Offre en lien avec l’Action/le Réseau : – — –/– — –

Laboratoire/Entreprise : Laboratoire IRISA
Durée : 5-6 mois
Contact : tristan.allard@irisa.fr
Date limite de publication : 2026-03-01

Contexte :
Health data, social networks, electricity consumption… Vast quantities of personal data are collected today by private companies or public organizations. Various legal, monetary, or visibility incentives push data holders to envision sharing versions of the collected datasets that provide both statistical utility and privacy guarantees. Indeed, sharing data at large, e.g., as open data, without jeopardizing privacy, is expected to bring strong benefits (strengthening, e.g., scientific studies, innovation, public policies). Synthetic data generation is a promising approach. First, synthetic data generation algorithms aim at generating datasets that are as close as possible to the original datasets. Either synthetically generated data or the generative models trained over the original data could be shared for supporting elaborate data analysis. Second, substantial progress has been made during the last decade about the privacy guarantees of synthetic data generation algorithms. For example, there exist today synthetic data generation algorithms that satisfy variants of differential privacy, one of the most prominent family of privacy models. However, the wealth of generative algorithms, of privacy models and algorithms, and of parameters makes it hard for non expert users to understand clearly the privacy implications of any given choice. Given the growing number of privacy attacks on machine learning models and especially on generative algorithms, an inappropriate choice can result in catastrophic consequences.

Sujet :
The main goal of this M2 thesis is to design an efficient approach for allowing a data holder to compute the most relevant privacy attacks given the data holder’s choice.

The main tasks of the Master student will be to:
• Study the state-of-the-art about privacy attacks (e.g., membership inference attacks [2, 4, 5]). We will focus on tabular data.
• Formalize the attackers (e.g., adversarial goals, background knowledge, impacts and costs of the attacks, vulnerable algorithms), structure the space of attackers (e.g., generalization/specialization of attackers, implications), and explore efficiently the resulting space for finding the attacks that best illustrate the privacy risks.
• Implement the approach and evaluate its performance.

In addition to the core tasks of the project, the successful candidate will also contribute to the organisation of competitions where the privacy guarantees of synthetic data generation algorithms are challenged.

Profil du candidat :
• The candidate must be in the second year of a master’s degree, or equivalent, in computer science or in a related field.
• The candidate must be curious, autonomous, and rigorous.
• The candidate must be able to communicate in English (oral and written). The knowledge of the French language is not required.
• The candidate must have a strong interest in cybersecurity.
• Skills in machine learning will be appreciated.

Formation et compétences requises :

Adresse d’emploi :
Campus de Beaulieu IRISA/Inria Rennes
263 avenue du Général Leclerc
35042 RENNES cedex

Document attaché : 202511171626_m2-attacks-25_26.pdf

Offre en lien avec l’Action/le Réseau : – — –/– — –

Laboratoire/Entreprise : Centre Génie Industriel, IMT Mines Albi
Durée : 6 mois
Contact : myriam.lamolle@mines-albi.fr
Date limite de publication : 2025-12-30

Contexte :
Ce stage s’inscrit dans le projet TABGHA qui se veut un outil de génération démultiplié de concepts hautement valorisables afin de promouvoir l’innovation avec l’aide du CGI (Centre de Génie Industriel) d’IMT Mines Albi (cgi.mines-albi.fr) et le LIASD, Université Paris 8.

Sujet :
Face à la méconnaissance des cultures innovations susceptibles de créer des ressources et le peu d’étude sur la valorisation économique des recherches (spin off, licence, prestation), le transfert de technologie devient crucial.
Pour cela, il faut, à partir d’une technologie générique (souvent sur étagère), multiplier les cas applicatifs (marchés d’atterrissage) pour accroitre les chances de valorisation (et aussi poursuivre les développements). À ce jour, de nombreuses méthodes existent pour générer ces différents cas applicatifs : design thinking, analyse fonctionnelle… Aucune d’entre elles ne faisant l’unanimité, le projet vise à développer une plateforme open-source en « Knowledge As A Service » (KaaS) comme écosystème digital fondé sur des composants logiciels qui se combinent pour collaborer de façon à permettre une évolution graduelle du système au travers de nouvelles contributions et de nouveaux composants fournis par la communauté.
L’objectif sous-jacent est de découvrir dans la base de connaissance constituée des connaissances cachées et de déterminer des analogies afin d’aider des décideurs à trouver de nouveaux domaines d’application de leur concept/produit ou à créer de nouveaux concept/produit.
Vous contribuerez à lever un des verrous scientifiques du projet : concevoir un ou des algorithmes d’alignement de méthodes de design thinking fondés sur l’analyse structurelle et sémantique de graphes de connaissances (ou ontologie modulaire).
Objectifs concrets :
1. Modéliser sous forme d’une ontologie les méthodes C-K, Vianeo (voire TriZ1) ;
2. En suivant une méthode scientifique, concevoir et implémenter des algorithmes permettant d’aligner semi-automatiquement les méthodes citées : les classes, les relations entre ces classes, etc. en explicitant les règles de transformation. Peupler la base de connaissance et évaluer les algorithmes sur le mini-cas d’étude réaliste ;
3. Créer les IHM pour offrir les services :
a) de choix de la méthode par laquelle le projet d’innovation sera développé ;
b) ou d’hybridation de méthodes pour une approche multi-points de vue ;
4. Rédiger un article de conférence internationale avec l’équipe d’encadrement.

Profil du candidat :
Bac+5 (Master 2 ou dernière année d’école d’ingénieur) en Informatique, Mathématiques Appliquées ou domaine connexe

Formation et compétences requises :
Compétences techniques indispensables :
• Maîtrise des concepts de l’IA neuro-symbolique ;
• Capaciter à conceptualiser des méthodes complexes
• Maîtrise d’outils et méthodes d’analyse de graphes
• Excellente compréhension écrite, bonne expression (lecture/rédaction d’articles) et bonne compréhension orale en anglais
Qualités personnelles attendues
• Capacité à proposer des idées algorithmiques dès l’entretien
• Être moteur dans les échanges scientifiques et la validation d’idées
• Goût pour la formalisation mathématique et l’expérimentation méthodique
• Appréciation pour la recherche bibliographique
Un plus :
• Expérience en recherche (stage labo, projet R&D…)
• Connaissance des ontologies et des logiques de description
• Intérêt pour les systèmes à base de raisonnement

Adresse d’emploi :
IMT Mines Albi (Centre de génie industriel), Albi (81)

Document attaché : 202511171603_offreStageM2_sujet1_2026.pdf

Offre en lien avec l’Action/le Réseau : – — –/– — –

Laboratoire/Entreprise : ICube
Durée : 5-6 mois
Contact : florence.leber@engees.unistra.fr
Date limite de publication : 2026-01-15

Contexte :
The restoration or naturalization of hydro-ecosystems is a major challenge for the coming years in order to protect and preserve the quality and quantity of river water. Many restoration works – both recent and historical – have generated large amounts of textual documentation (reports, archival documents, project plans, regulations, scientific articles) and visual material (maps, drawings, aerial/satellite imagery, photographs, cross-sectional charts). However, that material is often unstructured, scattered across institutions, in multiple languages, and not organized to support comparative analysis, learning, or decision-making effectively.

Sujet :
The main research task involves applying and refining VLMs to extract complementary information from visual and textual data. The VLMs should recognize and describe restoration structures, spatial configurations, and temporal stages (before, during, and after restoration) from images. They should extract objectives, methodologies, outcomes, and environmental parameters from text. A key scientific challenge lies in the multi-modal alignment of information linking visual elements and textual references to produce consistent and interpretable outcomes.

Building on these results, the internship will contribute to the enrichment of an already existing structured knowledge model (ontology), describing restoration cases through key properties including intervention type, environmental context, methods, results, constraints, and costs. In addition to enriching the knowledge model, another key point is populating the knowledge model by constructing knowledge graphs with information extracted from images and text, ensuring querying, comparison, and visualization by researchers and practitioners.

Profil du candidat :
Knowledge on data science methods, knowledge representation and reasoning, knowledge graphs.
Languages: Python, java, owl/sparql.
Interest in the application domain, ability to work with experts who are not computer scientists would be appreciated.

Formation et compétences requises :
Student about to graduate a Master or Engineer (Bac + 5) with a specialization in Computer Science.

Adresse d’emploi :
ICube — 300 bd Sébastien Brant – CS 10413 – F-67412 Illkirch Cedex
Meetings at ENGEES, 1 cour des cigarières, Strasbourg.

Document attaché : 202511141617_Sujet_stage_TETRA_VLM.pdf

Offre en lien avec l’Action/le Réseau : – — –/– — –

Laboratoire/Entreprise : Laboratoire Oceanographique de Villefranche-sur-me
Durée : 6 mois
Contact : enza.labourdette@imev-mer.fr
Date limite de publication : 2026-01-15

Contexte :
Le phytoplancton regroupe les micro-organismes marins dérivant au gré des courants et capables de
produire de la matière organique grâce à la photosynthèse.
Il joue un rôle essentiel dans le fonctionnement du système Terre : il contribue à environ la moitié de la
photosynthèse planétaire, participe au stockage du carbone à l’échelle climatique via la pompe
biologique, et constitue la base des réseaux trophiques marins.
Mieux comprendre la répartition et la diversité du phytoplancton, en lien avec les conditions physiques
et biogéochimiques de son environnement, est donc fondamental. Cela permet d’estimer plus finement
sa biomasse, son rôle dans les cycles biogéochimiques et d’anticiper les réorganisations induites par les
changements climatiques et océaniques globaux.
Cependant, l’observation directe du phytoplancton à grande échelle demeure limitée, car elle repose sur
des campagnes en mer et des prélèvements coûteux. Les données HPLC (High Performance Liquid
Chromatography), qui permettent d’analyser les pigments photosynthétiques pour quantifier
précisément les communautés phytoplanctoniques, offrent une référence fiable, mais restent
coûteuses et spatialement limitées.
Pour étendre ces observations ponctuelles à l’échelle globale, les satellites d’observation de la couleur
de l’océan constituent un outil clé. Les algorithmes empiriques actuels permettent d’estimer la
composition du phytoplancton à partir des réflectances optiques , mais ils présentent encore
une marge d’erreur importante et ne capturent pas toujours la complexité spatiale et temporelle des
structures océaniques (fronts, tourbillons, etc.).

Sujet :
Ce stage s’inscrit dans la continuité d’une thèse centrée sur la prédiction des Phytoplankton Functional
Types (PFTs) à partir de la couleur de l’eau et de variables satellitaires à l’aide de méthodes de deep
learning.
L’objectif principal est d’améliorer la robustesse et la précision des prédictions en intégrant explicitement
la dimension temporelle des données au moyen de méthodes avancées de deep learning.
Concrètement, cela consiste à exploiter les images satellites des jours ou des semaines précédentes pour
estimer la quantité et la composition du phytoplancton à un instant t donné. En effet, la croissance du
phytoplancton est un processus dynamique, et son état actuel dépend fortement des états antérieurs du
système.
Un pipeline opérationnel est déjà en place :
– les données satellitaires et in situ ont été extraites et prétraitées,
– des modèles de type MLP (Multi-Layer Perceptron) ont été développés et évalués,
– une première étude de la sensibilité, en termes de qualité des prédictions, aux variables d’entrée
et à l’architecture du modèle a été menée,
– une seconde étude, sur l’intégration de la dimension spatiale dans les données d’entrée, est en
cours.
Le stage consistera à poursuivre et à approfondir ces travaux, notamment en :
– explorant de nouvelles architectures de deep learning (par ex. RNN) permettant de mieux
capturer la structure temporelle des données ;
– évaluant l’impact de cette intégration sur la qualité des prédictions de PFTs ;
– affinant l’interprétation des modèles pour mieux comprendre les liens entre conditions
environnementales et composition des communautés phytoplanctoniques.
Ce travail contribuera directement à l’amélioration des produits satellitaires de distribution
phytoplanctonique à l’échelle globale, utiles pour le suivi de la santé des écosystèmes marins et la
gestion durable des ressources océaniques.

Objectifs du Stage :
– Compréhension de la problématique biologique.
– Étude des relations entre les propriétés optiques de l’eau et les PFTs.
– Prise en main et optimisation du pipeline existant, de la sélection et du traitement des données
jusqu’à leur exploitation dans un modèle de deep learning.
– Analyse de la variabilité spatio-temporelle : quantifier l’impact de l’intégration des dimensions
spatiale et temporelle sur la prédiction des PFTs.
– Mise en forme et valorisation des résultats pour une utilisation dans des travaux de recherche à
publier.

Profil du candidat :

Formation et compétences requises :
– Intérêt marqué pour la recherche et les sciences marines.
– Maîtrise du langage Python.
– Connaissances en deep learning (PyTorch).
– Des notions en écologie, biologie marine ou océanographie seraient un atout.

Adresse d’emploi :
181 chemin du Lazaret Villefranche-sur-Mer, France

Document attaché : 202511140931_STAGE 2026 – ANNONCE.pdf

Offre en lien avec l’Action/le Réseau : RECAST/– — –

Laboratoire/Entreprise : LIMOS, UMR 6158 / Mines Saint-Étienne
Durée : 4-6 mois
Contact : maxime.lefrancois@emse.fr
Date limite de publication : 2025-12-15

Contexte :
Physical quantities form an important part of what is represented in scientific data, medical data, industry data, open data, and to some extent, various private data.

Whether it is distances, speeds, payloads in transportation, concentrations, masses, moles in chemistry, powers, intensities, voltages in the energy sector, dimensions of furniture, weights, heights of people, durations, and many others in health, there is a need to represent physical quantities, to store them, to process them, and to exchange them between information systems, potentially on a global scale, often on the Internet and via the Web.

Sujet :
In this internship, we seek to precisely define a way to unambiguously represent physical quantities for the Web of Data. More precisely, we will study the proposals made to encode physical quantities in the standard data model of the Semantic Web, RDF. We will be particularly interested in the use of a data type dedicated to this encoding, probably adapted from the proposal of Lefrançois & Zimmermann (2018) based on the UCUM standard.

Having established a rigorous definition of the data type (possibly its variants, if relevant), we will focus on implementing a module that can read/write and process physical quantities and their operations within the RDF data manipulation APIs, for the management, querying and reasoning with knowledge graphs containing physical quantities.

The ambition is that, on the one hand, the specification will become in a few years a de facto standard, before perhaps becoming a de jure standard; and that, on the other hand, the implementation will be the reference allowing to compare the compliance levels of other future implementations.

This study should lead to the publication of a scientific paper in a high impact scientific journal.

References
Maxime Lefrançois and Antoine Zimmermann (2018). The Unified Code for Units of Measure in RDF: cdt:ucum and other UCUM Datatypes. In The Semantic Web: ESWC 2018 Satellite Events – ESWC 2018 Satellite Events, Heraklion, Crete, Greece, June 3-7, 2018, Revised Selected Papers, volume 11155 of the Lecture Notes in Computer Science, pp196–201, Springer.
Gunther Shadow and Clement J. McDonald. The Unified Code for Units of Measure. Technical report, Regenstrief Institute, Inc, November 21 2017.

Profil du candidat :
Master 2 students in computer science

To apply, please submit by email or in an online file repository your CV, motivation letter, university transcripts, and possibly letters of recommendation. The motivation letter must explain why you are interested in this topic and why you are qualified to work on this topic.

Formation et compétences requises :
Equivalent of a M2 level in CS, with knowledge of Semantic Web technologies. Also, the candidate must have either very good programming skills in Java, or very good aptitude in formal and abstract thinking.

Adresse d’emploi :
Mines Saint-Étienne, Institut Henri Fayol, 29 rue Pierre et Dominique Ponchardier, 42100 Saint-Étienne, France

Offre en lien avec l’Action/le Réseau : – — –/– — –

Laboratoire/Entreprise : LIP6, Sorbonne University
Durée : 5-6 months
Contact : rafael.angarita@lip6.fr
Date limite de publication : 2026-04-30

Contexte :

Sujet :
Participatory democracy platforms (Make, Decidim, Cap Collectif, Consul) enable thousands of citizens to propose and discuss ideas for public policies. However, the large volume of textual contributions produces severe information overload: citizens struggle to identify similar or opposing proposals, while decision-makers face difficulty in detecting consensus or disagreement.

Recent research at LIP6 has shown that Natural Language Processing (NLP) can detect argumentative relations between citizen proposals (equivalence, contradiction, neutrality). These relations can be structured into argumentative graphs, which help organize debates and improve navigation within large participatory datasets.

This internship aims to extend these ideas using Graph Retrieval-Augmented Generation (Graph-RAG). By combining graph-based retrieval with language generation, the project seeks to build intelligent tools capable of summarizing debates, identifying conflicting or redundant proposals, and assisting citizens in writing balanced contributions.

Profil du candidat :
Master 2 / Final-year engineering

Formation et compétences requises :
– Programming: Python, PyTorch or TensorFlow

– NLP / ML: Experience with large language models, embeddings, or NLP tasks

– Data Science: Text preprocessing, vector representations, evaluation metrics

– Research: Ability to conduct literature reviews, design small experiments, and analyze results

– Participatory democracy: Interest in participatory democracy or computational argumentation

Adresse d’emploi :
Sorbonne University, 4 place Jussieu 75005 Paris.

Document attaché : 202511121059_Stage_LIP6_2025_2026.pdf

Offre en lien avec l’Action/le Réseau : – — –/– — –

Laboratoire/Entreprise : LIPADE
Durée : 6 months
Contact : ayoub.karine@u-paris.fr
Date limite de publication : 2026-04-30

Contexte :
The performance of supervised deep learning methods in computer vision heavily depends on the availability of
labeled data, whose annotation is both time-consuming and requires expert knowledge. To overcome this limitation,
Self-Supervised Learning (SSL) has emerged as a promising alternative to address the challenge of limited annotations.
In this paradigm, models learn from unlabeled data by generating their own supervisory signals. The resulting pre-
trained models can then be fine-tuned on various downstream tasks such as image classification, object detection, and
semantic segmentation. However, achieving performance comparable to supervised learning often requires large-scale
datasets and high training costs, which significantly increase computational and storage demands. This internship
aims to alleviate these constraints by exploring data distillation techniques to make SSL training more efficient.

Sujet :
Dataset Distillation (DD) [1] aims to condense a large-scale training dataset into a much smaller synthetic one
such that models trained on the distilled data achieve performance comparable to those trained on the original
dataset (see figure 1). Most existing DD methods are designed for efficient supervised learning and can be broadly
classified into three main categories [2] : (1) Performance Matching, which minimizes the loss on the synthetic
dataset by aligning the performance of models trained on real and synthetic data, (2) Parameter Matching, which
trains two neural networks respectively on real and synthetic data and encourages similarity in their parameters and
(3) Distribution Matching, which generates synthetic data that closely mimics the distribution of the original dataset.
In this internship, we will focus on the Parameter Matching approach. Building upon the work of Cazenavette et al.
[3], the authors of [4] extended this concept to SSL using knowledge distillation [5, 6, 7], particularly employing SSL
methods such as Barlow Twins and SimCLR. In the same vein, this internship will explore the DINO (self-DIstillation
with NO labels, MetaAI) SSL method [8], which naturally produces teacher–student parameter trajectories that can
be leveraged for Parameter Matching. The different steps of the internship are :
▷ Step 1 – Literature review : Review recent dataset distillation methods applied to computer vision, with a
focus on parameter matching and SSL-based approaches.
▷ Step 2 – Trajectory Observation : Analyze and visualize the teacher–student parameter trajectories generated
by DINO during SSL training.
▷ Step 3 – Integration into Data Distillation Frameworks : Design a trajectory matching loss based on
DINO’s teacher–student dynamics and train a student model on synthetic data guided by these trajectories.
▷ Step 4 – Test on down-stream computer vision tasks : Assess the effectiveness of the proposed approach
on tasks such as image classification
– Bibliography
[1] Tongzhou Wang et al. “Dataset distillation”. In : arXiv preprint arXiv :1811.10959 (2018).
[2] Ruonan Yu, Songhua Liu et Xinchao Wang. “Dataset distillation : A comprehensive review”. In : IEEE transactions on pattern analysis and machine
intelligence 46.1 (2023), p. 150-170.
[3] George Cazenavette et al. “Dataset distillation by matching training trajectories”. In : Proceedings of the IEEE/CVF Conference on Computer Vision and
Pattern Recognition. 2022, p. 4750-4759.
[4] Siddharth Joshi, Jiayi Ni et Baharan Mirzasoleiman. “Dataset Distillation via Knowledge Distillation : Towards Efficient Self-Supervised Pre-training
of Deep Networks”. In : The Thirteenth International Conference on Learning Representations. 2025. url : https://openreview.net/forum?id=c61unr33XA.
[5] Geoffrey Hinton, Oriol Vinyals et Jeff Dean. “Distilling the knowledge in a neural network”. In : arXiv preprint arXiv :1503.02531 (2015).
[6] Ayoub Karine, Thibault Napoléon et Maher Jridi. “I2CKD : Intra- and inter-class knowledge distillation for semantic segmentation”. In : Neurocomputing
649 (oct. 2025), p. 130791. url : https://hal.science/hal-05144692.
[7] Ayoub Karine, Thibault Napoléon et Maher Jridi. “Channel-spatial knowledge distillation for efficient semantic segmentation”. In : Pattern Recognition
Letters 180 (avr. 2024), p. 48-54. url : https://hal.science/hal-04488459.
[8] Oriane Siméoni et al. “Dinov3”. In : arXiv preprint arXiv :2508.10104 (2025)

Profil du candidat :
The ideal
candidate should have knowledge in deep learning, computer vision, Python programming and an interest in efficient
machine/deep learning.

Formation et compétences requises :
Master 2 student or final year of MSc, or engineering school in computer science.

Adresse d’emploi :
45 rue des Saints-Pères, 75006, Paris

Document attaché : 202511111324_2025_Internship_DD_SSL.pdf

Offre en lien avec l’Action/le Réseau : – — –/– — –

Laboratoire/Entreprise : LIPADE
Durée : 6 months
Contact : ayoub.karine@u-paris.fr
Date limite de publication : 2026-04-30

Contexte :
Recently, several Large Vision Foundation Models (LVFMs) have been proposed in the literature [1]. They are
trained through a Self-Supervised Learning (SSL) paradigm on large-scale unlabeled datasets and evaluated on small
labeled datasets (fine-tuning). These models have achieved state-of-the-art performance across a wide range of
downstream computer vision tasks, including both non-dense tasks (e.g., image classification, image retrieval) and
dense tasks (e.g., semantic segmentation, object detection). However, the growing size and computational demands of
the LVFMs significantly constrain their applicability in resource-limited devices (e.g., drone, smarphone). For instance,
CLIP (Contrastive Language–Image Pretraining, OpenAI) [2] comprises up to 0.4 billion parameters, DINOv3 (self-
DIstillation with NO labels, MetaAI) [3] includes models with up to 7 billion parameters, and the SAM 2 (Segment
Anything Model, Meta AI) [4] exceeds 224 million parameters. To reduce the computational demands of such massive
architectures, this internship will focus on investigating knowledge distillation techniques.

Sujet :
The knowledge distillation (KD) technique [5, 6, 7] transfers knowledge from a powerful teacher network to a
smaller student model, enabling the student to achieve significantly improved performance with lower computational
cost. In this process, the student is trained on the same dataset as the teacher, allowing it to directly leverage the
teacher’s learned representations. However, directly applying KD to LVFMs presents several challenges. First, the
most performant LVFMs are developed by large tech companies, and their training datasets are often not publicly
available. Second, these LVFMs typically employ Vision Transformer (ViT) architectures [8] as encoders, whereas
convolutional neural networks (CNNs) are generally lighter and more computationally efficient, making them strong
candidates for student models on edge devices. Third, there are significant discrepancies in capacity between LVFMs
and smaller edge models. The latter two challenges are partially addressed by Lee et al. [9], who propose a method
to customize the well-generalized features of LVFMs for a given student model. Despite promising results, this work
does not thoroughly address the issues of unavailable source datasets and cross-architecture knowledge transfer.
Additionally, only the image classification task is considered. In this internship, we aim to tackle these challenges by
investigating state-of-the-art methods for cross-architecture KD [10], data-free KD [11] and adaptive KD [12]. As
illustrated in figure 1, we will focus on two dense down-stream tasks : semantic segmentation and object detection.
The different steps of the internship are :
▷ Step 1 – Literature review on KD from foundation models
▷ Step 2 – Compare different methods of cross-architecture KD, data-free KD and adaptive KD : The
teacher will be a LVFM such as CLIP, DINOv3 and SAM2. The student encoder should be a CNN one like ResNet18.
▷ Step 3 – Test the student model on different semantic segmentation and object detection datasets :
A comparison is to be done with classical KD methods dedicated to dense prediction.
– Bibliography
[1] Muhammad Awais et al. “Foundation models defining a new era in vision : a survey and outlook”. In : IEEE Transactions on Pattern Analysis and
Machine Intelligence (2025).
[2] Alec Radford et al. “Learning transferable visual models from natural language supervision”. In : International conference on machine learning. PmLR.
2021, p. 8748-8763.
[3] Oriane Siméoni et al. “Dinov3”. In : arXiv preprint arXiv :2508.10104 (2025).
[4] Nikhila Ravi et al. “Sam 2 : Segment anything in images and videos”. In : arXiv preprint arXiv :2408.00714 (2024)
[5] Geoffrey Hinton, Oriol Vinyals et Jeff Dean. “Distilling the knowledge in a neural network”. In : arXiv preprint arXiv :1503.02531 (2015).
[6] Ayoub Karine, Thibault Napoléon et Maher Jridi. “I2CKD : Intra- and inter-class knowledge distillation for semantic segmentation”. In : Neurocom-
puting 649 (oct. 2025), p. 130791. doi : 10.1016/j.neucom.2025.130791. url : https://hal.science/hal-05144692.
[7] Ayoub Karine, Thibault Napoléon et Maher Jridi. “Channel-spatial knowledge distillation for efficient semantic segmentation”. In : Pattern Recognition
Letters 180 (avr. 2024), p. 48-54. doi : 10.1016/j.patrec.2024.02.027. url : https://hal.science/hal-04488459.
[8] Alexey Dosovitskiy et al. “An Image is Worth 16×16 Words : Transformers for Image Recognition at Scale”. In : International Conference on Learning
Representations. 2021. url : https://openreview.net/forum?id=YicbFdNTTy.
[9] Jungsoo Lee et al. “Customkd : Customizing large vision foundation for edge model improvement via knowledge distillation”. In : Proceedings of the
Computer Vision and Pattern Recognition Conference. 2025, p. 25176-25186.
[10] Weijia Zhang et al. “Cross-Architecture Distillation Made Simple with Redundancy Suppression”. In : Proceedings of the IEEE/CVF International Confe-
rence on Computer Vision. 2025, p. 23256-23266.
[11] Qianlong Xiang et al. “Dkdm : Data-free knowledge distillation for diffusion models with any architecture”. In : Proceedings of the Computer Vision and
Pattern Recognition Conference. 2025, p. 2955-2965.
[12] Yichen Zhu et Yi Wang. “Student customized knowledge distillation : Bridging the gap between student and teacher”. In : Proceedings of the IEEE/CVF
International Conference on Computer Vision. 2021, p. 5057-5066.

Profil du candidat :
The ideal candidate should have knowledge in deep learning, computer vision, Python programming and an interest in efficient
deep learning.

Formation et compétences requises :
Master 2 student or final year of MSc, or engineering school in computer science

Adresse d’emploi :
45 rue des Saints-Pères, 75006, Paris

Document attaché : 202511111320_2025_Internship_KD_LVFM.pdf