User Tools

Site Tools


internships

This is an old revision of the document!


General requirements

For first and second year, internship topics are to be validated by Marie-Paule Cani and Erwan Scornet by sending an email mentioning the subject of the internship, the duration, the supervisor and the location of the internship. Once the internship topic is validated, you can put Erwan Scornet as the “Enseignant référent” at Ecole Polytechnique.

First-year internship

Period: from the beginning of April to the end of August (at least 4 months)

Defence: First week of September

An internship either in a private company or in a public lab. The research component of the internship is not mandatory but strongly encouraged. Students are requested to choose a topic related to the curriculum of their graduate degree program ie. involving either machine learning or visual computing, ideally both (see the first and second-year courses for details). This work can be theoretical (comprehension of a theorem and extension to a new setting), applied (developing and implementing a new solution) or mix theory and application.

The report must be at least 20 pages long. It must contain a general presentation of the topic and describe state-of-the-art methods in this field. The contributions of the student must be clearly identified and explained in details. There is no need for an exhaustive description of all codes produced during the internship. However, algorithms highlighting the challenging tasks solved by the student must be presented and explained in the report (if needed, relevant parts of the code can be included as well in an appendix).

Second-year internship

Period: from the beginning of April to the end of September (at least 5 months)

Defence: First week of September (the internship can be pursued after the defence).

A research internship either in the R&D department of a company or in a public research lab. The student must produce original work related to machine learning and/or to visual computing, ideally involving several different topics studied during the whole curriculum. This work can be theoretical or applied as soon as it contains novel ideas developed and validated by the student.

The report must be at least 30 pages long. It must contain a general presentation and motivation for the topic, point-out the related  challenges, and describe state-of-the-art methods in this field. The contributions of the student must be clearly highlighted and explained in details. There is no need for an exhaustive description of all codes produced during the internship. However, algorithms highlighting the challenging tasks solved by the student must be presented and explained in the report (if needed, relevant parts of the code can be included as well in an appendix).

Subjects for 2019-2020

Master 1 internships (4 to 5 months)

INRIA

Life Long Deep Learning (LLDL) Challenge Preparation

Abstract text.

Topic: Machine Learning, Deep Learning City and country: UPSud/INRIA Paris-Saclay (Gif-sur-Yvette) and Google (Zurich). Team or project in the lab: Isabelle Guyon (UPSud/INRIA), Michèle Sebag (CNRS), Zhengying Liu (PhD. Student). International collaboration with André Elisseeff (Google Zurich), Sergio Escalera (Univ. Barcelona), and Wei-Wei Tu (4Paradigm, China). Name and mail of the advisors: Isabelle Guyon (iguyon@lri.fr), Michèle Sebag (sebag@lri.fr), Zhengying Liu (zhengying.liu@lri.fr). Name and mail of the head of the department: Yannis Manoussakis (yannis.manoussakis@lri.fr) LRI, UPSud Paris-Saclay.

General presentation of the topic: Machine Learning problems (including Deep Learning) are often thought of as a “onetime” effort in which models are trained and tested on data coming from a common invariant source. Yet, in practice, data sources may drift and/or variants of an original problem may recur. Thus, it is important to capitalize on previously seen data and/or situations to adapt or solve new problems. Going beyond the usual setting of Transfer Learning or Domain Adaptation, researchers have been starting to investigate the possibility of Life-Long Machine Learning. In parallel, Deep Learning methods have known in the recent year tremendous success in applications, but also generated a lot of frustrations from users eager to try them out, but slowed down by the effort needed to tune hyper-parameters by trial and errors. To address the latter problem, we have engaged in a series of data science challenges with Google Zurich and 4Paradigm to foster research in the area of self-tuning machine learning algorithm (http://autodl.chalearn.org), which are part of the official selection of the prestigious NeurIPS conference. To prepare these challenges, we have formatted nearly 100 datasets coming from a variety of domains and including speech, images, videos, text, and tabular data, all formatted in a unified way, but NOT preprocessed into fixed-length feature vectors. Therefore, we have a wealth of data making it possible to organize an unprecedented LLDL challenge.

Objective of the internship: The aim of this internship will be to help prepare the LLDL challenge. The intern will produce baseline results for the challenge using state-of-the-art techniques to bootstrap participation in the competition. Available to him will be the solutions of the winners to past AutoDL challenges. However, such solutions do not involve “meta-learning”, i.e. learning from exposure to past datasets to perform better on new tasks. The intern with explore with meta-learning strategies, including reinforcement learning. The intern will join a dynamic and motivated team of challenge organizers and may contribute to a publication on the challenge design. Bibliographic references: [1] I. Guyon, K. Bennett, G. Cawley, H. J. Escalante, S. Escalera, T. K. Ho, B. Ray, M. Saeed, A. Statnikov, and E. Viegas. AutoML challenge 2015: Design and first results. [2] Liu Z, Bousquet O, Elisseeff A, Escalera S, Guyon I, Jacques J, Pavao A, Silver D, Sun-Hosoya L, Treguer S, Tu WW. AutoDL Challenge Design and Beta Tests-Towards automatic deep learning. InCiML workshop@ NIPS2018 2018 Dec. [3] B. Zoph, and L. Quoc. “Neural architecture search with reinforcement learning.” arXiv preprint arXiv:1611.01578(2016). [4] Elsken, Thomas, Jan Hendrik Metzen, and Frank Hutter. “Neural architecture search: A survey.” arXiv preprint arXiv:1808.05377 (2018). [5] Zhengying Liu, Isabelle Guyon, Julio Jacques,Jr, Meysam Madadi, Sergio Escalera, Adrien Pavao, Hugo Jair Escalante, Wei-Wei Tu, Zhen Xu, Sebastien Treguer, AutoCV Challlenge Design and Baseline Results. https://hal.archives-ouvertes.fr/hal-02265053 [6] Madrid, J.G., Escalante, H.J., Morales, E.F., Tu, W.W., Yu, Y., Sun-Hosoya, L., Guyon, I. and Sebag, M., 2019. Towards AutoML in the presence of Drift: first results. arXiv preprint arXiv:1907.10772.

Expected ability of the candidate: We are accepting candidates with background in machine learning, statistics, scientific modeling, signal processing and control (or at least a subset if those). The candidate should have the ability of working on cross-disciplinary problems, have a strong math background, and the experience or strong desire to work on practical problems. Good programming skills are also required. Experience with TensorFlow and/or PyTorch is preferred. Experience in GPU/TPU computing is a plus.

LJK-Inria Grenoble

Research internships at LJK-Inria Grenoble, in Computer Graphics, contact Fabrice Neyret Fabrice.Neyret@imag.fr

Link Abstract :

    M2R (+ X-M1): Playing with Fourier spectrum for real-time ocean explorer and texture generation 

http://www-evasion.imag.fr/Membres/Fabrice.Neyret/Etudiants/sujet1-2020.html

    M2R (+ X-M1): Procedural modeling of 3D realistic galactic dust and nebulas 

http://www-evasion.imag.fr/Membres/Fabrice.Neyret/Etudiants/sujet2-2020.html

    M2R (+ X-M1): Real-time walk-through the Milky-Way: continuum to points, on-the-fly generation, and mixing with catalog 

http://www-evasion.imag.fr/Membres/Fabrice.Neyret/Etudiants/sujet3-2020.html

    M2R: Revisiting volumetric ray-tracing to make it well-posed 

http://www-evasion.imag.fr/Membres/Fabrice.Neyret/Etudiants/sujet4-2020.html

Master 2 internships (4 to 5 months)

Ecole Polytechnique / LIX / STREAM - Master 2 level Artificial Intelligence and advanced Visual Computing https://www.lix.polytechnique.fr/stream/job-offer/

TOTAL Topological Data Analysis - 6-month internship - Master 2 level or end of engineering school

Abstract text. As part of a collaboration with Total, we are actively looking for a student with an interest in Topological Data Analysis for a 6-month internship at Master 2 level or at the end of an engineering school, with a view to continuing with a CIFRE thesis. Candidates already holding an M2 degree are invited to apply directly for the CIFRE thesis.

The internship and thesis focus on the exploitation, using topological methods, of volumes of probability of the presence of seismic faults generated by deep learning.

Desired start date: April 1, 2020. Location of the internship: Sorbonne University, Paris (metro Jussieu)

The detailed subject is available below: https://julien-tierny.github.io/stuff/openPositions/internship2020_su_total.pdf

TOTAL Topological Data Analysis - 6-month internship - Master 2 level or end of engineering school with a continuation in thesis

Abstract We are actively looking for a student (or two) with an interest in Topological Data Analysis for a 6-month internship at Master 2 level or at the end of an engineering school, with a view to continuing with a thesis. The algorithms designed during the internship will be implemented in C++ within the open-source library “Topology ToolKit” (https://topology-tool-kit.github.io/) that we are developing with our collaborators. Desired start date: April 1, 2020. Location of the internship: Sorbonne University, Paris (metro Jussieu) Detailed topics are available below: https://julien-tierny.github.io/stuff/openPositions/internship2020a.pdf https://julien-tierny.github.io/stuff/openPositions/internship2020b.pdf https://julien-tierny.github.io/stuff/openPositions/internship2020c.pdf

If you already have a Master's degree (or equivalent) and are interested in a thesis related to these subjects, please send us your application letter with an updated CV. Dr. Julien Tierny CNRS Researcher Sorbonne University http://lip6.fr/Julien.Tierny

INRIA Automatic Deep Learning (AutoDL) Benchmark with extensive GPU usage

Topic: Machine Learning, Deep Learning City and country: UPSud/INRIA Paris-Saclay (Gif-sur-Yvette) and Google (Zurich). Team or project in the lab: Isabelle Guyon (UPSud/INRIA), Michèle Sebag (CNRS), Zhengying Liu (PhD. Student). International collaboration with André Elisseeff (Google Zurich) and Wei-Wei Tu (4Paradigm, China). Name and mail of the advisors: Isabelle Guyon (iguyon@lri.fr), Michèle Sebag (sebag@lri.fr), Zhengying Liu (zhengying.liu@lri.fr). Name and mail of the head of the department: Yannis Manoussakis (yannis.manoussakis@lri.fr) LRI, UPSud Paris-Saclay. General presentation of the topic: Despite all the recent successes in Machine Learning in general and Deep Learning in particular, there is still a lot of room for improvement. Indeed, even though there has been significant advances both in hardware and software available for implementing and efficiently running large scale Machine Learning systems, a key bottleneck still remains: designing appropriate model architectures and tuning hyper-parameters is still a very tedious and labor-intensive process. Training a complex deep network is often a long and frustrating trial-and-error process involving lots of heuristics and requiring a fair amount of practical experience and expertise. Even worse, the difficulty of those tasks often translates into lack of reproducibility of the published results. To address this problem, we have engaged in a series of data science challenges with Google Zurich and 4Paradigm to foster research in the area of self-tuning machine learning algorithm (http://autodl.chalearn.org) which are part of the official selection of the prestigious NeurIPS conference. We also won an award to use the new supercomputer

Jean Zay http://www.idris.fr/annonces/annonce-jean-zay-eng.html for the purpose of conducting intensive simulations of the problem of AutoDL.

Objective of the internship: The aim of this internship will be to conduct systematic experiments to reproduce the results of the AutoDL competition and go beyond. One challenging aspect of the Google design is that the participants were exposed to data coming from a variety of domains and including speech, images, videos, text, and tabular data, all formatted in a unified way, but NOT preprocessed into fixed-length feature vectors. The winners have made their code publicly available and explained their methods in fact sheets, but clearly their methods involve a lot of ad hoc engineering. The purpose of this internship is then to identify the key ingredients of success of the methods and create a simplified , principled, and robust methodology. This internship may lead to a publication, in collaboration with the organizers.

Bibliographic references: [1] I. Guyon, K. Bennett, G. Cawley, H. J. Escalante, S. Escalera, T. K. Ho, B. Ray, M. Saeed, A. Statnikov, and E. Viegas. AutoML challenge 2015: Design and first results. [2] Liu Z, Bousquet O, Elisseeff A, Escalera S, Guyon I, Jacques J, Pavao A, Silver D, Sun-Hosoya L, Treguer S, Tu WW. AutoDL Challenge Design and Beta Tests-Towards automatic deep learning. InCiML workshop@ NIPS2018 2018 Dec. [3] B. Zoph, and L. Quoc. “Neural architecture search with reinforcement learning.” arXiv preprint arXiv:1611.01578(2016). [4] Elsken, Thomas, Jan Hendrik Metzen, and Frank Hutter. “Neural architecture search: A survey.” arXiv preprint arXiv:1808.05377 (2018). [5] Zhengying Liu, Isabelle Guyon, Julio Jacques,Jr, Meysam Madadi, Sergio Escalera, Adrien Pavao, Hugo Jair Escalante, Wei-Wei Tu, Zhen Xu, Sebastien Treguer, AutoCV Challlenge Design and Baseline Results. https://hal.archives-ouvertes.fr/hal-02265053 [6] Ying, Chris, Aaron Klein, Esteban Real, Eric Christiansen, Kevin Murphy, and Frank Hutter. “Nas-bench-101: Towards reproducible neural architecture search.” arXiv preprint arXiv:1902.09635 (2019). Expected ability of the candidate: We are accepting candidates with background in machine learning, statistics, scientific modeling, signal processing and control (or at least a subset if those). The candidate should have the ability of working on cross-disciplinary problems, have a strong math background, and the experience or strong desire to work on practical problems.

Good programming skills are also required. Experience with TensorFlow and/or PyTorch is preferred. Experience in GPU/TPU computing is a plus.

LJK-Inria Grenoble

Research internships at LJK-Inria Grenoble, in Computer Graphics, contact Fabrice Neyret Fabrice.Neyret@imag.fr

Link Abstract :

    M2R (+ X-M1): Playing with Fourier spectrum for real-time ocean explorer and texture generation 

http://www-evasion.imag.fr/Membres/Fabrice.Neyret/Etudiants/sujet1-2020.html

    M2R (+ X-M1): Procedural modeling of 3D realistic galactic dust and nebulas 

http://www-evasion.imag.fr/Membres/Fabrice.Neyret/Etudiants/sujet2-2020.html

    M2R (+ X-M1): Real-time walk-through the Milky-Way: continuum to points, on-the-fly generation, and mixing with catalog 

http://www-evasion.imag.fr/Membres/Fabrice.Neyret/Etudiants/sujet3-2020.html

    M2R: Revisiting volumetric ray-tracing to make it well-posed 

http://www-evasion.imag.fr/Membres/Fabrice.Neyret/Etudiants/sujet4-2020.html

Unity Technologies

Unity Technology est une entreprise internationale (+2500 employés) qui développe le moteur de jeux vidéo Unity.

Unity a une antenne de recherche à Grenoble spécialisée dans le domaine de l'image de synthèse et du machine learning. Nous recherchons régulièrement des candidats pour des stages recherche niveau ingénieur/M2 avec possibilité de poursuite en thèse CIFRE. Nous définissons les sujets de stage en fonction des intérêts des candidats et de ceux de l'équipe.

Plus d'informations sont disponibles sur le site de l'équipe de Grenoble. Personnel : http://unity-grenoble.github.io/website/people.html Liste des travaux et des publications : http://unity-grenoble.github.io/website/publications.html Offre de stage et d'emploi : http://unity-grenoble.github.io/website/news.html

ANATOSCOPE

Stage : Creation et analyse d’un format 3Da base de contours 2D empile s Niveau : Stage d'étude (M2) Duree : 6 mois Lieu : Grenoble (Montbonnot) Encadrement : Ulysse Vimont Remuneration :a discuter Mots clefs : 3D, C++, mode lisation surfacique Contact : vimont@anatoscope.com, job@anatoscope.com link :http://www.lix.polytechnique.fr/Labo/Marie-Paule.Cani/MasterAI/doku.php?id=start&image=sujet_stage_anatoscope_spr_compressed.pdf&ns=&tab_details=view&do=media

IMAGEENS Sujet de stage : Classification automatique des images médicales cardiovasculaires www.imageens.com linlk : http://www.lix.polytechnique.fr/Labo/Marie-Paule.Cani/MasterAI/doku.php?id=internships&image=sujetstage_classification-1_imageens.pdf&ns=&tab_details=view&do=media

PhD

Title

Abstract text.

Jobs

Title

Abstract text.

List of internships for 2018-2019

Master 1 internships (4 to 5 months)

Automatic landmark detection on 3D meshes using convolutional neural networks. Anatoscope.

Context : One mission of Anatoscope is to offer solutions to automatically transform medical imaging to complete 3D digital twins of the patients, with the objective of helping diagnostics and improving treatment. One method to construct personalized anatomical avatar is to combine a canonical avatar (3D model) with medical images of the person (CT, MRI, 3D scans…) by registration. Landmark detection is an essential early process in a typical pipeline of establishing registration. It refers to the localization of fiducial key points of the related object. The detected landmarks can be used to initialize parameters of the 3D morphable model. They can also be used directly to determine correspondences between the canonical model and the 2D image or 3D scan. In this work we focus on the case of automatic landmark detection on 3D scan. In practice, the captured 3D scans are often noisy and incomplete, which increases the difficulty of the landmark detection task. (see full proposal)

Link to the full proposal.

Mappiiing - Artificial Intelligence

Link to the full proposal

IDEMIA

See Link to the full proposal

sc.io*

Sujet 1: sc.io, startup spécialisée dans l’extraction automatique de données recherche un(e) stagiaire pour son équipe de “Data Science and Engineering”.

Nous combinons une approche traditionnelle de Deep Learning avec des techniques directement issues de la recherche de type weak supervision qui nous permettent de pratiquement nous affranchir du processus de préparation (labelling) de training sets.

Nous recherchons des stagiaires passionné(e)s par les possibilités du Machine Learning et intéressées par la reconnaissance d’images, le traitement du langage naturel (NLP), la calibration et le deploiement d’algorithmes de RNN et les mathématiques appliquées en général afin d'améliorer l‘approche existante et d’en étendre le champ d’action.

Nos stagiaires ont la possibilité de travailler directement sur (i) les algorithmes utilisés par nos clients et (ii) des résultats de recherche récente publiée lors de conférences de premier plan telles que NIPS par ex.

Ce stage est d'une durée minimale de 4 mois à partir de mars.

Responsabilités: Effectuer des recherches visant à intégrer des techniques de reconnaissance d’image au traitement de documents afin d’en ameliorer la comprehension par nos algorithmes. Validation des résultats a partir d’une base de données publiques ou à construire. Intégrer et collaborer sur des projets de recherche appliquée pour un déploiement chez nos clients.

Qualifications minimales: Expérience à résoudre des problèmes d’analyse à l’aide d’approches quantitatives. Capacité à manipuler et à analyser des données à partir de sources variées. Connaissance dans un langage de programmation de type Python, C# et / ou Java. Formation en ingenierie, mathematiques, informatique ou similaire.

Qualifications souhaitées: Expérience concrete des techniques de Machine Learning. Utilisation de TensorFlow ou Keras

Master 1 or master 2 internship

Ynsect Internship in image recognition and machine learning Ÿnsect is a pioneer and leading company in insect industrial technologies. We develop insect farms. This technology consists, on the one hand, of rearing insects at large scale. On the other hand, it consists of processing insects into proteins and lipids for pet food and fish feed.

R&D USP department is looking for an apprentice in image recognition and machine learning in order to develop methods and tools to contribute to the quality control of industrial insect rearing.

Link to the full proposal

Master 2 internships (5 to 6 months)

Apprentissage Profond pour le suivi systématique de cellules d’embryons vivants imagées par microscopie (LIRRM et CRBM, Montpellier).

L'apprentissage automatique et plus particulièrement l'apprentissage profond (Deep Learning) donnent des résultats très prometteurs dans divers domaines de la recherche en biologie. Le but du stage est d'appliquer l'apprentissage profond au domaine de l'imagerie biologique, et plus particulièrement à la segmentation cellulaire 4D (3D plus temps), la reconnaissance et le suivi de chaque cellule dans des embryons vivants en développement. L’apprentissage se fera sur une base de données préexistante de plus de 10 embryons entièrement segmentés (> 10000 cellules suivies). L’objectif est de créer un suivi des cellules basé uniquement sur l’apprentissage en profondeur. Durant ce stage, en fonction du profil et des intérêts du stagiaire, nous pourrons :
- Implémenter et adapter à notre base de données les méthodes d'apprentissage en profondeur basées sur l'identification d'objets individuels (instance segmentation).
- Explorer les aspects théoriques du concept de mémoire dans les réseaux neuronaux récurrents (LSTM) et les adapter au suivi cellulaire dans des images 3D + temps.
- Utiliser l'apprentissage en profondeur pour prédire et définir les règles de la division cellulaire au cours de l'embryogenèse.

Ce stage de 5 à 6 mois peut déboucher sur un projet de thèse, combinant potentiellement des approches informatiques avec la validation expérimentale des prédictions.

PROFIL ET COMPÉTENCES DE RECHERCHE :

Le stage est principalement destiné aux étudiants ayant une formation initiale en mathématiques et en informatique et un intérêt pour la biologie, mais aussi à des biologistes ayant de très bonnes compétences en programmation informatique et en analyse statistique. Esprit d'équipe, autonomie, dynamisme et créativité seraient appréciés. La maîtrise de l'anglais ou du français technique doit être suffisante pour rédiger la documentation technique et interagir verbalement quotidiennement.

Langage de programmation requis : python L'étudiant sélectionné aura des connaissances de base sur au moins l'un des sujets suivants : la théorie de l'apprentissage automatique, le traitement d'image, le calcul haute performance ou l'utilisation de bibliothèques d'apprentissage en profondeur (par exemple Keras).

Contact : Le stage sera co-encadré par un informaticien (Emmanuel Faure, LIRRM, Montpellier, emmanuel.faure@irit.fr) et un biologiste (Patrick Lemaire, CRBM, Montpellier, patrick.lemaire@crbm.cnrs.fr).

Two internship proposals KAIST, South Korea, contact Jean-Charles Bazin bazinjc@kaist.ac.kr

  • Artificial Intelligence (Deep Learning) Approach for Automatic Editing of Actor’s Facial Performance Videos and Movie Special Effects Link to project proposal
  • Deep Learning Approach for Automatic Capture and Editing of Drone Videos. Link to project proposal

Automatic landmark detection on 3D meshes using convolutional neural networks . Anatoscope.

Context : One mission of Anatoscope is to offer solutions to automatically transform medical imaging to complete 3D digital twins of the patients, with the objective of helping diagnostics and improving treatment. One method to construct personalized anatomical avatar is to combine a canonical avatar (3D model) with medical images of the person (CT, MRI, 3D scans…) by registration. Landmark detection is an essential early process in a typical pipeline of establishing registration. It refers to the localization of fiducial key points of the related object. The detected landmarks can be used to initialize parameters of the 3D morphable model. They can also be used directly to determine correspondences between the canonical model and the 2D image or 3D scan. In this work we focus on the case of automatic landmark detection on 3D scan. In practice, the captured 3D scans are often noisy and incomplete, which increases the difficulty of the landmark detection task. (see full proposal)

Link to the full proposal.

Reconnaissance automatique de la parole (Airbus Elancourt).

See Link to the full proposal

Contact: mohamed.bouaziz@airbus.com

Text-Mining et population de bases de connaissances (Airbus Elancourt).

See Link to the full proposal

Contact: mohamed.bouaziz@airbus.com

Exploring deep learning for Virtual Reality (Immersion, Bordeaux)

En charge de la recherche pour la société Immersion (leader Européen en réalité virtuelle), je recherche des candidats potentiels pour un stage orienté IA. Le stage consistera à élaborer une stratégie d’exploration des potentiels du deep learning dans notre contexte de développement. Il s’agira également de soutenir les développements actuels sur le sujet afin d’aboutir à des prototypes démonstratifs.

Contact: Julien Castet julien.castet@immersion.fr

Start-up FeetMe

-Envie de découvrir une véritable MedTech? qui développe son propre dispositif médical connecté, leader dans son domaine qui intègre à une plateforme centralisée des données de santé.

-Envie de travailler au développement de biomarqueurs digitaux de demain liés à la mobilité sur des données collectées avec les meilleurs centres experts.

Si tu souhaites découvrir l'environnement d'une startup en pleine croissance, avec une véritable ADN tech et avec des solutions technologiques de rupture, tu trouveras plus de détail dans nos offres de stage:

Link to the full proposal

IDEMIA

See Link to the full proposal

Laboratoire de météorologie dynamique: The power of deep learning applied to oceanic eddy detection

See Link to the full proposal

CRITEO

See Link to the full proposal

THALES

If you are interested by one of these internships, please contact Iwona Piskulska iwona.piskulska@thalesgroup.com or Frederic Barbaresco frederic.barbaresco@thalesgroup.com.

2019_stage_deeplearning_optimisation.pdf

2019_stage_denoising.pdf

descriptif_de_projet_masterx_thalesatm_beatricepesquetpopescu.pdf

stage_etude_intelligence_artificielle_pour_systeme_de_defense_multi-agents.pdf

SAFRAN

Deep Learning in the field of numerical experiments

offre_stageia_safrantech.pdf

EDF

Apprentissage par renforcement. Contact SCHMITT Damien damien.schmitt@edf.fr

stage_apprentissage_renforcement_2019_acceleration.doc

stage_apprentissage_renforcement_2019_interpretabilite.docx

stage_apprentissage_renforcement_2019_es.doc

stage_apprentissage_renforcement_2019_contraintes.doc

PhD thesis proposals

Automatiser une chaine de renseignement militaire (THALES, apprenti ou thèse sur 3 ans).

L’observation est permanente. On traite plus de 10 000 interceptions par jour soit plusieurs dizaines ou centaines de millions d’impulsions radar élémentaires. Le traitement est automatique mais il peut y avoir par exemple 1% d’échec ce qui représente entre 100 ou 1000 interceptions ou plus. Le traitement fait un diagnostic de chaque interception sur chaque paramètre pour savoir si il y a correspondance avec un signal cible. Le traitement analyse pour cela plus de 100 paramètres, avec une hiérarchie sur les types d’erreur rencontrés. Il peut être aussi amené à déclencher des actions pour compléter les faits à disposition. Les diagnostics élémentaires sont tous automatisés.

Quand tout ne « match » pas, il faut déterminer si c’est un signal inconnu ou une variante d’un mode d’émission connu. Et faire l’analyse des causes.Et compléter en conséquence la base de données. Tant qu’on a pas « légiférer » sur le type de problème rencontré, c’est un opérateur expert qui décide. On dispose donc d’une pseudo-réalité terrain.

Pour traiter automatiquement ces situations d’échec on envisage l’utilisation :

- D’un apprentissage d’arbres de décisions pour retrouver automatiquement le bon diagnostic. A ce niveau l’explication (retour aux causes) est indispensable.

- De régles locales définies par l’expert sur les noeuds terminaux.

Pour l’apprentissage, nous envisageons d’utiliser :

- une technique rapide ne donnant pas forcément d’explication pour approcher la performance décisionelle que nous pourions atteindre

- Une technique à base d’arbres de décision pour avoir la possibilité d’expliquer le raisonement (et à terme de le compléter)

Nous avons réalisé il y a 20 ans un procédé type CART avec apprentissage local de densités de probabilités par Kernel. Dans le cadre d’une thèse nous aimerions approfondir la pertinence de ce type de technique avec les forets aléatoires. Il faudra aussi compléter (voir plus loin).

Dans le cadre de l’apprenti, nous voulons réaliser un véritable prototype travaillant sur données réelles (d’où une nécessité d’habilitabilité de celui-ci au niveau confidentiel requis). Et mettant en œuvre les techniques que nous trouvons sur étagère, puis les techniques plus raffinées extraites de la thèse.

Sur les nœuds terminaux de l’arbre, il pourrait être nécessaire d’améliorer le traitement en lançant des diagnostics partiels plus élaborés, répetoriés ou construis adhoc à la premiere fois ou le cas serait rencontré. Pour cela des régles simples de déclenchement d’action et de décision semblent idéales. Se posera alors le problème du maintien en cohérence des systèmes locaux de régles et les problèmes de généralisation et spécialisation.

Contact : Jean-Francois Grandin jean-francois.grandin@fr.thalesgroup.com

THALES

See propositiondethese_tlasome_lip6.pdf

internships.1576504699.txt.gz · Last modified: 2019/12/16 14:58 by payan