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

Laboratoire/Entreprise : INRIA – EVERGREEN
Durée : 3 ans
Contact : diego.marcos@inria.fr
Date limite de publication : 2024-06-30

Contexte :
This PhD offer is funded by the GEO-ReSeT ANR project, representing a collaboration between Inria (team EVERGREEN, Montpellier) and Université de Paris Cité (team LIPADE, Paris).

Leveraging the large amounts of available geo-spatial data from different sources, the GEO-ReSeT (Generalized Earth Observation with Remote Sensing and Text) project has the objective to learn a rich representation of any geo-spatial location and convey a semantic representation of the information, by improving on existing models and providing a better experience to the end users. By using location on the Earth’s surface as the common link between different modalities, a geo-spatial foundation model would be able to incorporate a variety of data sources, including remote sensing imagery, textual descriptions of places, and other generic features.

Such a foundation model has the potential to open a set of all new possibilities in terms of Earth observation applications, by allowing for few or zero-shot solutions to classical problems such as land-cover and land-use mapping, target detection, and visual question answering. It will also be useful for a wide range of applications with a geo-spatial component, including environmental monitoring, urban planning and agriculture.
By leveraging several data modalities, this foundation model could provide a comprehensive and accurate understanding of the Earth’s surface, enabling informed decisions and actions. This will be particularly valuable for new potential users in sectors such as journalism, social sciences or environmental monitoring, who may not have the resources or expertise to collect their own training datasets and develop their own methods, thus moving beyond open Earth observation data and democratizing the access to Earth observation information.

Sujet :
The work to be conducted during the proposed PhD thesis will contribute to the ambition of the GEO-ReSeT ANR project by linking textual descriptions of places (e.g., collected from heterogeneous online sources, such as news articles or search engine results), to their approximate geo-location, a task known as geoparsing.

This text-location link will then be used in combination with other geospatial data modalities, with a focus on remote sensing data from sensors such as Sentinel-1 and -2, in order to train multi-modal models that are aware about the way in which people describe locations.

This will be done by first combining information stemming from different databases containing geographic named entities, such as Open Street Map, Wikipedia and gazetteers, such that geographic points or polygons can be linked to each named entity.

In a second step, a Natural Language Processing (NLP) pipeline will be developed to obtain the most likely geographic named entities that are referred to in any piece of text that describes a place.

With respect to existing Named Entity Recognition (NER) methodologies, in order to avoid restricting us to cases where entities’ names appear exactly as in the databases or gazetteers, we will leverage pre-trained Large Language Models (LLM) to resolve ambiguities and gather evidence towards the most likely entities that are being described in the text. Such an approach will be trained and validated by using the cases that do match the names in the gazetteer.

We will then move on, in collaboration with the rest of the GEO-ReSeT consortium, to train a multi-modal large language model (MMLLM) that will serve as a foundation model for Earth observation tasks.

This model will finally be evaluated on several agro-environmental tasks.

Application must be sent through the following link : https://recrutement.inria.fr/public/classic/en/offres/2024-07756

Profil du candidat :
Main activities
Description of the state-of-the-art in unstructured text geoparsing, with a focus on approaches leveraging LLMs.
Collection of a database of geographic named entities linked to their geographic footprint (e.g. point or polygon). Collection of a database of unstructured online text that is likely to contain a reference to a geographic location.
Development of an NLP pipeline to link each piece of geographic text to its likely geographic footprint.
Participate in the design and training of a multi-modal large language model (MMLLM) using remote sensing and geoparsed text.
Evaluation of the final model on two of the following case studies at a national or continental scale: ecosystem type mapping, crop type mapping or land-use mapping.

Formation et compétences requises :
Skills
Python programming.
Deep Learning with Python (preferably with Pytorch).
Experience with NLP.
Experience with GIS would be a plus.

Adresse d’emploi :
Montpellier, France

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

Laboratoire/Entreprise : LITIS Lab, Rouen
Durée : 3 years
Contact : paul.honeine@univ-rouen.fr
Date limite de publication : 2024-06-30

Contexte :
Keywords: Deep learning, generative (probabilistic) models, diffusion probabilistic models, normalizing flows, anomaly detection, time series

Sujet :
The broad interest in deep neural networks has driven recent advances in anomaly detection, also called out-of-distribution or novelty detection. Deep anomaly detection methods fall within three major categories: Deep one-class, variational autoencoders and generative adversarial networks [1, 2]. While these methods do not allow an exact evaluation of the probability density of new samples, they also suffer from notorious training instability (mode collapse, posterior collapse, vanishing gradients and non-convergence), as corroborated by many research studies [3]. For these reasons, we aim to go beyond the weaknesses of these methods, by investigating novel classes of generative models with deep learning to address anomaly detection.

The goal of this PhD thesis is to explore novel generative models, such as diffusion probabilistic models (DPM) and normalizing flows (NF). These classes consist of models that can generate, through a deep latent space, a probability distribution for a given dataset from which we can then sample. With solid theoretical foundations and often interconnections with Optimal Transport, several variants of generative models have been proposed based on different definitions of their main components, namely the forward and backward processes and the sampling procedure. Of particular interest are NF and DPM. NF are generative models where both sampling and density evaluation are efficient and exact, and where the latent representation is learned through an invertible transformation, thus providing explainable models [4, 5]. DPM rely on diffusion processes, inspired from nonequilibrium thermodynamics, with their flagship being denoising diffusion probabilistic models [6]. Diffusion models have been demonstrating record-breaking performance in many applications in computer vision, mainly for image synthesis [7, 8] and medical imaging [9].

The proposed PhD research program aims to investigate these recent advances in generative models with deep learning for anomaly detection. Recent studies have explored generative probabilistic models for anomaly detection, mainly in images [10-12] with some attempts in signal processing [13, 14], demonstrating preliminary results on their relevance in anomaly detection and bringing out new research questions [15]. The PhD candidate will investigate such generative probabilistic models in a more in-depth research study, in order to take full advantage of their underlying theory. Moreover, the PhD candidate will go beyond image processing, with a focus on anomaly detection in time series, by considering the specificities of time series. The proposed framework and devised methods will be assessed in a variety of scenarios and real-world time series datasets.

Research Environment

The PhD candidate will conduct her/his research within the Machine Learning group in the LITIS Lab, under the supervision of Prof. Paul Honeine, Dr. Fannia Pacheco and Dr. Maxime Berar. This PhD thesis is within a research project gathering 9 permanent researchers of the LITIS Lab and the PhD candidate will also interact with several PhD students and interns also working on deep anomaly detection with a focus on time series.

Application

Applicants are invited to send their CV and grade transcripts by email to:
paul.honeine@univ-rouen.fr, fannia.pacheco@univ-rouen.fr, maxime.berar@univ-rouen.fr.

References

[1] L. Ruff et al., “A unifying review of deep and shallow anomaly detection,” Proceedings of the IEEE, 2021.
[2] G. Pang et al., “Deep learning for anomaly detection: A review,” ACM Computing Surveys, 2021.
[3] D. Saxena and J. Cao, “Generative adversarial networks (GANs) challenges, solutions, and future directions,” ACM Computing Surveys, 2021.
[4] I. Kobyzev et al., “Normalizing flows: An introduction and review of Current Methods,” IEEE T PAMI, 2021.
[5] G. Papamakarios et al., “Normalizing flows for probabilistic modeling and inference,” JMLR, 2021.
[6] J. Ho et al., “Denoising diffusion probabilistic models,” NeurIPS, 2020.
[7] L. Yang et al., “Diffusion models: A comprehensive survey of methods and applications,” ACM Computing Surveys, 2023.
[8] F.-A. Croitoru et al., “Diffusion models in vision: A survey,” IEEE T PAMI, 2023.
[9] A. Kazerouni et al., “Diffusion models in medical imaging: A comprehensive survey,” Medical Image Analysis, 2023.
[10] J. Wolleb et al., “Diffusion models for medical anomaly detection.” MICCAI, 2022.
[11] W.H. Pinaya et al., “Fast unsupervised brain anomaly detection and segmentation with diffusion models,” MICCAI, 2022.
[12] A. Kascenas et al., “The role of noise in denoising models for anomaly detection in medical images.” Medical Image Analysis, 2023.
[13] Y. Chen et al., “ImDiffusion: Imputed Diffusion Models for Multivariate Time Series Anomaly Detection,” in Proc. of the VLDB Endowment, 2023.
[14] R. Hu et al., “Unsupervised Anomaly Detection for Multivariate Time Series Using Diffusion Model,” IEEE ICASSP, 2024.
[15] P. Kirichenko et al., “Why normalizing flows fail to detect out-of-distribution data,” NeurIPS, 2020.

Profil du candidat :
– Master or Engineering degree, in data science, AI, applied mathematics, or related fields.
– Strong skills in advanced statistics and Machine Learning, including Deep Learning
– Good programming experience in Python

Formation et compétences requises :

Adresse d’emploi :
LITIS Lab, Rouen

Document attaché : 202405242231_PhD – Deep learning with Generative Models for Anomaly Detection.pdf

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

Laboratoire/Entreprise : CReSTIC & MEDyC
Durée : 36 mois
Contact : nicolas.passat@univ-reims.fr
Date limite de publication : 2024-12-31

Contexte :
Projet ANR MODELAGE
Cette thèse est proposée dans le cadre du projet ANR MODELAGE (Modélisation de l’évolution des microstructures
vasculaires par imagerie synchrotron à très haute résolution – Prédiction du vieillissement
normal vs accéléré), mené en partenariat entre l’Université de Reims Champagne-Ardenne, l’Université de
Technologie de Troyes et le Synchrotron SOLEIL (Saclay).

Contexte
Le vieillissement vasculaire se caractérise par des altérations lentes et asymptomatiques des microstructures
vasculaires. Parmi celles-ci, les lamelles élastiques de la paroi vasculaire sont les premières concernées.
Néanmoins, les événements précoces prévoyant ces altérations restent pour la plupart non documentés.
En effet, les méthodes d’exploration actuelles n’atteignent pas une résolution suffisante. L’exploration des
caractéristiques vasculaires à l’aide de la microtomographie à rayons X synchrotron haute résolution (μCT) a
révélé l’existence d’un réseau en forme de treillis construit à l’intérieur des lamelles élastiques chez la souris.
Les images μCT acquises sur synchrotron peuvent ainsi fournir de nouveaux indices pour comprendre le
processus de vieillissement vasculaire [1]. En effet, leur résolution, leur contraste et leur champ de vision sont
si élevés qu’ils révèlent de nouveaux détails structurels fins dans la paroi aortique. Cependant,
la recherche, l’extraction et l’analyse de ces données massives et riches en informations constituent un
véritable défi.

Sujet :
Objectifs
Les images μCT sont des données 3D de très haute résolution (voxels de 0.65 μm de côté) de très grande
taille (4000 x 4000 x 2000 voxels) pouvant de plus être empilées jusqu’à former des structure de l’ordre du
tera-octet. Il est, en l’état, impossible de naviguer dans ces données et de les analyser dans leur globalité.
Les solutions actuellement développées les manipulent par coupes 2D et/ou par tranches 3D épaisses [2].
Le premier but de cette thèse est de développer de nouvelles structures de données hiérarchiques (arbres)
qui permettent de modéliser les images à différents niveaux d’échelle en adaptant le niveau d’échelle local
au niveau de détail dans les images. Une telle politique repose de manière conjointe sur deux paradigmes :
les espaces d’échelles [3] et les modèles de décomposition de type quadtree/octree [5]. Contrairement aux
stratégies usuellement considérées pour les espaces d’échelles (approximation gaussienne) et pour les octrees
(subdivisions régulières), l’idée est ici de tirer parti de la connaissance a priori sur le contenu des images pour
développer un modèle hiérarchique morphologique [4] qui puisse représenter les images avec un minimum
de perte d’information, tout en maximisant la compacité des structures, afin de permettre leur gestion en
mémoire et une navigation complète sans recours à des architectures matérielles lourdes.

Les objectifs de cette thèse seront ainsi d’explorer des stratégies pour :
• définir de tels modèles hiérarchiques ;
• les construire de manière efficace ;
• développer de nouveaux descripteurs d’images dédiés aux images μCT ;
• développer des politiques de calcul efficaces de ces descripteurs sur les modèles hiérarchiques développés.

Ces travaux viendront s’interfacer avec des méthodes et outils récemment développés pour l’analyse des
images synchrotron, dans le cadre du projet ANR MODELAGE. Le(la) candidat(e) aura aussi l’opportunité
de participer aux campagnes d’acquisition des images lors des expériences synchrotron.

Profil du candidat :
Compétences requises
Le(la) candidat(e) sera titulaire d’un diplôme de Master 2 et/ou d’un diplôme d’ingénieur. Il(elle) aura
des compétences solides en informatique, mathématiques, et une capacité à travailler dans un contexte
collaboratif et pluridisciplinaire.
Compétences impératives :
• Programmation C++ et Python
• Traitement et analyse d’images
Compétences souhaitées mais non-indispensables :
• Imagerie

Formation et compétences requises :
Compétences requises
Le(la) candidat(e) sera titulaire d’un diplôme de Master 2 et/ou d’un diplôme d’ingénieur. Il(elle) aura
des compétences solides en informatique, mathématiques, et une capacité à travailler dans un contexte
collaboratif et pluridisciplinaire.
Compétences impératives :
• Programmation C++ et Python
• Traitement et analyse d’images
Compétences souhaitées mais non-indispensables :
• Imagerie

Adresse d’emploi :
Lieu d’exercice
Université de Reims Champagne-Ardenne, Campus Moulin de la Housse
Laboratoires CReSTIC et MEDyC

Document attaché : 202405211157_MODELAGE_PhD.pdf