Jérémie Sublime

Contributions au clustering collaboratif et à ses potentielles applications en imagerie à très haute résolution: Cette thèse présente plusieurs algorithmes développés dans le cadre du projet ANR COCLICO et contient deux axes principaux :Le premier axe concerne l'introduction d'un algorithme applicable aux images satellite à très haute résolution, qui est basé sur les champs aléatoires de Markov et qui apporte des notions sémantiques sur les clusters découverts. Cet algorithme est inspiré de l'algorithme Iterated conditional modes (ICM) et permet de faire un clustering sur des segments d'images pré-traitées. La méthode que nous proposons permet de gérer des voisinages irréguliers entre segments et d'obtenir des informations sémantiques de bas niveau sur les clusters de l'image traitée.Le second axe porte sur le développement de méthodes de clustering collaboratif applicables à autant d'algorithmes que possible, ce qui inclut les algorithmes du premier axe. La caractéristique principale des méthodes proposées dans cette thèse est leur applicabilité aux deux cas suivants : 1) plusieurs algorithmes travaillant sur les mêmes objets dans des espaces de représentation différents, 2) plusieurs algorithmes travaillant sur des données différentes ayant des distributions similaires. Les méthodes que nous proposons peuvent s'appliquer à de nombreux algorithmes comme l'ICM, les K-Moyennes, l'algorithme EM, ou les cartes topographiques (SOM et GTM). Contrairement aux méthodes précédemment proposées, notre modèle permet à des algorithmes très différents de collaborer ensemble, n'impose pas de contrainte sur le nombre de clusters recherchés et a une base mathématique solide.

The rapid growth of deep learning has led to major successes in audio classification, but the “black box” nature of these models slows down their use in important areas like healthcare where trust is critical. This paper addresses this problem through a detailed study on the use of Explainable AI (XAI) for audio-based systems. Six distinct datasets are used to evaluate a consistent deep learning model, including speech emotion recognition, environmental sound classification, and healthcare (heart, lung, and cough sounds). The study analyzes both audio-only models and multimodal models that combine audio features (spectrograms and MFCCs) with demographic data. A set of three distinct XAI techniques, namely LIME, SHAP, and Grad-CAM, is used to explain how the models make their predictions. Our key findings show that XAI methods are very useful for checking how the models work, confirming that they learn to focus on clinically and acoustically relevant features in the audio signals. The analysis also shows the power of multimodal explanations, where tools like SHAP and LIME can trace a prediction back to both acoustic patterns and specific demographic data, showing how the model creates a complete, patient-specific picture. The main contribution of this work is a broad, comparative study that provides a more complete understanding of explainability in the audio domain. The results confirm that XAI is an important part of the entire process of building a model, helping with validation, debugging, and the creation of more reliable and human-aligned AI systems for real-world use.

In this work, we introduce m-FDPC, a mass-based variant of the Fast Density Peak Clustering (FDPC) algorithm, aimed at improving both performance and ease of use in unsupervised learning tasks. Traditional FDPC relies on Euclidean distance and requires careful parameter tuning and data normalization, which can significantly affect clustering outcomes—especially for heterogeneous or high-dimensional datasets. To address these challenges, m-FDPC replaces the conventional Euclidean metric with a mass-based distance measure derived from isolation forests, a method originally designed for anomaly detection. This substitution allows the algorithm to capture local data density and structure more naturally, while eliminating the need for normalization and simplifying the choice of key parameters such as cutoff distance and density thresholds. Comprehensive experiments on synthetic and real-world datasets demonstrate that m-FDPC not only matches or surpasses the performance of well-established clustering techniques such as DBSCAN, K-means, and Euclidean FDPC, but also offers greater robustness, scalability, and interpretability, particularly in high-dimensional or unevenly distributed data scenarios. Results evaluated through metrics like Matching Score and Silhouette Score confirm the algorithm’s superior ability to detect meaningful cluster structures with minimal user intervention. Overall, m-FDPC provides a more efficient, adaptive, and user-friendly framework for density-based clustering, making it a promising tool for diverse applications in data mining, anomaly detection, and exploratory data analysis.

The rapid adoption of Generative AI (GenAI) based on Large Language Models (LLMs) such as ChatGPT has recently and profoundly impacted education, offering transformative opportunities while raising significant concerns. In this study we present the results of a survey that investigates how 395 students aged 13 to 25 years old in France and Italy integrate LLMs into their educational routines. Key findings include the widespread use of these tools across all age groups and disciplines, with older students and male students demonstrating higher usage frequencies, particularly in scientific contexts. The results also show gender disparities, raising concerns about an emerging AI literacy and technological gender gap. Additionally, while most students utilise LLMs constructively, the lack of systematic proofreading and critical evaluation among younger users suggests potential risks to cognitive skills development, including critical thinking and foundational knowledge. The survey results underscore the need for educational institutions to adapt their curricula to integrate AI tools effectively, promoting ethical use, critical thinking, and awareness of AI limitations and environmental costs. This paper provides actionable recommendations for fostering equitable and effective cohabitation of LLMs and education while addressing emerging challenges.

In today's world, AI programs powered by Machine Learning are ubiquitous, and have achieved seemingly exceptional performance across a broad range of tasks, from medical diagnosis and credit rating in banking, to theft detection via video analysis, and even predicting political or sexual orientation from facial images. These predominantly deep learning methods excel due to their extraordinary capacity to process vast amounts of complex data to extract complex correlations and relationship from different levels of features. In this paper, we contend that the designers and final users of these ML methods have forgotten a fundamental lesson from statistics: correlation does not imply causation. Not only do most state-of-the-art methods neglect this crucial principle, but by doing so they often produce nonsensical or flawed causal models, akin to social astrology or physiognomy. Consequently, we argue that current efforts to make AI models more ethical by merely reducing biases in the training data are insufficient. Through examples, we will demonstrate that the potential for harm posed by these methods can only be mitigated by a complete rethinking of their core models, improved quality assessment metrics and policies, and by maintaining humans oversight throughout the process.

The ability to build more robust clustering from many clustering models with different solutions is relevant in scenarios with privacy-preserving constraints, where data features have a different nature or where these features are not available in a single computation unit. Additionally, with the booming number of multi-view data, but also of clustering algorithms capable of producing a wide variety of representations for the same objects, merging clustering partitions to achieve a single clustering result has become a complex problem with numerous applications. To tackle this problem, we propose a clustering fusion algorithm that takes existing clustering partitions acquired from multiple vector space models, sources, or views, and merges them into a single partition. Our merging method relies on an information theory model based on Kolmogorov complexity that was originally proposed for unsupervised multi-view learning. Our proposed algorithm features a stable merging process and shows competitive results over several real and artificial datasets in comparison with other state-of-the-art methods that have similar goals.

This work proposes a generalized learning architecture for vehicular channel estimation using the Ensemble Learning (EL) technique applied to a method based on the long-term memory network (LSTM). The challenge of estimating wireless vehicular channels is addressed, where few methods explore generalizing models for variations in wireless channel models. The proposed approach is robust to changes in Dopplerdelay characteristics across different environments and channel models, resulting in an estimator that can work under varying conditions without added online complexity. The results show the possibility of achieving a generalized model with improved performance compared to specific channel condition models.

Vehicular communications are an important focus of studies for 5G applications and beyond. However, in a scenario with doubly-selective and highly variable channel characteristics, tracking the wireless channel to ensure communication reliability is one of the main goals to provide communication efficiency. Moreover, multicarrier modulation schemes usually employed in these scenarios are susceptible to nonlinear distortions caused by high power amplifiers (HPA) at the transmitter, impairing the channel estimation and detection capability of the receivers. In view of these requirements and challenges, in the present work we propose a low complexity estimator based on the long short-term memory (LSTM) network, followed by a neural network (NN) in order to improve the data-pilot aided (DPA) estimation. In addition, we propose a new technique to exploit the characteristics of the vehicular channel, by sampling the subcarriers used at the input of the LSTM. Thus, besides tracking the variations of the wireless channel, the LSTM network is also used to interpolate the channel estimates for all subcarriers. The simulation results show the superiority of the proposed scheme in comparison with other state-of-the-art schemes, especially in high signal-to-noise ratio (SNR) regimes. Furthermore, the proposed scheme significantly reduces the computational complexity due to the subcarrier sampling procedure.