Incertitudes et méconnaissances dans les interpolations spatiales pour la cartographie des risques en milieu urbain
Contexte
L’interpolation des données spatiales est une opération courante dans le domaine des géosciences. Si les exemples d’application sont innombrables, un constat important est la faible proportion d’études proposant l’estimation des incertitudes (<5%). Or les incertitudes peuvent être multiples, de différentes natures, et plus spécifiquement, l’incertitude liée à l’imperfection des connaissances (épistémique) peut être significative dans des applications à forts enjeux sociétaux en milieu urbain. Une voie prometteuse pour une prise en compte exhaustive et transparente des incertitudes est celle de la théorie des probabilités imprécises (comprenant en particulier la théorie des possibilités, Dubois et Prade, 1988 et la théorie de Dempster Shafer, Shafer 1976, Dempster 1967). Ce cadre a des fondements dans la théorie classique des probabilités et peut être vu comme une généralisation du cadre Bayésien en apportant un degré de flexibilité supplémentaire pour exprimer les différentes types d’incertitudes.
Le stage
Dans le cadre de ce stage (en partenariat entre UTC, IRIT et BRGM), nous cherchons à comparer les résultats donnés par différentes approches (Loquin 2010), par exemple utilisant la théorie des possibilités ou encore la théorie Bayésienne standard. De telles comparaisons ne sont pas triviales, et demandent en général de développer des cadres dédiés. Les développements du stage seront d’abord testés au sein d’études simulées, par exemple pour comparer les cadres proposés dans des situations différentes (e.g., absence et présence d’erreur de modélisation, de données aberrantes). Ils seront ensuite appliqués au cas de l’estimation des fonds pédogéochimiques urbains (FPGU), notamment pour Toulouse Métropole (Belbèze 2019), dont le contexte de données / connaissances parcellaires, imprécises, clustérisées est représentatif d’un ensemble de situations en pratique. Ces données comprendront les résultats sols des analyses BRGM ainsi que les covariables utilisés pour le rapport FGU EXPLO (usage, géologie,segmentation du territoire).
Candidature
Helene Fargier, IRIT/CNRS, 3IA ANITI Senior
Romain Guillaume, IRIT/UPS & 3IA ANITI
Stage :
M2 / ingénieur
Durée : 3 à 6 mois – printemps 2022
Explainability of Preference Models for decision-aid system
Context
Decision aid systems, like on-line configurators or recommender systems, need to adapt themselves to each user in order to offer a better interaction and guide the user quickly to the best decision for her: the system should be able to gather a model of the preferences of the user, and be able to show her, almost instantly, during the interaction, what seems to be her best, most preferred possible alternatives. Several models of preferences have been developed in the literature on Artificial Intelligence and Operations Research, offering the possibility to represent complex preferences over multi-attribute domains in some rather compact form. The richness of the models comes at a cost: finding the optimal alternatives is, in general, a computationally hard problem – at least NP-nard – except for some quite restrictive models.
On the other hand, the recent expansion of the use of Artificial Intelligence in many fields is accompanied by a demand for interpretability of the models used. In order to be trusted by its users, an intelligent system should not only be able to solve problems, but also capable to explain its solutions ; the explanation should allow the user to identify elements on which the solution is based.
About the internship
The topic of this internship is to compare models of preferences from the point of view of their explainability. Explaining models of preferences is an emerging topic. Recent works have explored explainability of Multi-Criteria Decision Aiding models [Belahcene et al, 2016 ; Labreuche & Fossier, 2018] and of Bayesian Networks classifiers [Marques-Silva et al., NIPS 2020; Koopman & Renooij, 2021]. The aim of this internship is to study what kind of explanations one would expect for decision-aid systems based on preference models ; and to study the complexity of generating such explanations for existing preference models. Models of interest can be valued CSPs, Bayesian Network as a preference representation model, lexicographic models, CP-nets
Application
Helene Fargier, IRIT/CNRS, 3IA ANITI Senior
Jerome Mengin, IRIT/UPS & 3IA ANITI
Stage :
M2
Duration : 3 to 6 months – spring 2022
Engineer or master thesis internship : domain adaptation for automotive radar object detection, for ANITI partner : NXP
Context
NXP automotive business unit is in charge of developing IC for the connected car, covering areas as diverse as in-vehicle networking, automotive lighting, car radio and audio, automotive power and radar.
The 77GHz radar development team based in Toulouse oversees developing automotive transceiver IC used in short to long range applications, enabling use-cases from parking to cruise control, anti-collision and later autonomous vehicle.
About the position
We are looking for a master thesis student for a 6-month internship in computer science or mathematics to work on domain adaptation techniques based on AI for simulated radar data.
Domain Adaptation is a technique to bootstrap the AI model development which enables better generalization on real data of AI models trained on simulated data.
The intern will be in charge of the following missions :
- Generate a large amount of labelled dataset with dSpace simulation tool, with help of NXP people in California that have the simulator.
- Collect, process and annotate data to enhance Toulouse site multi-sensors dataset (including cameras, lidar and radar).
- Collaborate with an apprentice and a PhD student to collect data and annotate the captured real dataset.
- Work in contact with other NXP sites to be aware about ongoing and future projects for radar using A.I.
- Apply Domain Adaptation technique on existing AI models between simulated and real captured dataset: train models on simulated dataset and evaluate the models on real dataset.
- The targeted applications for the models are High resolution DOA (direction of arrival) estimation and point cloud perception.
- If time allows, find models architecture that give better performances with Domain Adaptation on the tasks listed above.
Profile
You are doing Master at University or last year in engineering school, with desirably knowledge and interest in computer science, deep learning and image/signal processing. You are familiar with computer languages ( Python, C, C++, Matlab, and Linux) and you have knowledge or interest in deep learning/artificial intelligence (Tensorflow, Pytorch…). A good level in English, written and spoken is required.
The internship may be continued by a thesis as part of the 3IA (Interdisciplinary Institute of Artificial Intelligence) of Toulouse ANITI
Application
Apply on NXP website
Contact : Didier Salle
Stage :
M2
Duration : 6 months starting asap
Profile : engineer or Master student in computer science, mathematics or artificial intelligence
Testing of a UAV software architecture based on skill and fault models
Description
The development of decisional autonomous systems, like drones or mobile robots, now makes it possible to perform tasks without human supervision for extremely varied environments. However, the failures of these systems can have unacceptable consequences for the mission. These failures may appear at low functional level of the architecture of the system (e.g., sensors), but also at higher levels where decision are taken (e.g., behaviour management).
When developing these systems, a major challenge is to check if all the possible failures have been considered and mitigated. To do so, a relevant approach is to combine (1) failure analysis (e.g. through Fault Tree models) to list the possible failures and how they are supposed to be handled, and (2) testing, to have an evidence that the architecture correctly implements the handling of these failures.
For this internship, we will focus on a three-layers architecture: (1) a functional layer developed with ROS (Robot Op- erating System) that manages all sensors, actuators, and basic control, (2) a decisional layer at the top where high-level behaviors are implemented, and (3) an intermediary skill management level between the functional and the decisional layer, that checks the correct preconditions and resource states of the functional layer when executing behaviours planned by the decisional layer. A previous work has performed a Fault Tree analysis of the skill- and functional- layers in order to identify the possible failures and how they are managed in the architecture. This methodology has been applied on a case study, consisting of a drone performing the mapping of a rescue area, in which several failures may occur (e.g., GPS failure, loss of communication with the pilot, or sensor failures). The objective of the internship is to define how test cases can be derived from the fault-tree analysis and from the skills specification. A first step will then be to compare and discuss the set of test cases we can derive from both models. A second step will be to execute these tests, using simulations, or data replay, with or without fault injection. This step will require to use the proposed autonomous architecture, perform fault injection in the ROS functional layer, and analyse test results. x
Location
- ONERA / LAAS CNRS, Toulouse, France
Student profile
Computer science, Embedded systems (robotics or dependability is a plus)
Application
Please send a cv + motivation + transcript (marks) of the last 2 university years to Charles Lesire, Jeremy Guiochet and Augustin Manecy.
Stage :
M2
Duration : 4 to 5 months between Feb – Sept 2022
Apprentissage multi tâche pour le traitement de la parole et de la langue dans le cadre de conversations spontanées multi-locuteurs
Contexte
L’équipe R&D de la société LINAGORA développe en open-source des outils d’assistance intelligente pour entreprises, y compris l’assistant vocal LinTO, et LinSTT, un outil de reconnaissance de la parole qui est capable de transcrire sous forme textuelle un signal vocal, ce qui nous permet de produire, de manière automatique, des transcriptions de réunion. Actuellement, nous travaillons sur un gestionnaire de conversation, Conversation Manager, une plateforme qui permettra à partir d’un enregistrement complet d’une réunion d’en déduire un résumé aussi pertinent que possible. L’idée est qu’un utilisateur du Conversation Manager va pouvoir d’abord visualiser, corriger et annoter une transcription proposée par notre système et ensuite exploiter le contenu de la transcription et ses annotations pour créer un résumé de manière semi-automatique.
Pour ce faire, il est impératif que la transcription proposée à l’utilisateur, avant l’étape de correction, soit aussi correcte et facile à visualiser que possible, ce qui peut être difficile pour les transcriptions de réunion où il y a plusieurs locuteurs et où les participants ont tendance à faire des interventions longues et mal structurées d’un point de vue grammatical. Pouvoir bien associer un tour de parole à son locuteur (segmentation et regroupement en locuteurs, ou diarisation en anglais) et ajouter les marques de ponctuation qui rendent le texte plus facile à lire sont très importants pour faire des transcriptions de haute qualité.
La diarisation et la ponctuation peuvent ensuite servir à améliorer les algorithmes de résumé automatique en aidant un système à découper le contenu d’une réunion en clauses individuelles — appelés segments discursifs. Ces segments fournissent des unités sémantiques qui seront passées ensuite aux algorithmes de résumé qui jugeront quels segments sont plus centraux à la conversation et du coup, au résumé final.
Missions
Pour ce stage, le stagiaire étudiera les trois tâches – la diarisation, la ponctuation, et la segmentation discursive – en parallèle avec une approche d’apprentissage multi-tâche. L’entraînement du modèle sera fait sur des données de conversation transcrites soit en français, soit en anglais. Nous commencerons avec des modèles existants de ponctuation et segmentation qui se basent sur une architecture de transformer + bi-LSTM ainsi qu’un modèle de diarisation. La nouveauté de ce stage consistera dans (a) l’approche multi-tâche pour étudier ces trois sujets en parallèle et (b) l’usage des informations acoustiques des enregistrements de conversation et de réunion (alors que les modèles de base pour la ponctuation et la segmentation discursive sont entraînés exclusivement sur du texte).
L’encadrement du stage : Le stagiaire sera encadré par Samir Tanfous de LINAGORA, mais travaillera en collaboration avec Julie Hunter de LINAGORA et plusieurs membres du laboratoire IRIT, notamment Philippe Muller de l’équipe Melodi (NLP) et Thomas Pellegrini et Hervé Bredin de l’équipe Samova (Traitement de la parole).
Localisation : LINAGORA, soit à Paris, soit à Toulouse
Compétences clés recherchées
- Étudiants de M2 ou d’école d’ingénieur en dernière année, en informatique et IA avec des compétences en machine learning
- Expérience en deep learning et PyTorch serait un plus
- Expérience en speech processing et/ou NLP serait un plus
Pour candidater, envoyez votre CV à Julie Hunter
Références
Bredin, H., Laurent, A. (2021) End-To-End Speaker Segmentation for Overlap-Aware Resegmentation. Proc. Interspeech 2021, 3111-3115.
Muller, P., Braud, C., Morey, M. (2019) ToNy: Contextual embeddings for accurate multilingual discourse segmentation of full documents. Proceedings of the Workshop on Discourse Relation Parsing and Treebanking 2019, 115-124.
Stage :
M2
Durée : 5 à 6 mois, dès que possible
Chargé(e) de mission mixité
Contexte
L’institut interdisciplinaire d’intelligence artificielle (IA) de Toulouse, appelé ANITI (Artificial and Natural Intelligence Toulouse Institute), porté par l’UFTMiP, est l’un des 4 instituts à la pointe de la recherche en intelligence artificielle en France, labélisé par le programme investissements d’avenir PIA3 dans le cadre de l’appel à projets 3IA. Rassemblant environ 200 chercheurs et ingénieurs issus de 12 partenaires académiques et d’une trentaine de partenaires industriels, ANITI vise à développer un modèle original de collaboration permettant d’assurer le développement et l’utilisation pérenne de technologies à base d’intelligence artificielle dans des secteurs d’application stratégiques, par exemple, mobilité / transports, robotique/cobotique pour l’industrie du futur, etc.. L’institut se situe au coeur des problématiques posées aujourd’hui par les technologies d’intelligence artificielle telles que la robustesse, l’explicabilité mais aussi l’acceptabilité par le public, l’éthique et la compréhension de leurs impacts sur la vie de nos concitoyens.
ANITI met un accent particulier sur la sensibilisation du public à ces technologies naissantes en multipliant les initiatives de communication et de diffusion, dans le but de donner une vision équilibrée du potentiel et des risques associés à l’utilisation de l’intelligence artificielle mais aussi d’attirer les jeunes et de renforcer la diversité de genre dans ce domaine d’avenir. En effet, historiquement, les femmes ont occupé une place importante en informatique, puis elles s’en sont éloignées, comme en témoignent des statistiques récentes. La spécialité Numérique et Science Informatique a été choisie en 2020 par 15 à 20% de lycéen.nes, dont moins de 3% de filles et il y a à peine 10% d’étudiantes à l’université en informatique. L’IA pâtit de cette absence de «vocations» féminines et il est urgent de ré-intéresser les filles aux sciences et au numérique en particulier.
Dans cette optique, ANITI a mis en place une commission en vue de contribuer à la diversité de genre en IA. Au-delà d’améliorer le taux de présence féminine dans ANITI, il s’agit de sensibiliser le grand public aux enjeux de l’IA éthique et d’inciter les femmes de tous âges à travailler dans les domaines de l’IA. La commission, qui comprend une dizaine de membres, définit et lance des actions visant à combattre les préjugés sur l’IA et à lutter contre les stéréotypes de genre. Ces actions, ciblant en priorité le milieu scolaire, filles et garçons, car on sait que les stéréotypes sont ancrés dès le plus jeune âge, se font en coordination avec le Rectorat, la Région, les établissements académiques et les associations qui oeuvrent contre les inégalités H/F, ainsi qu’avec nos partenaires industriels.
Missions
Mentor’IA : ce réseau de Mentorat cible les étudiantes en L3, M1 et M2 des universités et écoles du site toulousain. Lancé en juillet 2021, son objectif est de les accompagner durant leurs parcours et faciliter leur intégration dans la vie professionnelle, tout en créant de l’émulation et du lien social entre elles et avec les mentors. L’ambition est aussi de mener des actions communes avec les trois autres instituts labélisés 3IA pour amplifier notre impact.
- Organiser des rencontres du réseau
- Proposer des actions d’animation du réseau
- Réfléchir à des indicateurs d’évaluation et d’amélioration continue du réseau
- Participer aux activités de communication : publics cible, message, vecteurs de communication
Réseau des ambassadeur.rices « Mixité en IA » issu.es des entreprises partenaires d’ANITI a été lancé en février 2021 pour servir de relais et être force de propositions pour des actions collectives.
- Mettre en place une enquête sur les attendus de ce réseau auprès des membres
- Préparer des actions permettant de répondre aux attendus identifiés
- Organiser des rencontres du réseau des ambassadeur.rices «Mixité en IA»
Stages découverte de l’IA en 3e
- Organiser avec les laboratoires de recherche et les partenaires industriels un programme sur mesure pour des collégiennes souhaitant découvrir l’IA dans le cadre du stage découverte de métiers en 3e
- Évaluer l’impact de cette action sur les collégiennes, leur famille, les tuteur.rices de stage
Stratégie diversité en IA : participer à la réflexion de la commission mixité d’ANITI pour évaluer les actions déjà lancées (réseau Mentor’IA, réseau «Mixité en IA», stages découverte de l’IA en 3e, jeu pédagogique pour les lycéen.nes et le grand public sur l’IA et incluant la sensibilisation aux biais, dispositif ASTEP dans les classes de primaire) et définir une ligne d’action pertinente tenant compte de cette évaluation et de nos objectifs.
Capitalisation : rassembler des ressources (MOOC, rapports d’études, enquêtes, sites web, etc.) pouvant être des sources d’information pour les acteurs de la mixité en IA – en collaboration avec les autres 3IA ; éventuellement alimentation de la page web ANITI mixité avec les liens sur ces ressources.
Connaissances et compétences requises
- Connaissance en sociologie du genre
- Connaissance des politiques publiques et des politiques sociales en matière de promotion de l’égalité et de la lutte contre les discriminations
- Concevoir et élaborer des méthodes d’enquête, piloter des outils et des méthodes d’investigation sociologique pour mener à bien la réalisation de ces enquêtes
- Impulser et mettre en place des bilans d’actions promouvant la diversité de genre, évaluer leur impact, diagnostiquer les atouts et les limites de ces actions en tenant compte de la situation et du contexte
- Rendre compte de la conduite des travaux
- Aptitudes au travail en équipe, bon relationnel
- Rigueur et sens de l’organisation
- Maîtrise des outils bureautiques
- Bonnes qualités rédactionnelles
Informations pratiques
Stage conventionné avec gratification – durée souhaitée : 6 mois au 1er semestre 2022
Temps plein – horaires en journée
Localisation géographique : B612 3 Rue Tarfaya, 31400 Toulouse
Pour candidater, envoyez votre CV à Corinne Joffre
Internship :
M2
Durée : 6 mois, au 1er semestre 2022
Spectro-Temporal Feature Extraction for the classification on Satellite Image Time Series with Variational Gaussian Processes
Context
In the last years, the advent of Earth observation satellite missions with short revisit time and increased spatial resolution has led to an unprecedented amount of remote sensing images of heterogeneous physical nature (e.g., optical & radar Sentinel time series . . . ) at various scales (e.g., submetric, decametric . . . ). Furthermore, satellite image archives, such as Spot Heritage, are made available by many space agencies. Such massive data extend existing satellite and in-situ acquisition systems used to understand, explain and predict the states and trends of our environment. However, the novel complexity of the data makes the conventional analytical methods not adapted, and therefore not suitable for extracting and for processing all the relevant information from this massive flow of data.
In order to address challenges raised by such applicative domains, the interdisciplinary institute in artificial intelligence of Toulouse, named the Artificial and Natural Intelligence Toulouse Institute (ANITI) from which the CNES is partner, has been proposed to develop innovative solutions using theoretical advances in core AI scientific areas. The CESBIO lab, with J. Inglada and M. Fauvel, is part of the ANITI Chair entitled “Data-driven approximate Bayesian computation for fusion-based inference from heterogeneous (remote sensing) data” hold by Prof. N. Dobigeon.
A PhD thesis has started in September 2020 to work on large scale Gaussian Processes for the classification of land use and land cover using satellite image time series (SITS). Such methodology offers the possibility of optimizing the model parameters using standard stochastic techniques, and scales well with the size of the data to be processed. Results currently obtained on a large data set show promising results.
Objectives
The objective of this master internship is to benefit from the current architecture to learn feature representations during the training process. Several feature reduction techniques will be investigated to account for the structure of SITS, i.e., the spectro-temporal information, such as random matrix models, functional statistics, deep representation etc . . .
The recruit will implement and test the different approaches using the processing chain developed by Valentine Bellet during her PhD work, and a critical comparison will be done on a large data set.
Developments will be done in Python, using the PyTorch and GPyTorch machine learning libraries.
Resources made available
The recruit will have access to :
- Full stacks of pre-processed Sentinel-2 satellite image time series. — Access to the CNES HPC.
- A laptop and an office at the CESBIO-lab.
Profile
The recruit should have strong knowledge and skills in :
- Machine learning and/or statistical learning,
- Remote sensing,
- Programming and computer science.
Practical information
- Grant will be approximately around 620 euros per month.
- The internship will be located at CESBIO lab (Rond-Point du Professeur Francis Cam-bou 31400 Toulouse).
- Internship duration 6 months, starting from February or March 2022.
- To apply, send a a detailed resume (including grades) and a motivation letter to Valentine Bellet, Jordi Inglada and Mathieu Fauvel
Internship :
M2
Period : This internship shall take place in 2022. The precise starting and ending dates can be adjusted according to the availability of the selected candidate.
Profile & requirements : Master 2 or Engineering school students with major in applied mathematics, computer science or electrical engineering.
Multiband image fusion for astrophysics
Abstract
Recently developed astronomical data fusion methods combine the benefits of multispectral and hyperspectral images to en- hance scientific interpretation of the data. Multispectral images have a high angular (spatial) resolution but are composed of few spectral bands, while hyperspectral images contain very detailed spectra (up to a few thousands spectral points) but measured at lower spatial resolution. The fused product, combining those images, reconstructs the observed scene at high spatial and spectral resolutions. The approach described in those papers consists in formulating an observation forward model of an imager and a spec- trometer, providing respectively multispectral and hyperspectral images. Several conventional penalization have been considered, e.g., based on Sobolev regularization or sparse representation over a dictionary.
The main objective of this internship is to extend these methods to more sophisticated data-driven regularizations. In particular, a particular attention will be paid to so-called deep priors and related approaches.
The performance of the developed methods will be assessed through experiments conducted on simulated data sets or real images provided by ground (e.g., MUSE) and space (e.g., JWST and HST) observatories dedicated to astronomical observations.
Scientific environment
This M.Sc. trainee period will be part of a longstanding collaboration between the “Syste`mes de ́cisionnels” (DISC) group from ISAE-Supae ́ro, the “Signal & Communications” (SC) group from IRIT (CNRS and Toulouse INP) and the “Milieu Interstellaire, Cycle de la Matie`re, Astro-Chimie” (MICMAC) group from IRAP (CNRS, CNES and University of Toulouse). The SC and DISC groups bring their expertise in the development of state-of-the-art signal & image processing methods, in particular for multivalued images for various applications (medical imaging, remote sensing, microscopy). The MICMAC group will bring its expertise on astronomy and astrophysics, in particular in the context of analyzing interstellar matter. Besides, the Early Release Program (ERS 1288) led by Dr. Olivier Berne ́ will favor an easy access to the instrument specifications, as well as privileged interactions with instrument designers.
The M.Sc. student will therefore benefit from a favorable context and will be able to rely on the most recent results and advances in signal & image processing for astronomical data. He/she will be mainly co-advised by
- Thomas Oberlin, Assistant Professor within the DISC group at ISAE-Supaéro
- Nicolas Dobigeon, Professor within the SC group at IRIT laboratory (UMR CNRS 5505, Toulouse) and AI Research Chair at the Artificial and Natural Intelligence Toulouse Institute (ANITI)
- Olivier Berné, CNRS Researcher within the MICMAC group at IRAP (UMR CNRS 5277, Toulouse)
and with a possible collaboration with
- Thierry Contini, CNRS Researcher within the GAHEC group at IRAP (UMR CNRS 5277, Toulouse)
Profile & requirements
Master or Engineering school students with major in applied mathematics, computer science or electrical engineering.
The knowledge needed for this work includes a strong background in signal & image processing and/or machine learning.
Experience and/or interests in astrophysics will be appreciated.
Contact & application procedure
Applicants are also invited to send (as pdf files)
- a detailed curriculum,
- official transcripts from each institution you have attended (in French or English).
to the co-advisors
You will be contacted if your profile meets the expectations. Review of applications will be closed when the position is filled.
Internship :
M.Sc. proposal in signal/image processing
Period : This internship shall take place in 2022. The precise starting and ending dates can be adjusted according to the availability of the selected candidate.
Profile & requirements : Master 2 or Engineering school students with major in applied mathematics, computer science or electrical engineering.
Sparse unmixing for active molecular imaging
Abstract
The capabilities of spectral imaging are certainly attractive in many applications since they offer the opportunity to address complex physical/chemical questions. However, interpreting the resulting images requires a “dual” lens to extract the relevant information encoded in both spatial and spectral domains. While spatial information can be fully exploited through image pro- cessing techniques, this M.Sc. project aims at providing a meaningful spectral decomposition of the imaging data. This will be achieved by analyzing the reduced spectral data set provided by sparse, targeted sampling, based on the rationale that information is typically highly redundant across channels when using any kind of spectral imaging. Such a redundancy is inherent to the nature of the measurement, i.e. the bandwidth of the molecular spectroscopy signals, and introduces high correlations among spectral channels. More importantly, the spectral variation observed in the imaged sample is likely to be attributed to a limited number of sources (individual components), which allows a whole set of high-dimensional measurements to be modeled with a small number of degrees-of-freedom, i.e. to lie in a low-dimensional subspace. We propose to leverage the later property to reach two levels of characterization, namely signal subspace learning and spectral unmixing. The main objective of this M.Sc. project is to design subspace learning and/or spectral unmixing methods dedicated to sparsely sampled images, i.e., from partial and targeted measurements provided by our proposed IMAGIN acquisition protocol.
Scientific environment
This M.Sc. trainee period will be part of the IMAGIN project, funded by ANR. The two main teams involved are the “Signal & Communications” (SC) group from IRIT (CNRS and Toulouse INP) and the “Dynamics, Nanoscopy & Chemometrics” (DyNaChem) group from LASIRE (CNRS and University of Lille). The SC group brings its expertise in the development of state- of-the-art signal & image processing methods, in particular for multivariate images for various applications (medical imaging, remote sensing, microscopy). The DyNaChem group is interested in micro- and nano-imaging of photoactive bio-systems, with a particular focus on hyperspectral and super-resolved nanometer-scale imaging of ultrafast processes, and on the development of new instrumentation and methodologies for the analysis of these hyperspectral and super-resolved data.
The M.Sc. student will therefore benefit from a favorable context and will be able to rely on the most recent results and advances in signal & image processing for molecular imaging. He/she will be mainly co-advised by
- Henrique Goulart, Assistant Professor within the SC group at IRIT laboratory (UMR CNRS 5505, Toulouse)
- Nicolas Dobigeon, Professor within the SC group at IRIT laboratory (UMR CNRS 5505, Toulouse) and AI Research Chair at the Artificial and Natural Intelligence Toulouse Institute (ANITI) in collaboration with
- Cyril Ruckebusch, Professor within the DyNaChem group at LASIRE laboratory (UMR CNRS 8516, Villeneuve d’Ascq)
The physical location for the project is the INP-ENSEEIHT campus (Signal & Communications group), in a lively neighbou- rhood of the Toulouse city center.
Funding
Fully funded by ANR, this M.Sc. position is part of the IMAGIN project. A fully funded Ph.D. position is available for a possible continuation of this M.Sc. training period.
Contact et application procedure
Applicants are also invited to send (as pdf files)
- a detailed curriculum,
- official transcripts from each institution you have attended (in French or English).
to the co-advisors
You will be contacted if your profile meets the expectations. Review of applications will be closed when the position is filled.
Internship :
M.Sc. (M2 or final year engineering project) proposal
in machine learning and/or signal/image processing (with possible continuation as fully funded Ph.D. position)
Period : This internship shall take place in 2022. The precise starting and ending dates can be adjusted according to the availability of the selected candidate.
Profile & requirements : Master 2 or Engineering school students with major in applied mathematics, computer science or electrical engineering.
Smart and sustainable management of karst aquifers
Contexte
Les aquifères karstiques constituent des systèmes au comportement fortement non line ́aire. Cependant, ils sont présents dans toute la région Occitanie notamment et peuvent être le siège de crues dévastatrices ou bien soumis à des étiages de plus en plus sévères en lien avec une pression croissante des besoins en eau. Le stage de master 2 devrait permettre d’initier le de ́veloppement de nouvelles me ́thodes d’apprentissage automatique en vue d’une gestion intelligente et durable des ressources en eau en milieu karstiques en re ́gion Occitanie. Cette pre ́occupation s’est en effet manifeste ́e avec de nombreux conflits d’usage car :
- les eaux souterraines assurent le maintien hydrique des ruisseaux ou cours d’eau et donc par la même assure un support de vie des écosystèmes (conservation des zones humides, biodiversité),
- elles contribuent a` une large part de l’alimentation en eau potable et permettent l’essor d’une activité économique (en particulier dans le secteur tertiaire, industriel, touristique ou énergétique) dont le développement est conditionne par la qualité comme par la quantité de ressources disponibles,
- l’irrigation fortement pre ́sente en re ́gion Occitanie demeure largement tributaire des eaux souterraines.
Objectifs
Ce sujet de stage de master 2 interdisciplinaire s’appuie sur les compétences développées au sein de l’IRIT (Nicolas DOBIGEON) pour la partie apprentissage automatique et au sein du GET (David LABAT) pour la partie hydrogéologie karstique. Ce stage de master 2 s’attachera notamment à proposer des nouvelles méhodes d’analyse supervisées incluant les donne ́es hydro- biogéochimiques mais aussi géophysiques ou spéléologique obtenues sur le bassin du Baget (09). Ainsi, la mise en place de méthodes originales et innovantes en IA permettra d’inclure à la fois des donne ́es temporelles continues et des donne ́es isole ́es tels que des trac ̧ages artificiels mais aussi des donne ́es spatiales issues de la géologie, de la géophysique ou de la spéléologie entre autres.
Environnement scientifique
L’étudiant(e) bénéficiera d’un contexte favorable grâce à l’appui d’une équipe d’encadrants aux compétences complémentaires. Il ou elle sera encadré(e) principalement par
- NicolasDobigeon, Professeur au sein de l’IRIT(UMR CNRS 5505, Toulouse) et Porteur d’une chaire en IA au sein d’ANITI
- David Labat, Professeur au sein du GET (UMR CNRS 5563, Toulouse)
Profil et expertise
Master 2 ou Ecole d’Ingénieur en mathématiques appliquées, informatique ou discipline connexe.
Les connaissances requises pour ce travail incluent une solide formation en apprentissage automatique (statistics, linear algebra,optimization), sciences des données ou traitement du signal. Une expérience et/ou un intérêt pour l’hydrologie serait appréciée.
Contact et candidatures
Les candidat(e)s sont invité(e)s à envoyer (au format pdf) — un CV détaillé,
- les relevés de notes des années de Master 1 et Master 2 aux deux encadrants
- Nicolas Dobigeon
- David Labat
Tout(e) candidat(e) dont le profil répond aux attentes sera contacté(e). L’étude des candidatures s’achèvera dès que le stage sera pourvu.
Stage :
M.Sc. (M2 or final year engineering project) proposal
in machine learning and/or signal/image processing (with possible continuation as fully funded Ph.D. position)
Durée : à définir
Niveau d’études : Master 2 ou MSc
From Biology to Computer Vision: Learning abstract concepts like a honeybee
Although deep neural networks demonstrated a remarkable success in several visual tasks, some significant limitations are becoming increasingly evident in these models. In a recent study, we showed that convolutional neural networks, de facto the current computational models of vision, struggle to accomplish simple visual reasoning tasks, such as judging whether two items are the same or not. Arguably, implementing neural networks able to solve these same-different tasks would be an essential step toward realizing models truly able to learn abstract representations and eventually conceptual learning. One way to address this challenge is to take advantage of biological-inspired mechanisms to obtain better visual reasoning models. Interestingly, previous studies demonstrated that the honeybee’s brain could learn to associate equal (or different) stimuli, and generalize this association to novel stimuli, thus learning the concept of the sameness. These results have been replicated in biologically plausible models, which were able to perform the same-different tasks implementing architectures inspired by structures present in the honeybee’s brain.
This project aims at translating these computational solutions, inspired by the honeybee’s brain, into current state-of-the-art deep neural networks. We aim specifically at leveraging previous computational studies modeling the mushroom bodies and translate these biologically plausible models in current deep neural networks, eventually proposing a novel architecture capable of abstract learning.
This M2-level internship is funded for up to 6 months. It will take place at CerCo, in the NeuroAI team, in a very stimulating environment for both machine learning and computational neuroscience. The project also involves a collaboration with CRCA in Toulouse, a lab with expertise in honeybee cognition and neuroimaging. The candidate must have programming skills in python (familiarity with machine learning and pytorch is a plus), and an interest in biologically inspired AI.
Contact (In case of interest, please send your CV and motivation letter to Rufin VanRullen.
Intership
Durée : up to 6 months
Niveau d’études : Master 2
Implementation of a Big Data processing platform for the quantification of the safety impacts of HW failures on FPGA-based CNN implementation
Machine learning-based applications, and in particular neural networks, are becoming widespread in the automotive domain. For some embedded systems, such as obstacle detection for autonomous cars, the neural network is involved in safety critical applications and executed on board. This means that the implementation must ful ll several properties such as real-time guarantees (i.e. the WCET { Worst Case Execution Time { must be computable) and safety (i.e. hardware random failure must be detected and mitigated). In the automotive domain, those properties are summarized in the standard ISO 26262 [3].
To address this challenge, chip makers like NXP prepare the next generation of chips to be used for vision-based computer, automotive and autonomous driving or radar / lidar computing. Those chips will integrate neural network hardware IP and should ensure the same level of safety as expected by the ISO 26262. To demonstrate this level of safety, the applicant must ensure that the likelihood of hazardous outcomes due to HW failures complies to the requirement provided in the ISO26262. The quantification of this likelihood can typically be achieved through a thorough fault injection process on the IP executing the CNN. Such a process has been developed and implemented in the context of [2]. Once the campaign has been conducted, a post-processing phase should be conducted to analyze the impact of injected faults on the CNN execution.
The objective of the internship is to implement an efficient processing platform extracting the relevant indicators out of the fault injection results. To do so, the first step of the internship will be to identify and formalize the indicators to be extracted from the fault injection results. The second step will be to design the data processing platform to compute these indicators. Due to the large amount of fault injection locations and tested inputs, the processing platform must be able to deal with a huge amount of fault injection results thanks to the Big Data framework Apache Spark [1]. Eventually, the processing platform will be validated on a comprehensive fault injection campaign (100Go) performed on the LeNet5 CNN trained on the MNIST data-set.
Collaborations: this internship is funded by ANITI and conducted within a collaboration with NXP and ONERA.
Application procedure
Formal applications should include detailed CV, a motivation letter and transcripts of bachelors’ degree.
Applications should be sent by email to: advisor email
Références
[1] Apache Spark framework. URL: https://spark.apache.org.
[2] Certi ableAI chair of ANITI . URL: https://aniti.univ-toulouse.fr/chaire-claire-pagetti.
[3] ISO. ISO 26262 Road vehicles | Functional safety, 2018. URL: https://www.iso.org/obp/ui/#search.
Implémentation du jeu d’instruction Kalray K3 dans l’analyseur statique OTAWA
L’analyse de pire-temps d’exécution (WCET) de programmes est une étape nécessaire à la construction et à la validation temporelle des systèmes embarqués temps-réel. Pour les applications les plus critiques nécessitant une certification, cette analyse est réalisée statiquement et directement sur la version du programme compilée pour son processeur cible (fichier binaire). Pour réaliser cette analyse, le programme est désassemblé automatiquement et son graphe de flot de contrôle est régénéré à partir des informations contenues dans le binaire. Cette étape nécessite que le jeu d’instructions (ISA) du processeur cible soit décrit dans le programme de désassemblage.
L’objectif de ce stage est de permettre l’analyse du jeu d’instructions du cœur K3 présent dans le processeur Kalray Coolidge [1] par l’outil d’analyse WCET OTAWA [2] développé et maintenu à l’Institut de Recherche en Informatique de Toulouse (IRIT). Il s’agira donc premièrement d’encoder le jeu d’instructions (correspondance binaire/assembleur et sémantique) dans le langage du désassembleur d’OTAWA à partir des informations contenues dans le manuel du cœur K3, ou d’interfacer directement OTAWA avec un simulateur Qemu du cœur et de compléter les propriétés sémantiques des instructions. Dans un second temps, il s’agira de modéliser les principales caractéristiques temporelles du cœur (profondeur du pipeline, latence des étages, latence des mémoires, comportement des accélérateurs ML, etc.) dans OTAWA.
Suite à ces travaux, il devrait être possible de désassembler et d’analyser des programmes implémentant des réseaux de neurones [3] sur le processeur Kalray Coolidge.
Profil du candidat
Nous recherchons des candidats F/H niveau M1/M2 motivés et possédant des compétences en programmation C/C++, en assembleur (par exemple x86 ou ARM) et en architecture des processeurs.
Le stage sera réalisé à l’IRIT sur le campus de l’université Toulouse 3 Paul Sabatier, et co-encadré par des chercheurs de l’ONERA, d’Airbus et de l’IRIT.
Contacts
Références
[1] Benoît Dupont de Dinechin, Julien Hascoët, Julien Le Maire, Nicolas Brunie. Deep Learning inference on the MPPA3 Manycore Processor. In Embedded world, 2020.
[2] Clément Ballabriga, Hugues Cassé, Christine Rochange, Pascal Sainrat. OTAWA : An Open Toolbox for Adaptative WCET Analysis. In SEUS, 2010.
[3] Iryna De Albuquerque Silva, Thomas Carle, Adrien Gauffriau, and Claire Pagetti. Automatic predictable C code generation of machine learning models for avionics systems. In French Real-Time Summer School, 2021.
Internship
Duration : 5-6 months
Jeux et décision sous incertitude : application aux jeux de coordination
La théorie des jeux se propose d’étudier des situations (appelées « jeux ») où des agents (les « joueurs ») prennent des décisions (on parle de « strategies ») , chacun étant conscient que le résultat de son propre choix (son « gain ») dépend de celui des autres. Si les joueurs peuvent passer des accords le jeu est dit jeu coopératif. Si c’est non (par exemple parce que les agents ne peuvent pas communiquer, ou parce qu’il n’est pas possible de garantir un accord ), le jeu est dit non coopératif. La distinction est fondée sur le contexte, pas du tout sur la manière de jouer. Il ne faut pas confondre jeu de coordination et jeu de coopéation. Dans un jeu de coopération, les joueurs ont la possibilité de se concerter afin d’obtenir le meilleur paiement. Ce n’est pas le cas dans un jeu de coordination, où les joueurs ont interêt à jouer la même action (c’est à dire à se coordonner), car ils ont la même fonction de paiement, mais sans pouvoir se concerter. L’utilisation d’un réseau social est un exemple typique de jeu de coopération dans une grand graphe de joueurs, certains ne connaissant même pas l’existence d’autres : on a intérêt à utiliser le même outil que ses proches, lesquels ont des accointances avec l’autres etc.
Dans ce stage on considérera des jeux à issue incertaine, où les gains dependent non seulement des stratégies/décisions des joueurs mais aussi circonstances extérieures sur lesquelles ils ont une connaissance limitée (par exemple le niveau sécurité des données dans l’outil utilisé, son exposition aux attaques, etc), commune ou pas – typiquement, le résultat des actions jointes est incertain. .
Dans le cadre de ce stage, on mènera une étude des jeux de cooperation sous incertitude probabiliste. Il s’agira dans une premier temps de developper des algorithmes pour les jeux de cooperation Bayesiens où les interactions entre joueurs sont pair à pair – dans ce cas, il est en effet théoriquement possible de diminuer la complexité du problème de recherche d’équilibres de Nash en donnant une formulation plus compacte du problème.
Dans ce second temps, on étendra cette étude aux jeux de cooperation non Bayesiens – il existe en effet de nombreuses situations où l’utilité espérée n’est pas (ou ne peut pas être) utilisée, car la connaissance ne peut être exprimée par une unique distribution, comme mis en evidence par le paradoxe de Ellsberg . On se placera alors dans le cadre de la théorie des fonctions de croyance pour developper une nouvelle approche des jeux de coordination sous incertitude.
Nous recherchons donc, pour un stage de 4 à 6 mois, un/e candidat/e en cours de Master II ou de dernière année d’ecole d’ingénieur, intéressé par la décision sous incertitude ; des connaissances en optimisation (e.g.PLNE, Solveurs SAT/CSP) seraient appréciees
Ces recherches prennent place dans le cadre de la chaire « Compilation de Connaissances »
Contact (Hélène Fargier (helene.fargier@irit.fr)) – IRIT/ ADRIA –
Stage de recherche
Durée : 4 à 6 mois
Niveau d’études : Master 2
Scheduling robotic tasks for industrial applications

The automation of robotic tasks in the aviation industry requires the development of techniques combining the automatic generation of movement elements to execute tasks with the optimal scheduling of these tasks. The Gepetto team at LAAS has a long experience in motion planning for systems with a large number of degrees of freedom in environments occupied by obstacles.
The ROC team has a long experience in combinatorial optimization problems such as task scheduling, which is also the subject of the KC@ANITI chair.
The automation of robotic tasks in the aviation industry requires the development of techniques combining the automatic generation of movement elements to execute tasks with the optimal scheduling of these tasks. The Gepetto team at LAAS has a long experience in motion planning for systems with a large number of degrees of freedom in environments occupied by obstacles. The ROC team has a long experience in combinatorial optimization problems such as task scheduling, which is also the subject of the KC@ANITI chair.
The objective of this internship is to combine these techniques in order to produce
movements performing robotic tasks in optimal order. Two difficulties specific to this combination are:
- the a priori ignorance of the cost of transition from one task to another,
- the fact that a given task can be performed by an infinite number of configurations of the robot due to its kinematic redundancy.
The work will be illustrated by an application of drilled holes deburring in an Airbus A 320 reactor mast (picture above).
REQUIRED SKILLS
The candidate must have programming skills in python. Knowledge of C++ programming and combinatorial optimization will be appreciated.
Stage de recherche
Durée : 4 à 6 mois
Niveau d’études : Master 2
Détection et délinéation automatique des parcelles de culture à partir des données Sentinel2
Pour les besoins des activités de recherche en cartographie de l’occupation du sol et de suivi des cultures réalisées au CESBIO (www.cesbio.cnrs.fr), il serait utile de disposer d’une méthodologie automatique permettant, à partir d’une série temporelle d’images Sentinel2, de produire une base de données géographique des parcelles agricoles.
> Consultez l’annonce en ligne sur le site du Cnes
Stage ingénieur
Durée : 4 à 6 mois, temps plein
Niveau d’études : Master, DESS, DEA, Bac+5/
Prototyping tools for digital democracy
We are looking for Master students interested in the development of technologies for civic participation. Candidates should be passionate about the design and development of online systems, and should think critically about the social impact of technologies. Candidates should be interested in working in a multidisciplinary environment combining computer scientists, lawyers, UI/UX designers, natural and social scientists. We are looking for candidates with web development skills (e.g. Javascript, React, Node, etc.), design sensibilities, and/or the ability to critically analyze data using statistical and graphical methods. The working language of the group is English.
Formal applications should include a CV, letter of motivation, a list of three possible recommenders, and links to previous work. They should be directed to professor Cesar Hidalgo (cesar.hidalgo@univ-toulouse.fr). The work will be performed at ANITI’s Collective Learning group and is supported by ANITI’s Augmented Society chair.
Stage ingénieur /
Durée : 4 à 6 mois, temps plein
Niveau d’études : Master 2
Advanced data-driven techniques for the Lasserre hierarchy
A large number of problems from diverse fields such as optimization, probability and statistics, dynamical systems and control or quantum physics can be tackled within the unied and powerful framework of the Lasserre hierarchy [1, 2], which allows one to solve challenging nonconvex and nonlinear problems by a sequence of convex optimization problems, in a uni_ed and very systematic fashion. Additional research [5] investigated the ability of Christo_el-Darboux kernels to capture information about the support of an unknown probability measure. A distinguishing feature of this approach is to allow one to infer support characteristics, based on the knowledge of infinitely many moments of the underlying measure. A major open question remains whether this approach can be used in a data-driven setting, where the underlying model is unknown and only observed data are available. This project will investigate this direction, building on the very recent work [3]. Progress in this direction would be an enabling factor in bringing the elegant and powerful tools of the Lasserre hierarchy to the realm of the present-day big-data applications, which are currently typically tackled using ad-hoc heuristic techniques with limited mathematical foundation.
The goal of this M2 internship is the extension of the Lasserre hierarchy and Christo_el-Darboux kernels to a data-driven setting, developing new methods furnished with a theoretical analysis (convergence rate, non-asymptotic out-of-sample error etc). One first possible investigation track will consist of studying Christo_el-Darboux kernels to extend the approach from [4] for measures supported on specic classes of mathematical varieties. We intend to apply this framework to deep learning network models, for which latent representation correspond to such low-dimensional varieties. Numerical experiments will be performed on several benchmark suites, including MNIST, CIFAR10 or fashion MNIST. Other main steps may include, but are not limited to, the investigation of adaptive sampling techniques and basis choice for the approach developed in [3] as well as extension of the proposed methodology beyond the problem class considered, e.g., to data-driven optimal control. Other directions include complexity reduction based on sparsity, symmetries and other more advanced structural properties of the problem at hand. A significant degree of freedom will be given to the student to create and pursue his/her ideas broadly within this scope.
Requirements: A successful M2 candidate will have a strong background in applied mathematics or physics, having a very good knowledge of probability and statistics as well as a working knowledge of convex optimization, real analysis and basic measure theory. The candidate should be highly motivated and creative.
Funding: This M2 internship will be funded by the arti_cial intelligence center of Toulouse (ANITI) and co-supervised with LAAS CNRS and University Toulouse 3, Paul Sabatier. A PhD can be foreseen. The M2 candidate will be hosted by LAAS and the DEEL team (DEpendable and Explainable Learning), a joint project team between Toulouse and Montr_eal, in building B612, IRT Saint Exupery.
References
[1] J.B. Lasserre (2001). Global optimization with polynomials and the problem of moments. SIAM
Journal on optimization, 11(3), 796-817.
[2] J.B. Lasserre (2010). Moments, positive polynomials and their applications. World Scientific, 2010.
[3] M. Korda (2019). Data-driven computation of the maximum positively invariant set for nonlinear dynamical systems. FEANICSES workshop, https://cavale.enseeiht.fr/feanicses/files/
W2019/4-korda.pdf.
[4] E. Pauwels, M. Putinar and J.-B. Lasserre (2019). Data analysis from empirical moments and the Christo_el function. hal-01845137
[5] E. Pauwels and J.-B. Lasserre (2019). The empirical Christo_el function with appication in data analysis. To appear in Advances in Computational Mathematics.
Stage ingénieur /
Durée : 4 à 6 mois, temps plein
Niveau d’études : Master 2
Estimation of large-dimensional tensor models and applications in machine learning

SUPERVISORS
The internship will be supervised by Prof. H. Goulart (henrique.goulart@irit.fr) at IRIT/ENSEEIHT, in remote collaboration with two members of the LargeDATA (DataScience) chair at 3IA MIAI/Univ. Grenoble-Alpes : Prof. R. Couillet (head, romain.couillet@gipsa-lab.grenobleinp.fr) and Dr. P. Comon (pierre.comon@gipsa-lab.grenoble-inp.fr).
FUNDING
This internship will be funded by the ArtiFicial and Natural Intelligence Toulouse Institute (3IA ANITI), as part of the AI Research Chair lead by N. Dobigeon.
CONTEXT
Tensor models are powerful tools for addressing many problems in signal processing, machine learning and beyond [1]–[3]. Yet, their use in these applications typically requires estimating a low-rank tensor from a set of observations corrupted by noise, which is often a difficult task. Moreover, in most cases there is currently no theory for predicting the actual estimation performance that can be attained.
To overcome this gap, in recent years several researchers have studied the asymptotic statistical performance of ideal and practical estimators in the large-dimensional regime,
where the size of the tensor grows large [4]–[6]. In particular, these works have uncovered the abrupt phase transition that the performance of an ideal estimator may undergo as the
signal-to-noise ratio grows (see Figure 1 below for an illustration). While some important advancements have been achieved, many scenarios of practical interest remain unexplored, as well as the practical implications of the existing results in applications.
OBJECTIVES
The overall goal of this internship is to study extensions and applications of the existing results, as a first step for pushing the existing theory beyond its current limits. We will in particular
consider extensions to more general tensor models that apply to larger classes of realworld problems, including e.g. asymmetric models. Application to practical machine learning problems—such as community detection in hypergraphs [7], latent variable model estimation [2] and high-order co-clustering [8]—will also be considered.

The intern will initially perform computer simulations aimed at understanding the behavior of ideal and practical estimators in the target scenarios/applications. Some theoretical results may then be derived on the basis of these experimental findings. Scientific dissemination of these findings will also be encouraged, via publication of papers and/or participation
in scientific events.
A PhD thesis may be proposed to the intern at the end.
CANDIDATE PROFILE
We look for strongly motivated candidates with a solid background on mathematics and statistics, having good programming skills in scientific computing languages (Python, Matlab, Julia). Basic knowledge or interest in random matrix theory is a strong plus.
PRACTICAL INFORMATION
• The intern will be hosted at the ENSEEIHT site of IRIT, located at 2 rue Charles Camichel,Toulouse.
• The monthly internship gratification is of 600€.
• Application procedure: please send your CV, your report card and a motivation letter to Henrique Goulart
Contact : Henrique Goulart – henrique.goulart@irit.fr
References
[1] N. D. Sidiropoulos, L. De Lathauwer, X. Fu, et al., “Tensor decomposition for signal processing and machine
learning,” IEEE Transactions on Signal Processing, vol. 65, no. 13, pp. 3551–3582, 2017.
[2] A. Anandkumar, D. Hsu, S. M. Kakade, et al., “Tensor decompositions for learning latent variable models,”
Journal of Machine Learning Research, vol. 15, pp. 2773–2832, 2014.
[3] S. Rabanser, O. Shchur, and S. Günnemann, “Introduction to tensor decompositions and their applications
in machine learning,” arXiv preprint arXiv:1711.10781, 2017.
[4] A. Jagannath, P. Lopatto, and L. Miolane, “Statistical thresholds for Tensor PCA,” The Annals of Applied Probability,
vol. 30, no. 4, pp. 1910–1933, 2020.
[5] A. Perry, A. S. Wein, and A. S. Bandeira, “Statistical limits of spiked tensor models,” Annales de l’Institut Henri
Poincaré, Probabilités et Statistiques, vol. 56, no. 1, pp. 230–264, 2020.
[6] A. Montanari and E. Richard, “A statistical model for tensor PCA,” Advances in Neural Information Processing
Systems, vol. 4, pp. 2897–2905, 2014. eprint: 1411.1076.
[7] C. Kim, A. S. Bandeira, and M. X. Goemans, “Community detection in hypergraphs, spiked tensor models, and
sum-of-squares,” in 2017 International Conference on Sampling Theory and Applications (SampTA), Tallinn,
Estonia, Jul. 2017, pp. 124–128.
[8] E. E. Papalexakis, N. D. Sidiropoulos, and R. Bro, “From k-means to higher-way co-clustering: Multilinear
decomposition with sparse latent factors,” IEEE Transactions on Signal Processing, vol. 61, no. 2, pp. 493–
506, 2013.
Master 2 internship proposal
Spring 2021 (5 to 6 months)
Representation learning for satellite image
time series classification

CONTEXT
In the last years, the advent of Earth observation satellite missions with short revisit time and increased spatial resolution has led to an unprecedented amount of remote sensing images of heterogeneous physical nature (e.g., optical & radar Sentinel time series . . . ) at various scales (e.g., submetric, decametric . . . ). Furthermore, satellite image archives, such as Spot Heritage, are made available by many space agencies. Such massive data extend existing satellite and in-situ acquisition system used to understand, explain and predict the states and trends of our environment.
In order to address challenges raised by such applicative domains, the interdisciplinary institute in artificial intelligence of Toulouse, named the Artificial and Natural Intelligence Toulouse Institute (ANITI) of whom CNES is partner, has proposed to develop innovative
solutions using theoretical advances in core AI scientific areas. The CESBIO lab, with J. Inglada and M. Fauvel, is part of the ANITI Chair entitled “Data-driven approximate Bayesian computation for fusion-based inference from heterogeneous (remote sensing) data” hold by Prof. N.Dobigeon. This chair’s objective is to develop learning algorithms able to extract meaningful information from multi-source, multi-scale and multi-temporal data.
For 4 years now, CESBIO has used machine learning tools to provide automatically the fist land cover map of France from satellite image time series (http://osr-cesbio.ups-tlse. fr/~oso/ and https://www.theia-land.fr/en/ceslist/land-cover-sec/). The master internship is related to the integration of massive multi-source/scale satellite image time series
in learning algorithms for the production of such land use and land cover maps.
OBJECTIVES
The goal of the internship is to investigate the definition of an auto-encoder architectures adapted to satellite image time series. Such data are structured into three dimensions:
1 spatial, spectral and temporal. Hence, each pixel of a given data cube can be represented as a collection of spectral measurements over the time and spatial domains (see figure 1 for instance).

to clouds and shadows.
While several neural architectures have been proposed in the remote sensing literature, most of them have been applied to hyperspectral images of reduced size. Thus, they hardly take into account the temporal dimension and do not scale well w.r.t. the number of pixels.
Therefore, there is a need for specific work in satellite image time-series analysis. The recruited candidate will do a review of auto-encoder architectures used in satellite image times series and will implement the appropriate ones. Tests on large scale data sets will be performed for the classification of land cover types. Depending on the advancement, modification of existing architectures will be considered.
The candidate will be located at the “Maison de la recherche et de la valorisation” close to the CESBIO lab. He/She will be with other master students within the ANITI project. He/She will benefit of a full access to CESBIO and CNES high performance computers.
CANDIDATE PROFIL
The candidate must have a solid background in at least one of the following subjects:
- Statistical signal and image processing,
- Machine learning,
- Remote sensing data processing.
A good knowledge of English and scientific programming (Python, C/C++) is required.
PRACTICAL DETAILS
- The grant is approximately 620 euros per month.
- Beginning of the internship: February or March 2021, for 6 months.
- To apply, send a CV and a motivation letter to Mathieu Fauvel, mathieu.fauvel@inrae.fr.
To apply, send a CV and a motivation letter to Mathieu Fauvel : mathieu.fauvel@inrae.fr.
Contacts : Matthieu Fauvel (UMR CESBIO, INRAE Toulouse) & Jordi Inglada (UMR CESBIO, Cnes Toulouse)
Master internship
Blind inverse problems in microscopy
TOPIC
Most of the significant advances in imaging in recent years rely on the use of numerical computation. Examples include super-resolution in microscopy, compressed sensing in magnetic resonance imaging or radar interferometry for earth observation. These techniques
share the prior need to design a mathematical model of the observation system. The acquired data is often described by an equation of the form y = Ax where A is a linear operator that describes the physics of the acquisition. Recovering x then requires inverting the operator. Although these imaging modalities have allowed solving problems beyond reach in the past, their expansion is often severely limited by an important problem: it is impossible to accurately control imaging conditions, resulting in errors on the operator A that make the numerical methods unreliable.
The main objective of this project is to develop theories and algorithms to overcome these difficulties and apply them to microscopy and astronomical imaging.
This problem is one of the most important current challenges in the field of inverse problems. Depending on the candidate interest, it will be possible to concentrate the efforts on rather theoretical issues or practical issues.
THEORY
The supervisors have begun investigating various open questions in this field. Two of them will be explored.
Recently [4], Kech and Krahmer provided necessary and sufficient conditions guaranteeing the ability to recover the operator A and the signal x from the measurements y. This implies studying the global injectivity of a bilinear mapping. The setting is however too
restricted for practical applications and we plan to study the local injectivity of the mapping.
A second trail is about the design of optimization algorithms to recover A and x. A standard approach [1] consists in lifting the problem on the space of rank-1 matrices. A traditional approach to solve the resulting problem consists in relaxing the rank 1 constraint by penalizing the nuclear norm. We recently observed that simpler projected gradient descents [3] tend to produce better results. Following [5], we plan to explore the use of continuous relaxations, which discard spurious minimizer and simplify the minimization.
APPLICATION
The supervisors of this project have begun a collaboration with the Centre de Biologie Intégrative in Toulouse (CBI), to improve the resolution of microscopes of the imaging facility. In particular, they are interested in the field of Single Molecule Localization Microscopy (SMLM) [2] and Structured Illumination microscopy [6].
The dominant approach to solve the associated inverse problems consists in assuming that the observation operator has been perfectly calibrated before the acquisition. This hypothesis is seldom satisfied and we expect to significantly improve the reconstructions by
using a finer modelling of the system and an identification of the operator. In addition, real improvements of the microscopes might result in new collaborations with the biologists to answer practical problems.
PRACTICAL ASPECTS
We are looking for a highly motivated student, willing to continue with a PhD thesis, with a background in electrical engineering (signal/image processing, harmonic analysis) or mathematics (optimization, probability an statistics, geometry). Strong abilities in mathematics or computer sciences will be appreciated. A taste for optics or biological
problems is a plus since applied projects might emerge.
If the candidate is successful, this internship will be pursued by a PhD. Various funding sources will be considered (ADI, FRM, Labex CIMI, doctoral school, ANR, ANITI).
This internship will take place either at IRIT or at INSA (in Toulouse, France), depending on the candidate’s interest. It will be co-supervised by Emmanuel Soubies (CR Toulouse) and Pierre Weiss (CR Toulouse). In addition, the candidate will have access to the CBI (Centre de Biologie Intégrative) with Thomas Mangeat for an access to the microscopy resources and biological problems. Do not hesitate to contact us for more information.
BIBLIOGRAPHY
[1] Ali Ahmed, Benjamin Recht, and Justin Romberg. Blind deconvolution using convex programming. IEEE Transactions on Information Theory, 60(3):1711–1732, 2013.
[2] Valentin Debarnot, Paul Escande, Thomas Mangeat, and Pierre Weiss. Learning low- dimensional models of microscopes. IEEE Transactions on Computational Imaging, 2020.
[3] Valentin Debarnot and Weiss Pierre. Blind inverse problems with isolated spikes. ArXiv, 2020.
[4] Michael Kech and Felix Krahmer. Optimal injectivity conditions for bilinear inverse problems with applications to identifiability of deconvolution problems. SIAM Journal on Applied Algebra and Geometry, 1(1):20–37, 2017.
[5] Emmanuel Soubies, Laure Blanc-Féraud, and Gilles Aubert. A continuous exact \ell_0 penalty (cel0) for least squares regularized problem. SIAM Journal on Imaging Sciences, 8(3):1607–1639, 2015.
[6] Emmanuel Soubies and Michael Unser. Computational super-sectioning for single-slice structuredillumination microscopy. IEEE Transactions on Computational Imaging, 5(2):240–250, 2018.
APPLICATION PROCEDURE
Formal applications should include detailed cv, a motivation letter, and transcripts of Bachelor’s degree.
Applications should be sent by email to: advisor emails
Contacts : Emmanuel Soubies – emmanuel.soubies@irit.fr & Pierre Weiss – Pierre.armand.weiss@gmail.com
Stage de recherche
Durée : 4 à 6 mois
Niveau d’études : Master 2