. To proceed, we design a novel multi-view OT loss function that can encompass both node feature data and the underlying topological structure of the graph by utilizing multiple graph representations. We then utilize this novel loss to train efficiently a Graph Convolutional Neural Network (GCN) architecture capable of imputing all missing values over the graph at once. We evaluate the interest of our approach with experiments both on synthetic data and real-world graphs, including different missingness mechanisms and a wide range of missing data. These experiments demonstrate that our method is competitive with the state-of-the-art in all cases and of particular interest on weakly homophilic graphs. (10.1007/978-3-031-70365-2_16)
  DOI : 10.1007/978-3-031-70365-2_16
• Code stylometry vs formatting and minification
  
  • Balla Stefano
  • Gabbrielli Maurizio
  • Zacchiroli Stefano
  PeerJ Computer Science, PeerJ, 2024, 10, pp.e2142. The automatic identification of code authors based on their programming styles—known as authorship attribution or code stylometry—has become possible in recent years thanks to improvements in machine learning-based techniques for author recognition. Once feasible at scale, code stylometry can be used for well-intended or malevolent activities, including: identifying the most expert coworker on a piece of code (if authorship information goes missing); fingerprinting open source developers to pitch them unsolicited job offers; de-anonymizing developers of illegal software to pursue them. Depending on their respective goals, stakeholders have an interest in making code stylometry either more or less effective. To inform these decisions we investigate how the accuracy of code stylometry is impacted by two common software development activities: code formatting and code minification. We perform code stylometry on Python code from the Google Code Jam dataset (59 authors) using a code2vec-based author classifier on concrete syntax tree (CST) representations of input source files. We conduct the experiment using both CSTs and ASTs (abstract syntax trees). We compare the respective classification accuracies on: (1) the original dataset, (2) the dataset formatted with Black, and (3) the dataset minified with Python Minifier. Our results show that: (1) CST-based stylometry performs better than AST-based (51.00%→68%), (2) code formatting makes a significant dent (15%) in code stylometry accuracy (68%→53%), with minification subtracting a further 3% (68%→50%). While the accuracy reduction is significant for both code formatting and minification, neither is enough to make developers non-recognizable via code stylometry. (10.7717/peerj-cs.2142)
  DOI : 10.7717/peerj-cs.2142
• Efficient learning of ground and thermal states within phases of matter
  
  • Rouzé Cambyse
  • Stilck França Daniel
  • Onorati Emilio
  • Watson James
  Nature Communications, Nature Publishing Group, 2024, 15 (1), pp.7755. We consider two related tasks: (a) estimating a parameterisation of a given Gibbs state and expectation values of Lipschitz observables on this state; (b) learning the expectation values of local observables within a thermal or quantum phase of matter. In both cases, we present sample-efficient ways to learn these properties to high precision. For the first task, we develop techniques to learn parameterisations of classes of systems, including quantum Gibbs states for classes of non-commuting Hamiltonians. We then give methods to sample-efficiently infer expectation values of extensive properties of the state, including quasi-local observables and entropies. For the second task, we exploit the locality of Hamiltonians to show that M local observables can be learned with probability 1 − δ and precision ε using N = O (log (M / δ) epolylogðε1)) samples — exponentially improving previous bounds. Our results apply to both families of ground states of Hamiltonians displaying local topological quantum order, and thermal phases of matter with exponentially decaying correlations. (10.1038/s41467-024-51439-x)
  DOI : 10.1038/s41467-024-51439-x
• Modeling, Learning, and Transferring Anatomical Representations in Medical Imaging using AI
  
  • Gori Pietro
  , 2024. Recent advances in computer vision and statistical learning, particularly deep learning, have spurred research in (anatomical) medical imaging in recent years. However, it has been reported that state-of-the-art (SOTA) algorithms from computer vision do not necessarily perform equally well when applied to natural versus medical imaging tasks. Moreover, a simple supervised pre-training using ImageNet—widely employed for natural images—does not always yield optimal results for medical imaging tasks. Given these challenges, my research has focused on developing AI methods tailored to the specific needs and constraints of medical imaging data (e.g., limited data, data biases, lack of labeled data), leveraging clinical knowledge and (healthy) unlabeled data. From a methodological standpoint, my research has followed three main directions: 1) modeling medical knowledge and anatomy, and integrating this into machine learning models; 2) learning compact, relevant, and explanatory representations of anatomical imaging data; and 3) transferring anatomical representations across domains (i.e., different modalities, datasets, and populations) to improve downstream performance or discover new biomarkers. In terms of clinical applications, I have worked exclusively with anatomical data, including MRI and CT scans, focusing on the brain, chest, abdomen, and pelvic regions. In this manuscript, I present some of my work using brain MRI data, based on the three clinical applications I have worked on most extensively. I conclude by discussing future research directions, potential novel clinical applications, and the research valorization of my previous work.
• From Model Complexity Reduction to Feature Selection in Deep Learning: a Regularization Story
  
  • Tartaglione Enzo
  , 2024. In the last decade, the scientific community has witnessed the blooming and the massive exploitation of deep neural networks (DNNs). This is fueled by multiple factors: the inherent flexibility offered by DNNs to learn input-output functions from big data and the computation scaling-up in computational capability offered by computing devices, which grasped the attention of a growing research community that progressively further improves their performance. Such a tsunamic trend finds its (momentary) apex in foundation models: DNNs trained on a dementially large quantity of data, are able to extract rich features that adapt to a broad range of downstream tasks. Siding this enthusiasm in achieving generality for these models, multiple problems were encountered, especially when dealing with the model’s usability in resource-constrained environments, or when learning from biased data. In this work, an analysis of these two aspects will be conducted, also showing how the two are intrinsically linked. This work summarizes my research conducted after the PhD, divided mainly into two parts: - In the first part, model compression and efficiency will be treated, with a special emphasis on deep neural network pruning; - In the second part, model debiasing is treated, with openings also to privacy for DNN (intended as feature hiding); All along these, new research trends currently under study are suggested: the employment of adapters for efficient fine-tuning of a large pre-trained model, training with a selection of a subset of neurons to update, on-device learning with improved selection of a DNN to fine-tune, DNN depth reduction, and more. In all the presented approaches, the recurrent underlying theme, either in an implicit or explicit fashion, will be the design of regularization for DNNs.
• Vulnerability Assessment for the Rowhammer Attack Using Hardware Performance Counters and Machine Learning
  
  • Kolić Bogdana
  • Mushtaq Maria
  , 2024. <div><p>Numerous machines using DRAM chips as main memory are vulnerable to the Rowhammer attack, which can be used as a tool for privilege escalation. The existing mitigation techniques either require complex hardware implementation or have a high performance cost. A potential improvement would be to implement a detection mechanism and trigger performance-costly mitigation only in the case of attack detection. In this paper, we study this defence method on three systems using Intel Skylake, Tiger Lake and Alder Lake processors, and DDR4 and DDR5 DRAM chips as main memory. We execute four variants of the attack code on these machines and observe their traces in the hardware. We use the PAPI library and perf to periodically read the generated traces from the machines' hardware performance counters. Finally, we train machine learning models such as logistic regression and decision trees to distinguish attack and no-attack behaviour. Our best models achieve accuracy above 99.6% and perform the classification of both 50µs and 1ms samples in software fast enough (less than 0.5µs per sample) to detect the attack before completion.</p></div>
• Material Permittivity and Conductivity Estimation from 2 to 260 GHz and Extension of the ITU-R P.2040 Model at Frequency above 100 GHz
  
  • Conrat Jean-Marc
  • Aliouane Mohamed
  • Cousin Jean-Christophe
  • Begaud Xavier
  , 2024. This paper analyzes the frequency-dependent electromagnetic characteristics of usual building materials such as mortar, glass or wood. Reflection and transmission losses are measured from 2 to 260 GHz and the related permittivity and conductivity are estimated. Results are compared with the ITU-R 2040-3 model that is mainly defined for frequency below 100 GHz. As suggested by the ITU model, the permittivity does not depend on the frequency and the conductivity increases with the frequency. However, reflection and transmission losses for frequency above 100 GHz may be strongly impacted by the material inhomogeneity or surface roughness that generates a frequency small-scale fading not predicted by the ITU model.
• Post-Quantum Cryptographically-Secured Trusted Node for Quantum Key Distribution in a Deployed Network
  
  • Huang Heming
  • Jaouën Yves
  • Fabre Nicolas
  • Alléaume Romain
  • Pegon Jean-Sébastien
  • Camus Thomas
  • Zuber Martin
  • Faugère Jean-Charles
  • Verdier Pierre-Enguerrand
  • Lacour Baptiste
  • Gautier Maxime
  • Rivera Thomas
  • Piétri Yoann
  • Schiavon Matteo
  • Rhouni Amine
  • Diamanti Eleni
  , 2024. Objective: Reduce the security risks associated with the usage of trusted nodes in a QKD network. Conclusion: The transported QKD key is secure agains honest-curious nodes at a lower key-rate cost than state-of-the-art.
• Multi-Agent Proximal Policy Optimization for Dynamic Multi-Channel URLLC Access
  
  • Robaglia Benoît-Marie
  • Coupechoux Marceau
  • Tsilimantos Dimitrios
  , 2024, pp.1-7. This work addresses the challenge of Dynamic Multi-Channel Access (DMCA) in the context of Ultra Reliable Low Latency Communications (URLLC), a framework subjected to notably stringent constraints, required by numerous Internet of Things (IoT) applications across various sectors. We introduce a theoretically grounded approach, leveraging Deep Multi-Agent Reinforcement Learning (MARL) to tackle this problem. While prior research has not fully addressed the DMCA problem in URLLC networks under time-varying heterogeneous channels and traffic profiles, nor provided robust theoretical guarantees in the multi-agent context, this paper adapts the recent theoretical framework of Trust Region Policy Optimization (TRPO) in MARL to meet the specific challenges and requirements of the URLLC-DMCA problem. Specifically, we introduce Multi Channel Access - Proximal Policy Optimization (MCA-PPO), a MARL algorithm that benefits from theoretical guarantees and effectively handles the partial observability and the combinatorial nature of the DMCA challenge. We validate the superiority of our proposed method across a variety of heterogeneous scenarios, in terms of traffic models and system parameters, and show that we outperform the traditional multiple access benchmark and learning algorithms. (10.1109/PIMRC59610.2024.10817242)
  DOI : 10.1109/PIMRC59610.2024.10817242
• Mitigating Bias in Facial Recognition Systems: Centroid Fairness Loss Optimization
  
  • Conti Jean-Rémy
  • Clémençon Stéphan
  , 2025, 15614, pp.371-385. The urging societal demand for fair AI systems has put pressure on the research community to develop predictive models that are not only globally accurate but also meet new fairness criteria, reflecting the lack of disparate mistreatment with respect to sensitive attributes (e.g. gender, ethnicity, age). In particular, the variability of the errors made by certain Facial Recognition (FR) systems across specific segments of the population compromises the deployment of the latter, and was judged unacceptable by regulatory authorities. Designing fair FR systems is a very challenging problem, mainly due to the complex and functional nature of the performance measure used in this domain (i.e. ROC curves) and because of the huge heterogeneity of the face image datasets usually available for training. In this paper, we propose a novel post-processing approach to improve the fairness of pre-trained FR models by optimizing a regression loss which acts on centroid-based scores. Beyond the computational advantages of the method, we present numerical experiments providing strong empirical evidence of the gain in fairness and of the ability to preserve global accuracy. (10.1007/978-3-031-87657-8_26)
  DOI : 10.1007/978-3-031-87657-8_26
• RIR-in-a-Box: Estimating Room Acoustics from 3D Mesh Data through Shoebox Approximation
  
  • Kelley Liam
  • Carlo Diego Di
  • Nugraha Aditya Arie
  • Fontaine Mathieu
  • Bando Yoshiaki
  • Yoshii Kazuyoshi
  , 2024. This paper describes a method for estimating the room impulse response (RIR) for a microphone and a sound source located at arbitrary positions from the 3D mesh data of the room. Simulat- ing realistic RIRs with pure physics-driven methods often fails the balance between physical consistency and computational ef- ficiency, hindering application to real-time speech processing. Alternatively, one can use MESH2IR, a fast black-box estima- tor that consists of an encoder extracting latent code from mesh data with a graph convolutional network (GCN) and a decoder generating the RIR from the latent code. Combining these two approaches, we propose a fast yet physically coherent estimator with interpretable latent code based on differentiable digital sig- nal processing (DDSP). Specifically, the encoder estimates a vir- tual shoebox room scene that acoustically approximates the real scene, accelerating physical simulation with the differentiable image-source model in the decoder. Our experiments showed that our method outperformed MESH2IR for real mesh data ob- tained with the depth scanner of Microsoft HoloLens 2, and can provide correct spatial consistency for binaural RIRs.
• Predefined Prototypes for Intra-Class Separation and Disentanglement
  
  • Mariotte Théo
  • Almudévar Antonio
  • Ortega Alfonso
  • Tahon Marie
  • Vicente Luis
  • Miguel Antonio
  • Lleida Eduardo
  , 2024, pp.3809-3813. Prototypical Learning is based on the idea that there is a point (which we call prototype) around which the embeddings of a class are clustered. It has shown promising results in scenarios with little labeled data or to design explainable models. Typically, prototypes are either defined as the average of the embeddings of a class or are designed to be trainable. In this work, we propose to predefine prototypes following human-specified criteria, which simplify the training pipeline and brings different advantages. Specifically, in this work we explore two of these advantages: increasing the inter-class separability of embeddings and disentangling embeddings with respect to different variance factors, which can translate into the possibility of having explainable predictions. Finally, we propose different experiments that help to understand our proposal and demonstrate empirically the mentioned advantages. (10.21437/Interspeech.2024-825)
  DOI : 10.21437/Interspeech.2024-825