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

Laboratoire/Entreprise : Sorbonne Universite – Institut des Systèmes Intell
Durée : 36 mois
Contact : patrick.gallinari@sorbonne-universite.fr
Date limite de publication : 2024-12-30

Contexte :
Physics-aware deep learning aims at investigating the potential of AI methods to advance scientific research for the modeling of complex natural phenomena. This is a fast-growing research topic with the potential to boost scientific progress and to change the way we develop research in a whole range of scientific domains. An area where this idea raises high hopes is the modeling of complex dynamics characterizing natural phenomena occurring in domains as diverse as climate science, earth science, biology, fluid dynamics. A diversity of approaches is being developed including data-driven techniques, methods that leverage first principles (physics) prior knowledge coupled with machine learning, neural solvers that directly solve differential equations. Despite significant advances, this remains an emerging topic that raises several open problems in machine learning and application domains. Among all the exploratory research directions, the idea of developing foundation models for learning from multiple physics is emerging as one of the fundamental challenges in this field. This PhD proposal is aimed at exploring different aspects of this new challenging topic.

Sujet :
Foundation models have become prominent in domains like natural language processing (GPT, Llama, Mistral, etc) or vision (CLIP, DALL-E, Flamingo, etc). Trained with large quantities of data using self-supervision, they may be used or adapted for downstream tasks while benefiting through pre-training from large amounts of training data. Initial attempts at replicating this framework in scientific domains is currently being investigated in fields as diverse as protein, molecule, weather forecasting. Is the paradigm of foundation models adaptable to more general physics modeling such as the complex behavior of dynamical systems? Large initiatives are emerging on this fundamental topic (http://micde.umich.edu/SciFM24, https://iaifi.org/generative-ai-workshop). This high stake, high gain setting might be the next big move in the domain of data-driven spatio-temporal dynamics modeling. The objective of the PhD is to explore different directions pertaining to the topic of foundation models for physics, focused on the modeling of dynamical systems.

**Solving parametric PDEs

A first step is to consider solving parametric partial differential equations (PDEs), i.e. PDEs from one family with varying parameters including initial and boundary conditions, forcing functions, or coefficients. Current neural solvers operate either on fixed conditions or on a small range of parameters with training performed on a sample of the parameters. A first direction will be to analyze the potential of representative NN solvers to interpolate and extrapolate out of distribution to a large range of conditions when learning parametric solutions. A key issue is then the development of training techniques allowing for fast adaptation on new dynamics.

**Tackling multiple physics

The foundation approach is particularly interesting in the case of scarce data, provided physics primitive could be learned from related but different PDE dynamics that are available in large quantities and then transferred to the case of interest. Learning from multiple PDEs raises algorithmic challenges since they operate on domains with different space and time resolutions, shapes and number of channels. We will consider an Encode-Process-Decode framework so that the commonalities between the dynamics are encoded and modeled in a shared latent space and the encoding-decoding process allows to project from and to the observation space for each PDE. This framework will be evaluated with selected backbones.

**Generalization and few shot capabilities

Generalization to new dynamics is the core problem motivating the development of foundation models in science. This is a key issue for the adoption of data-driven methods in physics and more generally in any context were the data is scarce. We will consider the general framework of few shot learning aiming at fine tuning pre-trained models for downstream tasks. In this context the objective will be to develop frameworks for the fast adaptation of foundation models to target tasks. Different strategies will be analyzed and developed including parameters sampling, meta-learning for adaptation and strategies inspired from the developments in semantics and language applications like in-context.

Profil du candidat :
Computer science or applied mathematics. Good programming skills.

Formation et compétences requises :
Master degree in computer science or applied mathematics, Engineering school. Background and experience in machine learning.

Adresse d’emploi :
Sorbonne Université (S.U.), Pierre et Marie Campus in the center of Paris. The candidate will integrate the MLIA team (Machine Learning and Deep Learning for Information Access) at ISIR (Institut des Systèmes Intelligents et de Robotique).

Document attaché : 202404261431_2024-04-20-PhD-Description-Foundation-models-Physics.pdf

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

Laboratoire/Entreprise : IMT Toulouse et IPAL/ASTAR SIngapour
Durée : 3 ans
Contact : caroline.chaux@cnrs.fr
Date limite de publication : 2024-05-15

Contexte :
Nous proposons un sujet de thèse France – Singapour pour lequel la moitié de la thèse aura lieu en France (ITM Toulouse) et l’autre moitié à SIngapour (IPAL IRL CNRS 2955).

Sujet :
The goal of this PhD thesis is to deploy the 3D PAT scanner designed by
Jérôme Gateau at the Laboratoire d’Imagerie Biomédicale (LIB) for a routine use in biomedical studies.
This will be achieved by designing fast reconstruction methods that provide high-quality results. Depending on the candidate interests, the following axes would be considered:
• Designing implementations of the forward and adjoint models (modeled by matrix-vector products) that are fast and that incorporate the SVIR of the detector. During their preliminary works, the
partners have identified a promising approximation method of A together with the actual reconstruction algorithms. The method based on the Fourier Integral Operator form of the wave propagation equation, should be able to scale high-quality reconstructions to real data. Other types of approximation
could also be considered such as Hierarchical matrices or tensor-train decomposition.
• Designing reconstruction algorithms based on deep neural networks: such as Plug-and-Play methods or algorithm unrolling.
• Implementing an optimized high-parallel (GPU) version of the algorithms to meet with the time requirements of routine use.
• Designing automatic fine-tuning methods of the hyper-parameters involved in these reconstruction algorithms and the calibration of the parameters of A.

One outcome of this PhD project is a photoacoustic scanner that simultaneously combine, compared to standard reconstruction methods, (i) shorter acquisition times, (ii) a reconstructed image of higher resolution and contrast, and (iii) shorter computation times. This could have a great impact on the PAT community which in turn will benefit the clinical and biological communities. The candidate will be trained and could develop skills in optimization, image processing, machine learning, high performance computing and approximation theory. These competences are actively being in demand in the industry and the academic research.

Profil du candidat :
Master of computer science or applied mathematics with strong skills in signal/image processing, optimization, machine learning and numerical computations. Languages: Python/Matlab, C++/ CUDA.

Formation et compétences requises :
Master of computer science or applied mathematics with strong skills in signal/image processing, optimization, machine learning and numerical computations. Languages: Python/Matlab, C++/ CUDA.

Adresse d’emploi :
IMT Toulouse
IPAL Singapour

Document attaché : 202404260923_2024_PhD_offer_IPAL.pdf

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

Laboratoire/Entreprise : PRISME Lab Universiy of Orléans / BRGM Orléans
Durée : 36 months
Contact : yves.lucas@univ-orleans.fr
Date limite de publication : 2024-06-30

Contexte :
Scientific Background

In remote sensing, the intensive production of multi-sensor satellite and airborne data of ever-increasing spatial resolution (visible, IR, hyperspectral, lidar, radar, topography, spectral material libraries, etc.) enables very detailed observation of the earth. In particular, ENMAP satellite has opened up to the scientific community a considerable field of investigation for earth observation with a spatial resolution of 30m. This potential remains under-utilized, however, as conventional methods are unable to absorb such a mass of data, especially hyperspectral imagery, which extends over hundreds of bands.
Artificial intelligence techniques, which have revolutionized the field of Computer Vision, are opening up a new avenue in remote sensing for semantic segmentation, with the automatic extraction of characteristics of features exposed to natural hazards. With climate change, natural disasters are on the increase, demonstrating the urgent need to establish up-to-date risk scenarios.

The aim of this thesis is to evaluate the contribution of artificial intelligence to better assess vulnerability in the face of natural hazards, by unfolding impact scenarios from a multi-risk, multi-scale perspective.

The highly multimodal and heterogeneous nature of the data collected by remote sensing to characterize a territory has given rise to a new methodological challenge: developing suitable network architectures for the classification and semantic segmentation of this massive and complex data. What’s more, the lack of remote sensing training databases is driving work towards semi-supervised approaches with partially annotated data. It is through the pooling of heterogeneous data proposed in this thesis that ground truth will be substantially enriched. Network models will also have to adapt to degraded situations abroad, where some data are unavailable.

This work is closely linked to the Région Centre Val de Loire CERES project – Mapping and characterizing exposed elements in the CVL region from satellite images – with application prospects for the region’s economic operators, concerned by the growing risks associated with flooding and building cracking. CERES is in charge of access to paid data and intensive online computing on the deep learning models developed during the thesis. This work is also in synergy with actions carried out at BRGM (H2020 COCLICO, VIGIRISKS, ANR RESIFLEX) and the ANR-IA, where joint work is underway between PRISME and BRGM.

Funding

Region Centre – Val de Loire thesis grant (36 months) co-financed with BRGM Orléans.
The Region CERES project launched in oct. 2023 also provides a substantial budget for experimentation, with access to paid-for satellite data and online computing.

Sujet :
Work Schedule

The first year of the thesis will begin with a state-of-the-art review of deep learning algorithms applied to remote sensing. This will be followed by an inventory of exploitable data sources, the implementation of a data collection and processing platform, and experimentation with the extraction of a few relevant features from deep learning models derived from the state of the art in semantic segmentation. For the characterization of exposed elements, the aim is to identify the spatial, geometric, spectral and documentary characteristics of interest, which can be exploited in the various data sources and are relevant to the prevention of natural hazards.
We already have a database of images acquired in the Loiret region during the AGEOTHYP program, covering a wide range of terrain (crops, forests, urban areas, rivers, etc.), as well as satellite and documentary data on study sites abroad. The PhD student will have to familiarize himself/herself with the risk theme by consulting BRGM risk experts in order to list the criteria to be analyzed in order to build up a multi-risk issues database. The proximity of the study site will facilitate on-site surveys to enrich the ground truth. The cost of access to certain satellite data sources will be covered by the CERES regional project.
For data processing, the aim is to evaluate the performance of various online or local computing solutions, and to experiment with a few advanced state-of-the-art deep learning models for extracting the characteristics of elements exposed to climatic hazards. The power of the CaSciModot supercomputing infrastructure integrated into the DataCentre Régional Centre Val de Loire on the Grand Campus Orléans, including BRGM and the university, will be used to run the deep learning algorithms. The Region CERES project will also enable the models to be tested using pay-per-use online computing solutions.

During the second year, a ground truth database will be set up to train the algorithms. Network architectures adapted to heterogeneous modalities will be proposed. The possibility of multi-scale processing (building or urban aggregation) will be studied. Experimental validation will be carried out in the Loiret pilot area, where a large number of image modalities with good spatial resolution are available. Other experiments may be carried out on foreign sites, where the absence or scarcity of certain data will lead to a degraded mode.

The third year will be devoted to applying the results of semantic segmentation to one or more risk scenarios (floods, earthquakes, landslides, etc.), and to finalizing the analysis and evaluation of the contribution of deep learning methods to the mapping of issues. The CERES project’s economic partners will enhance the applicative vocation of the work, with a focus on the Blois conurbation and taking into account the concerns of the insurance sector.

The work will be promoted through participation in national and international conferences on general or specialized image processing (artificial intelligence approaches, remote sensing, etc.) and the publication of a journal article.

PRISME – BRGM Collaboration

Computer Vision – AI at PRISME

He has acquired expertise in Deep Learning image processing and hyperspectral imaging:

Deep learning approaches have been used for the semantic segmentation of images, with spectacular results compared with conventional methods, first in the medical field and then in public image databases (e.g. CityScapes: urban scenes or common objects: SBD). Parallel work in this area has also been applied to precision agriculture, animal video surveillance (sheepfolds, zoos, etc.) and heritage (monuments, paintings, etc.).
The hyperspectral imaging modality, which is highly discriminating but generates huge volumes of data, was first tested in the medical field (visualization of tissue spectra in the operating room), then in remote sensing (image segmentation using active contours on a graph), in particular with aerial images acquired by BRGM Orléans during the AGEOTHYP program (detection of clay soils at risk).

Risks and Prevention Division – BRGM
The team makes available heterogeneous data acquired in the field or collected in its databases, and has the expertise to exploit them in multi-risk scenarios:
BRGM has a hyperspectral dataset. The images are centered on a study area west of Orléans covering some 300 km², i.e. 170 images of 408 spectral bands (400 – 2500 nm) with a spatial resolution of 1 to 2 m. A geospatial database is also available to serve as ground truth (laboratory and in situ spectral libraries, spatialized geotechnical data and mineralogical analyses). Another foreign study site will be selected to work in degraded mode. Other Open Source data will be collected on the Loiret study site, depending on availability and quality: thermal infrared emissivity, LIDAR topography, SAR subsurface, IGN database.
BRGM’s risk specialists have the expertise to assess the vulnerability to natural hazards. As part of the RISQNAT research program “Building impact scenarios for the prevention of natural hazards”, they are looking into cost-effective solutions for the production of spatialized information over vast study areas, and the development of platforms integrating predictive models from a multi-risk, multi-scale perspective.

References

Le Cozannet, G., Kervyn, M., Russo, S., Ifejika Speranza, C., Ferrier, P., Foumelis, M., Lopez, T., Modaressi, H., 2020. Space-Based Earth Observations for Disaster Risk Management. Surv. Geophys. 41, 1209–1235. https://doi.org/10.1007/s10712-020-09586-5

Z. Ma, G. Mei, Deep learning for geological hazards analysis: Data, models, applications, and opportunities, Earth-Science Reviews, Volume 223, 2021,103858,ISSN 0012-8252

J. Jakubik, M. Muszynski, M. Vössing, N. Kühl and T. Brunschwiler, Toward Foundation Models for Earth Monitoring: Generalizable Deep Learning Models for Natural Hazard Segmentation, 2023, arXiv 2301.09318

Jia, J.; Ye, W. Deep Learning for Earthquake Disaster Assessment: Objects, Data, Models, Stages, Challenges, and Opportunities. Remote Sens. 2023, 15, 4098. https://doi.org/10.3390/rs15164098

A. Lacoste, N. Lehmann, P. Rodriguez, E. D. Sherwin, H. Kerner , B. Lutjens, J. A. Irvin, D. Dao, H. Alemohammad, A. Drouin, M. Gunturkun, G. Huang, D. Vazquez, D. Newman, Y. Bengio, S. Ermon and X. X. Zhu GEO-Bench: Toward Foundation Models for Earth Monitoring, 37th Conf. on Neural Information Processing Systems Datasets and Benchmarks, 2023

Yuri Boykov, Fatih Porikli, Antonio Plaza, Nasser Kehtarnavaz, and Demetri Terzopoulos Image Segmentation Using Deep Learning: A Survey Shervin Minaee, arXiv:2001.05566v4 [cs.CV] 10 Apr 2020

Prakash, N., Manconi, A., Loew, S., 2020. Mapping Landslides on EO Data: Performance of Deep Learning Models vs. Traditional Machine Learning Models. Remote Sens. 12, 346. https://doi.org/10.3390/rs12030346

Yang, H., Yu, B., Luo, J., Chen, F., 2019. Semantic segmentation of high spatial resolution images with deep neural networks. GIScience Remote Sens. 56, 749–768. https://doi.org/10.1080/15481603.2018.1564499

Jia Song Shaohua Gao, Yunqiang Zhu & Chenyan Ma A survey of remote sensing image classification based on CNNs Big Earth Data, Vol.3, N°3, 232-254, 2019|

X.X. Zhu, D.Tuia, L.Mou, G-S. Xia,L. Zhang, F.Xu, F.Fraundorfer, Deep Learning in Remote Sensing, IEEE Geoscience and Remote Sensing magazine, dec. 2017

L. Ma, Y. Liu, X. Zhang, Y. Ye, G. Yin, B.A. Johnson, Deep learning in remote sensing applications : a meta-analysis and review, ISPRS Journal of Phtogrammetry and Remote Sensing, 1552 (2019) 166-177

E. Colin Koeniguer, G. Le Besnerais, A. Chan Hon,Tong, B. Le Saux, A. Bouich, P. Trouvé, R. Caye Daudt, N. Audebert, G. Brigo, P. Godet, B. Le Teurnier, M. Varvalho, J. Castillo-Navaro, Recent examples of deep learning contributions for earth observation issues , AeroscpaceLab journal, issue 15, sept. 2020

D. Hong, L. Gao, N.Yokoya, J.Yao, J. Chanussot, Q. Du, B. Zhang, More diverse means better : multimodal deep learning meets remote sensing imagery classification, IEEE transactions on geoscience and remote sensing, vol.59, n°5, may 2021

Vali, A., Comai, S., Matteucci, M., 2020. Deep Learning for Land Use and Land Cover Classification Based on Hyperspectral and Multispectral Earth Observation Data: A Review. Remote Sens. 12, 2495. https://doi.org/10.3390/rs12152495

Signoroni, A., Savardi, M., Baronio, A., Benini, S., 2019. Deep Learning Meets Hyperspectral Image Analysis: A Multidisciplinary Review. J. Imaging 5, 52. https://doi.org/10.3390/jimaging5050052

Paoletti, M.E., Haut, J.M., Plaza, J., Plaza, A., 2019. Deep learning classifiers for hyperspectral imaging: A review. ISPRS J. Photogramm. Remote Sens. 158, 279–317. https://doi.org/10.1016/j.isprsjprs.2019.09.006

K. Tabia, X. Desquesnes, , S. Treuillet « A multiphase level set method on graphs for hyperspectral image segmentation” Lecture Notes in Computer Science LNCS 10016, Springer, p, 559-569

K.Tabia, X.Desquesnes, Y.Lucas, S.Treuillet, Influence of spectral metrics on the graph-based segmentation of hyperspectral images, 9th Workshop on Hyperspectral Image and Signal Processing: Evolution in Remote Sensing, WHISPERS 2018, 23-26 sept 2018, Amsterdam, Hollande.

Etienne Ducasse, Karine Adeline, Xavier Briottet, Audrey Hohmann, Anne Bourguignon, et al.. Montmorillonite Estimation in Clay-Quartz-Calcite Samples from Laboratory SWIR Imaging Spectroscopy: A Comparative Study of Spectral Preprocessings and Unmixing Methods. Remote Sensing, MDPI, 2020, ⟨10.3390/rs12111723⟩.

D. Nouri, Y. Lucas, S. Treuillet «Hyperspectral interventional imaging for enhanced tissue visualization and discrimination combining band selection methods» Int. Journal of computer assisted radiology and surgery, Springer Verlag, ISSN 1861-6410, Vol. 11, n°12 p. 2185–2197, déc 2016

D. Nouri, Y. Lucas, S. Treuillet «Efficient tissue discrimination during surgical interventions using hyperspectral imaging » Int. Confrence on Information Processing in Computer –Assisted Interventions (IPCAI) Fukuoka, Japan, 28 june 2014 

R. Niri, H. Douzi,Y. Lucas and S. Treuillet, Fully convolutional networks for diabetic foot ulcers diagnosis, Int. conf. on Medical Diagnostic Imaging and Radiology (ICMDIR 2020), Barcelona, Spain 05-06 march 2020

R. Niri, Y. Lucas, S. Treuillet and H. Douzi, Deep Learning for Multispectral Tissue Analysis applied to Diabetic Foot Ulcer Monitoring, The European Conference on Controversies in Diabetic Foot Management, Vienna, Austria, May 02 – 03, 2019 

R. Niri, E. Guttierez, H. Douzi, Y. Lucas, S. Treuillet, B. Castaneda, I. Hernandez, Multi-View Data Augmentation to Improve Wound Segmentation on 3D Surface Model by Deep Learning, IEEE Access, vol.9, pp. 157628-157638, 2021, doi: 10.1109/ACCESS.2021.3130784.

O. Zenteno, T. V. Pham, S. Treuillet, Y. Lucas, Markerless tracking of micro-endoscope for optical biopsy in stomach, EMBC July 23-27, 2019, Berlin, Germany

T.V. Pham, Y. Lucas, S. Treuillet, L. Debraux, Object contour refinement using instance segmentation in dental images, Int. conf. on Advanced concepts for intelligent vision systems ACIVS 2020, 10-14 Feb 2020, Auckland, New-Zealand,

T.V. Pham, Y. Lucas, S. Treuillet, L. Debraux, Improvement in design and training of feature pyramid network for contour refinement, Pattern Recognition Letters, vol. 155, march 2022, p1-8

M Dian Bah, Eric Dericquebourg, Adel Hafiane, Raphael Canals, Deep Learning based Classification System for Identifying Weeds using High-Resolution UAV Imagery, Chapter in Volume 857 of the Advances in Intelligent Systems and Computing, Jan 2019

M. Kerkech, A. Hafiane, R. Canals, Deep leaning approach with colorimetric spaces and vegetation indices for vine diseases detection in UAV images, Computers and Electronics in Agriculture 155, pp. 237–243, Oct 2018

Hohmann, A., Dufréchou, G., Grandjean, G., Bourguignon, A., 2013. Mapping of swelling and shrinking clays from airborne hyperspectral data: Presentation of a coming comparison of two approaches, in: NIR2013 proceedings. La Grande Motte, France, p. ?

Graff, K., Lissak, C., Thiery, Y., Maquaire, O., Costa, S., Medjkane, M., Laignel, B., 2019. Characterization of elements at risk in the multirisk coastal context and at different spatial scales: Multi-database integration (normandy, France). Appl. Geogr. 111, 102076. https://doi.org/10.1016/j.apgeog.2019.102076

Profil du candidat :
Profile required

Candidates with a research Master’s degree in computer science.

Formation et compétences requises :
Candidates should have extensive knowledge of image processing, including deep learning techniques and their implementation in software and hardware. Fundamental notions of remote sensing are also welcome. Fluency in English is essential. Autonomy, scientific rigor and great motivation for the proposed subject will be undeniable assets for the successful completion of the thesis.

Adresse d’emploi :
Orleans University – Polytech Orleans
12 Rue de Blois, 45100 Orléans

BRGM
3 Av. Claude Guillemin, 45100 Orléans

Candidates must submit the following documents in a single pdf file:
CV + cover letter + Master’s grades – optional letters of recommendation.

Contact us :

yves.lucas@univ-orleans.fr c.gracianne@brgm.fr c.negulescu@brgm.fr

Document attaché : 202404211103_PhD_PRISME_BRGM_IASIS_2024.pdf

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

Laboratoire/Entreprise : IRISA – Rennes
Durée : CD 3 ans
Contact : nathalie.girard@irisa.fr
Date limite de publication : 2024-06-30

Contexte :
L’équipe de recherche Shadoc (anciennement IntuiDoc) (https://www-shadoc.irisa.fr/) de l’IRISA travaille sur l’analyse et la reconnaissance de tracés et de gestes manuscrits réalisés sur surfaces 2D : tablettes et écrans tactiles. Nous nous intéressons notamment à la conception de moteur de reconnaissance de formes et aux nouveaux usages autour de l’interaction gestuelle sur des surfaces tactiles.

L’équipe travaille sur le développement d’environnements numériques innovants sur tablette stylet pour l’éducation, avec le pilotage de plusieurs travaux récents sur l’apprentissage de l’écriture manuscrite pour les classes de primaire et de maternelle, ou sur la production de schémas de géométrie pour les classes de collège.

Sujet :
Ce sujet de thèse s’inscrit dans la thématique des enjeux sociétaux autour de l’IA pour l’éducation. Il fait suite aux travaux de recherche effectués sur la conception de Systèmes Tutoriels Intelligents (STI) pour l’aide à l’apprentissage par le dessin. Nos précédents travaux ont notamment porté sur les thématiques pédagogiques de l’aide à l’apprentissage de la géométrie au collège et sur les schémas d’anatomie dans les formations de Santé. Ces travaux reposent sur des études qui ont démontré que l’introduction d’activités de dessin scientifique dans des cours permettait d’améliorer les performances d’apprentissage des étudiants. Les systèmes tutoriels intelligents permettent de développer des stratégies d’apprentissage personnalisées très efficaces en produisant automatiquement des feedbacks correctifs ou de guidage qui sont adaptés.

Les systèmes tutoriels intelligents [7, 8, 9] sont nés du couplage de deux domaines : l’intelligence artificielle et l’e-éducation. Pour les concevoir, le principe est de modéliser la connaissance experte qui permettra au système d’analyser ensuite automatiquement les actions de l’apprenant. L’analyse porte à la fois sur la reconnaissance des tracés manuscrits semi-structurés, et sur l’analyse de la validité de l’action relativement aux contraintes du problème (protocole de résolution de problème, étapes de dessin).

Dans ce travail de recherche, nous explorerons un nouveau challenge qui consiste à travailler sur un module de génération automatisée de règles expertes (mode auteur) pour appréhender la modélisation structurelle (et compositionnelle) semi-supervisée de schémas. L’ambition est de pouvoir prendre en entrée de l’apprentissage du STI pour générer les règles, aussi bien des schémas structurés (comme pour la géométrie) que des schémas semi-structurés, tel que les schémas d’anatomie.
En facilitant par l’apprentissage semi-supervisé la création des modèles de connaissances adossés aux STI, on permettra d’étendre leurs champs applicatifs à d’autres disciplines pour appréhender par exemple des schémas décrivant des processus ou encore des schémas scientifiques (chimie, biologie, physique…).

Contacts : eric.anquetil@irisa.fr; nathalie.girard@irisa.fr

Profil du candidat :

Formation et compétences requises :

Adresse d’emploi :
laboratoire IRISA, équipe Shadoc, Rennes, France

Document attaché : 202404201028_sujetTèse_EA_Shadoc_2024 V2.pdf

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

Laboratoire/Entreprise : IRISA/ATERMES
Durée : 36 mois
Contact : minh-tan.pham@irisa.fr
Date limite de publication : 2024-05-15

Contexte :
Please find the full PhD topic here: http://www-obelix.irisa.fr/files/2024/04/PhD_Cifre2024_IRISA_ATERMES.pdf

ATERMES is an international mid-sized company, based in Montigny-le-Bretonneux with a strong expertise in high technology and system integration from the upstream design to the long-life maintenance cycle. It specializes in offering system solution for border surveillance. Its flagship product BARIER™ (“Beacon Autonomous Reconnaissance Identification and Evaluation Response”) provides ready application for temporary strategic site protection or ill-defined border regions in mountainous or remote terrain where fixed surveillance modes are impracticable or overly expensive to deploy. As another example, SURICATE is the first of its class optronic ground “RADAR” that covers very efficiently wide field with automatic classification of intruders thanks to multi-spectral deep learning detection.

Sujet :
The project aims at providing deep learning-based methods to detect objects in outdoor environments using multispectral data in a low supervision context, e.g., learning from few examples to detect scarcely-observed objects. The data consist of RGB and IR (Infra-red) images which are frames from calibrated and aligned multispectral videos.

Few-shot learning [1][2], semi-supervised learning [3][4] and continual learning [5][6] are among the most widely-used frameworks to tackle this task. For the first approach based on few-shot object detection (FSOD), the recent trend has relied on using meta learning or transfer learning approaches [1:1]. Yet, realistic settings including scarce objects may exist a domain shift that makes the task more challenging. The second approach based on semi-supervised learning considers a large amount of unlabeled data in the training process to foster the representation capacity of deep models, improving the peformance of object detection from a small amount of labeled samples. As the third approach, continual learning [5:1] aims to maintain the performance of the deep models on old categories and avoid the “catastrophic forgetting” phenomenon when learning new object categories. It has been also integrated into a FSOD task [7] to ensure that few-shot object detectors could learn new object concepts without forgetting previous object categories that still exist in prediction phase. Last but not least, with the dramastically rapid evolution of research in AI, another challenge to tackle is the investigation of modern AI models, and more specifically foundation models which involves multimodal transformers [8][9]. Indeed, these large machine learning models trained on a vast quantity of data at scale have been designed to be adapted to a wide range of downstream tasks (including object detection, see for instance UniDetector [10]) or CLIP2 [11]. These models leading to zero-shot object detection could very well be the ultimate answer for the task of having a true scene understanding.

Profil du candidat :
MSc or Engineering degree with excellent academic track and proven research experience in the following fields: computer science, applied maths, signal processing and computer vision;

European nationality required

Formation et compétences requises :
Experience with machine learning, in particular deep learning;

Skills and proved experience in programming (Python is mandatory and knowledge about frameworks such as Pytorch is a real plus);

Excellent communication skills (spoken/written English) is required ;

Ambition to publish at the best level in the computer vision community (CVPR, ICCV, TPAMI, …) during the thesis.

Adresse d’emploi :
IRISA, Université Bretagne Sud, 56000 Vannes

Document attaché : 202404161424_PhD_Cifre2024_IRISA_ATERMES.pdf

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

Laboratoire/Entreprise : Laboratoire GREYC
Durée : 36 mois
Contact : jean-luc.lamotte@unicaen.fr
Date limite de publication : 2024-05-15

Contexte :

Sujet :
Les laboratoires de recherche GREYC(informatique) et CERMN (pharmacologie) travaillent depuis de nombreuses années au sein d’un groupe commun pour développer des méthodes informatiques innovantes afin de traiter des données liées aux médicaments et notamment d’essayer de prédire et de comprendre le mode d’action d’une nouvelle molécule à partir de connaissances extraites sur un ensemble de molécules.

Les molécules pouvant être modélisées par des graphes, il est possible d’apprendre à un GNN (Graph neural network) à classifier ou prédire l’action de molécules, mais actuellement, aucune explication sur la prise de décision du réseau n’est donnée. L’explicabilité des réseaux est fondamentale dans la prédiction des propriétés thérapeutiques de molécules. En effet, celle-ci permet~:

– de valider les prédictions avant d’engager des efforts et des moyens sur les synthèses d’une série moléculaire

– d’obtenir une intuition sur les propriétés physico-chimiques clés que doit posséder une molécule pour avoir une action biologique ciblée.

L’objectif de cette thèse est de proposer des solutions pour expliquer les décisions que prend un réseau de neurones opérant sur des graphes en vue de l’appliquer sur des données des molécules chimiques. Aucune connaissance en chimie thérapeutique n’est requise.

Profil du candidat :
La personne candidate doit être inscrit en dernière année d’un Master ou d’un diplôme d’ingénieur, ou être titulaire d’un tel diplôme, dans un domaine lié à l’informatique ou aux mathématiques appliquées, et posséder de solides compétences en programmation. Une expérience en informatique pour la Science des Données, l’apprentissage profond, … sera un plus. La personne doit avoir des capacités à rédiger des rapports scientifiques et à communiquer des résultats de recherche lors de conférences en anglais.

Formation et compétences requises :

Adresse d’emploi :
Equipe CODAG, laboratoire GREYC, Université de Caen Normandie

Document attaché : 202404161409_theseExplicabilite.pdf