Stage M2 Informatique à coloration recherche
Résumé de section
-
-
-Titre : Explainable Evaluation of Biometrics for Bias Discovery
Encadrants : Romain Giot (romain.giot@u-bordeaux.fr), Boyu Zhu (boyu.zhu@u-bordeaux.fr)
Résumé : La description est disponible ici.
Titre : Using Code-LLMs to enhance performance models for GPU kernels
Encadrants : Benjamin Negrevergne, Olivier Aumage (olivier.aumage@inria.fr)
Résumé :
* Context & purpose of the internship
This internship is a collaboration between the STORM Team at Inria Bordeaux France (High performance computing) and the Miles Teams (Machine Learning), PSL University, Paris, France. The internship will take place at Inria Bordeaux. Team STORM develops methodologies and tools to statically and dynamically optimize computations on High Performance Computing ( HPC) architectures, ranging from task-based parallel runtime systems to vector processing techniques, from performance-oriented scheduling to energy consumption reduction. In particular, STORM's StarPU runtime system brings task parallelism programming to heterogeneous, accelerated computing platforms. To properly schedule different tasks on the different computing units available, StarPU relies on simple performance models in order to predict the execution time of a task. Although simple, the performance of these models is inherently limited by the fact that they do not ``look'' into the code of the task to be executed. As a consequence, they are unable to generalize across distinct but similar tasks, or across various runs of the same task with similar inputs, yielding uninformed prediction for most executions. Recently Code-LLMs (LLMs that can generate code) have demonstrated impressive code generation and code understanding capabilities. Enhanced with such abilities, we hypothesize that performance models could significantly improve their quality and predict a variety of metrics useful for efficient scheduling. Such metrics include execution time but also memory footprint, bandwidth usage, or cache affinities between tasks. Ultimately, such knowledge could be used to better schedule tasks on complex heterogeneous platforms and dramatically improve the performance of the computations.
* Objectives
In a first phase of the internship, the student will design, train and evaluate a neural-network based regression models capable of predicting execution times (or other relevant metrics) for various GPU kernels. The models will be inspired from modern open source Code LLM such as Code LLama, and will make predictions based on the source code of the kernel, the input data and the hardware specification of the targeted computing unit. The focus will be on linear algebra kernels such as the ones available in the PaStiX library (https://solverstack.gitlabpages.inria.fr/pastix/index.html). This first part will require the student to develop a strong understanding of the performance and limitations of existing machine learning models when it comes to understanding the execution of programs. In a second phase, the student will investigate how these models can be used to improve the scheduling and execution of GPU kernels on complex high performance heterogeneous execution platforms. This phase will involve experiments to determine which metrics can be accurately predicted and which ones are most valuable for effective task scheduling. Given that LLM based models are also GPU intensive, it will also require evaluating the cost vs. benefit of different models of various sizes and performance.
Titre : Ordonnancement dynamique de chaînes de communication numérique pour la radio logicielle
Encadrants : Olivier Aumage (olivier.aumage@inria.fr)
Résumé :
La radio logicielle est au coeur de la prise en charge des communications numériques contemporaines telles que la téléphonie ou la transmission par satellite, elle pose un défi majeur en termes de performances et d'efficacité.
L'équipe [STORM](https://team.inria.fr/storm/), du centre Inria de l'Université de Bordeaux, développe le logiciel [AFF3CT](https://aff3ct.github.io) conjointement avec le laboratoire IMS à Bordeaux et le laboratoire LIP6 de Sorbonne Université, pour l'exécution optimisée de chaines de communication numériques dans le domaine de la radio logicielle (en anglais, SDR pour Software Defined Radio). Le but de ce domaine est de réaliser en logiciel et sur des processeurs multicoeurs généralistes l'essentiel des traitements des couches 'physique' et 'liaison de données' des communications numériques, autrefois gérés par des circuits dédiés.
AFF3CT est en particulier utilisé pour le développement, la validation et l'exploitation de codes correcteurs d'erreurs, utilisés dans les communications sans fil comme la téléphonie 5G ou les transmissions satellitaires DVB-S2, avec de forts enjeux de performances à la fois en débit et en latence, et de maitrise de la consommation énergétique. AFF3CT s'appuie sur un support d'exécution léger appelé [StreamPU](https://github.com/aff3ct/StreamPU), et dédié aux applications de type 'streaming' dominées par le flux des données. StreamPU est chargé répartir les calculs de ces chaines de communication sur des processeurs multicoeurs afin de bénéficier du parallélisme matériel. Cette répartition repose sur l'utilisation de techniques de pipeline, dont l'objectif est la constitution et le placement des étages sur l'architecture de calcul considérée pour maximiser les performances et minimiser le nombre de ressources utilisée.
L'objectif de ce stage est de faire évoluer la gamme d'ordonnanceurs proposés par StreamPU pour gérer une plus grande diversité de chaines de communications, en relâchant une partie des hypothèses faites sur la structure de ces chaines, comme permettre la présence de conditions ou de boucles ou autoriser une certaine variabilité de la durée des tâches, et au-delà, d'ordonnancer dynamiquement de multiples chaines concurrentes.
Mots-clés : parallélisme de tâches, ordonnancement, pipeline, streaming
Titre : Validation and Evaluation of a Federated Learning Optimization Framework in Fog Computing Environments
Encadrants : Hicham LAKHLEF (hicham.lakhlef@labri.fr)
Résumé : La description est disponible ici.
Titre : Decoding Spatial Ambiguity : Addressing Dendritic mRNA Localization in Spatial Transcriptomics
Encadrants : Nazim Mechkouri (nazim.mechkouri@u-bordeaux.fr), Loann Giovannangeli (loann.giovannangeli@u-bordeaux.fr)
Résumé : La description est disponible ici.
Title: Pseudo-Spectral Imaging
Advisors: Pascal Barla, Morgane Gérardin
Keywords: Computer graphics, spectral rendering, material measurement and modeling, environment lighting
Context: In Computer Graphics, there are two main classes of rendering engines. RGB rendering is the most common and is based on RGB colors emitted by light sources, reflected by objects according to their materials, and absorbed by sensors. This is clearly an approximation of physical light transport, since the properties of interest should not be colors but spectra; only at the very end of the rendering process should spectra be converted to colors. The other main class of rendering engine is thus spectral: it is not only much more costly in terms of computation and memory resources, but also makes asset capture (i.e., textures, environments) much more tedious. Yet it remains the solution of choice when it comes to produce physically-realistic (even predictive) images: it properly accounts for a variety of complex effects such as metamerism, iridescence, fluorescence, etc.
Objectives: In recent years, research has been done in our team to handle materials with complex spectral effects in RGB rendering engines, with a focus on iridescence [1,2,3]. Even more recently, we have shown that fluorescence, an optical property of materials thought to be only achievable through spectral rendering, could be closely approximated in a modified RGB rendering engine [4,5]. Now our goal is to generalize this approach to the measurement, modeling and rendering of spectral effects in a non-spectral framework, which we call Pseudo-Spectral Imaging (PSI). Depending on the interest of the candidate, two different sets of objectives may be pursued: pseudo-spectral measurements of real materials and lighting environments; or pseudo-spectral rendering using material and lighting models.
Pseudo-spectral measurements:
First, we will study to which extent pseudo-spectral material properties (reflectance, fluorescence) may be captured using a limited amount of measurements. We anticipate that as few as three color measurements under three known but arbitrary illuminants should be sufficient in the simplest case. However, more accurate results may be achieved using controlled lighting and additional measurements. Analytical material models may then be fitted to the measured material properties for editing and rendering.
Second, we will evaluate to which extent spectral environment (i.e., distant) lighting may be approximated by pseudo-spectral captures made with a consumer-grade camera of known sensitivity functions. Here again, we will investigate whether more accurate results may be achieved using a controlled setup, for instance using color charts or color filters on the camera. Known spectral lighting models (e.g., of clear or overcast skies) may also be used as apriori to refine the pseudo-spectral approximation.
Pseudo-spectral rendering:
We will start by revisiting our existing RGB iridescence material models in the context of PSI to evaluate whether their accuracy compared to spectral rendering may be improved. A particular case of interest will be accuracy in multiple scattering scenarios, where RGB approximations are prone to show the strongest approximations. We will then explore how additional effects such as metamerism or vathochromism [6] might be achieved within the PSI framework. The idea will be to let artists control the colors potentially reflected by different illuminants, or potentially transmitted through thick colored transparent objects, all without resorting to spectral representations. These developments will require implementing dedicated light transport solutions, which may be of two kinds: either based on real-time rendering (e.g., using Godot [7]) with support of live pseudo-spectral material editing; or based on off-line rendering (e.g., using Mitsuba [8]) with adapted importance sampling strategies.
Required Skills:Scientific computation in Python, Computer graphics techniques
For pseudo-spectral measurements:- radiometric measurements
- Knowledge in computer vision would be a plus
For pseudo-spectral rendering:- GLSL shader programming- knowledge in path tracing would be a plus
References:
[1] "A Practical Extension to Microfacet Theory for the Modeling of Varying Iridescence", Laurent Belcour, Pascal Barla. ACM Transactions on Graphics, 2017, 36 (4)
[2] "Interactive Exploration of Vivid Material Iridescence using Bragg Mirrors", Gary Fourneau, Romain Pacanowski, Pascal Barla. Computer Graphics Forum, 2024, 43 (2)
[3] " Efficient Modeling and Rendering of Iridescence from Cholesteric Liquid Crystals", Gary Fourneau, Pascal Barla, Romain Pacanowski. EGSR 2025 - 36th Eurographics Symposium on Rendering, Eurographics Association, Jun 2025, Copenhagen, Denmark
[4] " Non-Orthogonal Reduction for Rendering Fluorescent Materials in Non-Spectral Engines", Alban Fichet, Laurent Belcour, Pascal Barla. Computer Graphics Forum, 2024, 43 (4)
[5] "A Fluorescent Material Model for Non-Spectral Editing & Rendering", Laurent Belcour, Alban Fichet, Pascal Barla. SIGGRAPH Conference Papers 2025, Aug 2025, Vancouver BC, Canada
[6] "One‐to‐Many Spectral Upsampling of Reflectances and Transmittances", L. Belcour, Pascal Barla, Gaël Guennebaud, Computer Graphics Forum, 2023, 42 (4)
[7] https://godotengine.org/
[8] https://www.mitsuba-renderer.org/Title: Computational modeling of Alzheimer's disease effect on neural oscillations
Advisors: Amélie Aussel (amelie.aussel@inria.fr)
Keywords: Computational modeling, Neuroscience, Alzheimer's disease
Context: At the microscopic level, Alzheimer’s disease has several detrimental effects on individual neurons' dynamics and on their synaptic interactions. At the macroscopic level, Alzheimer's disease alters neural oscillations (the synchronized activity of a population of neurons at a specific frequency) within several brain regions involved in memory, such as the hippocampus. These changes in neural oscillations are most likely the cause of the cognitive deficits reported in patients. However, the way single neuron and synaptic abnormalities give rise to pathological neuronal oscillations is not well understood.
Computational models are a powerful tool to investigate this question. First, at a fundamental level, it can help disentangle the impact of the many complex disease-related neuronal changes (which co-occur in patients) on neural oscillations, and study the transition from physiological to pathological neural activity in a fully controllable and measurable environment. Second, modeling can guide experimental neuroscience and help reduce the number of (costly, time-consuming and sometimes ethically problematic) animal and human experiments needed to achieve results, and suggest or evaluate new potential therapeutic strategies.Objectives: The NeuroDTx team of Inria has developed a computational model of the hippocampus that is capable of reproducing healthy neural oscillations. In this context, the goal of this internship will be to incorporate Alzheimer's related changes into this model, in order to study how changes at this microscopic scale can lead to pathological oscillations. This study will be focused on only one type of oscillations (called theta-gamma oscillations), typically appearing in memory encoding. More precisely, the student will:
- understand the code and underlying mathematical equations of the model, then modify them according to experimental and computational literature to represent changes in ion channels and synapses dynamics
- run simulations and perform a parameter sensitivity analysis to study their impact on the power and frequency of the network oscillations.Required Skills:
Programming: Python
Mathematical concepts: differential equations, statistics.
Though curiosity about neuroscience is important, prior knowledge in it is not mandatory.Titre : Perception multi-sensorielle et suivi temporel d’objets déformables pour la robotique agricole autonome
Encadrants : Mélodie Daniel (melodie.daniel@u-bordeaux.fr)
Résumé : La description est disponible ici.
Titre : Développement d’une architecture de contrôle de la pose 2D d'un robot non-holonome par l’apprentissage par renforcement incluant la gestion de manœuvres et l’évitement d’obstacles
Encadrants : Mélodie Daniel (melodie.daniel@u-bordeaux.fr)
Résumé : La description est disponible ici.
Titre : Développement d’une architecture de contrôle en corps complet par apprentissage par renforcement pour la manipulation d’objets articulés distants
Encadrants : Mélodie Daniel (melodie.daniel@u-bordeaux.fr)
Résumé : La description est disponible ici.
Titre : Multimodal Image Segmentation for Cancer Diagnosis.Encadrants : Fabien Baldacci (fabien.baldacci@u-bordeaux.fr), Jenny Benois-Pineau (jenny.benois-pineau@u-bordeaux.fr)
Résumé : The context of this internship is a collaborative project between the LaBRI and the BRIC (Bordeaux Institute of Oncology) laboratories. The overall goal of the project is to define new tools related to the understanding of pancreatic cancer, based on XAI, in order to improve prognosis.
Diagnosis and prognosis can rely on different microscopic modalities, and there is actually no automatic method for cross-modalities segmentation of cancer lesions, which is mandatory for reliable prognosis. During this internship, we will focus on two widely used images for tissue analysis: mass spectrometry imaging (MSI) providing precise molecular information, and haematoxylin and eosin-stained imaging (H&E) for morphological information.
During this internship, the student will have to study the state-of-the-art segmentation models for both modalities and propose a solution to create a multimodal fusion model combining all the features. Later on, depending on the remaining time, the XAI methods developed in the laboratory may be used to assess how the proposed network works to enhance the quantitative and qualitative study of the results.
Titre : Modélisation dynamique des conséquences socio-environnementales du déploiement à large échelle de technologies numériques
Encadrants : Aurélie Bugeau (aurelie.bugeau@labri.fr), Gaël Guennebaud (gael.guennebaud@inria.fr)
Résumé : La description est disponible ici.
Titre : Modélisation basée données de scénarios de transition du système agricole français pour en estimer les impacts environnementaux
Encadrants : Gaël Guennebaud (gael.guennebaud@inria.fr), Olivier Gauwin (olivier.gauwin@labri.fr)
Résumé : La description est disponible ici.
Titre: Comment les agents de programmation sont-ils utilisés?Encadrants: Romain Robbes (romain.robbes@u-bordeaux.fr), Thomas Degueule (thomas.degueule@labri.fr)Résumé: Les agents de programmation (e.g. Claude Code, Codex, Cursor) sont des outils utilisant les LLMs pour accomplir des tâches de programmation telles que la résolution de bugs ou l’implémentation de nouvelles fonctionnalités. Popularisés en 2025, ces agents sont désormais largement adoptés (dans une étude, nous estimons que 15-20% des projets sur GitHub utilisent des agents).Notre étude sur l’adoption des agents est un exemple de “Mining Software Repositories”, ou nous utilisons les données sur le dépôt GitHub afin d’étudier comment les logiciels sont développés. Elle porte sur plus de 100,000 dépôts GitHub, parmi lesquels des milliers ont été identifiés comme utilisateurs d’agents de programmation. Certains les utilisent de façon expérimentale, d’autres génèrent la quasi-totalité de leur code via des agents!L’objectif de ce stage est d’étudier l’utilisation des agents en profondeur. En partant d’une liste de dépôts utilisateur d’agents, il faudra analyser ces dépôts (fichiers, commits, pull requests, bugs) pour extraire des informations utiles sur les changements occasionnés dans le développement logiciel par les agents. Des questions possibles sont:- Quelles sont les tâches faites par les agents, quelles sont celles faites par les humains?- Le code crée par les agents est-il de moins bonne, ou de meilleure qualité?- Quelles sont les caractéristiques des dépôts qui font que les agents marchent mieux, ou moins biens?- Quel est l’impact des fichiers d’instructions (les “prompts”) donnés aux agents?- La durée de vie du code généré par les agents est-elle plus courte?Les questions précises à adresser seront déterminées au début du stage. Le sujet est vaste, potentiellement plusieurs personnes peuvent travailler sur des questions différentes.Mots clés: Coding Agents, Étude Empirique, Mining Software RepositoriesTitre: Repenser la révision de code pour la programmation orientée agentEncadrants: Romain Robbes (romain.robbes@u-bordeaux.fr), Thomas Degueule (thomas.degueule@labri.fr)Résumé: La révision de code est une phase clé du développement logiciel pour les logiciels matures. Chaque changement qu’un développeur propose est soumis au projet (par exemple par une “Pull Request” sur GitHub), afin que d’autres développeurs puissent commenter les changements, avant de décider si les intégrer, si demander des changements supplémentaires (e.g. fixer des bugs, améliorer la qualité), ou bien refuser leur intégration.Les agents de programmation (e.g. Claude Code, Codex, Cursor) sont des outils utilisant les LLMs pour accomplir des tâches de programmation telles que la résolution de bugs ou l’implémentation de nouvelles fonctionnalités. Popularisés en 2025, ces agents sont désormais largement adoptés (dans une étude, nous estimons que 15-20% des projets sur GitHub utilisent des agents). Leur capacités sont impressionnantes, par exemple, le projet “sublayer_actions” a trouvé une façon complètement automatique de générer de nouvelles actions, et 3 agents codent, chaque nuit, des nouvelles actions pour le projet.Si bien les agents de programmation orientée agent peuvent automatiser un nombre croissant de tâches, cela a des conséquences. Notamment, le processus de révision de code peut s’alourdir car les contributions des agents peuvent être nombreuses et volumineuses. Si bien des agents peut contribuer des actions à “sublayer_actions” très facilement, le goulot d’étranglement est le processus de révision: il y a désormais plus de 1000 “pull requests” qui attendent d’être relues dans ce projet: https://github.com/sublayerapp/sublayer_actions/pullsL’objectif de ce stage est de travailler sur de meilleures interfaces pour permettre de rendre le processus de révision de code plus efficace, en prévision d’une forte augmentation des contributions générées par les agents, qui peuvent de plus être volumineuses. Dans ces cas, une interface purement textuelle, sans indication d’importance des changements dans de multiples fichiers, et sans indications de leur relations, est loin d’être optimale. Le projet explorera des façons alternatives de présenter les informations pertinentes pour la révision de code, par exemple en changeant leur ordre, en mettant en valeurs les éléments importants, ou autres.Mots clés: Coding Agents, Révision de code, Interfaces Homme MachineTitre : Leveraging Local Build Artifacts to Optimize Continuous Integration Pipelines
Encadrants : Jean-Rémy Falleri (jean-remy.falleri@labri.fr), Thomas Degueule (thomas.degueule@labri.fr)
Mots-clés : continuous integration, DevOps, build systems, energy consumption
Résumé :
Context
Continuous Integration and Continuous Deployment (CI/CD) have become key components of modern software development. They enable development teams to automate building, testing, and deployment processes, ensuring that code changes are continuously validated and integrated. This approach improves software reliability, shortens release cycles, and facilitates collaboration between developers and operations teams. As of 2021, more than half of all GitHub repositories use some form of CI system, which highlights how deeply these practices are embedded in modern development workflows [1]. However, this automation comes with tangible financial and infrastructure costs. CI servers often re-build, re-analyze, and re-test entire projects from scratch, duplicating work already performed by developers locally. The resulting computational overhead not only slows down feedback loops but also increases cloud expenses and energy consumption at scale.
As projects grow and accumulate complexity, these repeated computations become a significant bottleneck. Every commit can trigger a full cycle of compilation, linting, testing, and packaging, although some of these processes may have already been executed locally by developers. This redundancy highlights an opportunity for optimization: reducing waste by reusing results from local developer activity while maintaining the reliability and reproducibility guarantees central to CI/CD.
Objectives
This internship aims to investigate ways to reduce the computational, financial, and energy costs of CI pipelines by reusing the results of operations already carried out on developers’ machines. Developers commonly perform local builds, run linters and static analysis tools, execute unit tests, and generate intermediate artifacts before submitting code. If these artifacts could be securely shared with and reused by the CI system, a large portion of redundant work could be avoided. Achieving this, however, requires addressing challenges in verifying the correctness and trustworthiness of externally provided artifacts, managing differences between environments, and ensuring that the reuse mechanism itself remains efficient and scalable.
The research will explore strategies for artifact certification, reproducible build verification, and distributed caching mechanisms. It will also analyze existing technologies such as Gradle and Bazel, which already support shared build caches, to identify their strengths and limitations. Particular attention will be paid to the integration of locally produced outputs (e.g., static analysis reports, linting results, test outputs) into CI workflows while preserving traceability and build integrity. The expected outcome is a prototype or proof of concept demonstrating how CI systems can safely leverage local developer computations to reduce build times, lower infrastructure costs, and minimize energy consumption without compromising reliability.
[1] Golzadeh, Mehdi, Alexandre Decan, and Tom Mens. "On the rise and fall of CI services in GitHub." IEEE International Conference on Software Analysis, Evolution and Reengineering (SANER). IEEE, 2022.
[2] Perez, Quentin, et al. "Software frugality in an accelerating world: the case of continuous integration." arXiv preprint arXiv:2410.15816 (2024).Titre : Dimensionality Reduction Techniques for Human Gut Microbiome data
Encadrants : Guilhem Sommeria-Klein (guilhem.sommeria-klein@inria.fr), Clémence Frioux (clemence.frioux@inria.fr)
Résumé : La description est disponible ici.
Titre : Augmented-Reality Data Visualizations for Presenting Environmental Data
Encadrants : Leni Yang (leni.yang@inria.fr)
Résumé : La description est disponible ici.
Titre : Towards a complete processing chain of 6D pose estimation for assistance of upper limb neuro-prostheses control
Encadrants : Jenny Benois-Pineau (Labri, U Bordeaux), Renaud Péteri (MIA, U La Rochelle)
Résumé : La description est disponible ici.
Titre : Curiosity-driven automated discovery of interference patterns in multi-core architectures
Encadrants : Ludovic Matar (ludovic.matar@inria.fr), Clement Moulin-Frier (clement.moulin-frier@inria.fr)
Résumé : La description est disponible ici.
-