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

Laboratoire/Entreprise : Mathématiques et Informatique Appliquées – UMR 518
Durée : 3 ans
Contact : antoine.cornuejols@agroparistech.fr
Date limite de publication : 2021-09-15

Contexte :
Analysing heterogeneous remote sensing time series using supervised methods requires that the classes sought are perfectly known and defined and that the expert is able to provide a sufficient learning data set both in number and quality. Faced with the difficulty of obtaining sufficient examples within the context of the analysis of time series of remote sensing images, we propose to develop an innovative method of interactive multi-paradigm collaborative learning. The aim is to enable the expert to add “on the fly” information (labels, constraints, etc.) used to guide the learning process in order to produce clusters and models closer to the expert’s “intuitions”, i.e. potential thematic classes. To do this, the expert will be actively assisted by the system, which will for example offer advice or proposals for new constraints or labelling of objects. We will validate our work in several fields of application chosen in agreement with partners of the HERELLES project.

Sujet :
With the launch and entry into production of the European satellites in the Sentinel or Franco-Israeli constellation Venųs, satellite data are now arriving in massive, almost continuous flows. This massive influx of temporal data should lead to major advances in various Earth and environmental science disciplines for the study and modelling of complex phenomena (agricultural or urban dynamics, deforestation, anthropogenic actions on biodiversity, etc.). However, faced with this overabundance of temporal data, arriving almost continuously, the labelling phase of supervised learning can no longer be carried out by experts, as it is too tedious and time-consuming. Moreover, the supervised learning methods classically used in Earth observation assume that the learning data sufficiently and completely describe the classes to which they are attached. In other words, these methods require that the desired classes are well known and defined and that the expert is able to provide a sufficient set of learning data both in number and quality. In the case of temporal analysis in remote sensing, this assumption is no longer realistic. Indeed, the technological revolution of high-frequency image acquisition is still too recent for thematic knowledge to have adapted. Thus, there are currently no typologies (or nomenclatures) of changes that can really be used for this type of supervised analysis and therefore no associated quality learning data.
To compensate for this lack of formalization and examples, the expert must be able to rely on other types of information such as partially labelled data, formalized knowledge, constraints on data or results. At the same time, there are also numerous methods capable of analyzing this data. Combining these data and methods seems indispensable. Thus, approaches such as boosting, clustering or collaborative clustering take advantage of the complementarity between different methods, each with its own biases and its own analysis strategy but capable of processing its own data in a privileged way.
However, with the increase in the volume of data and the number of potential evolution classes, the highlighting and formalization of information that is really relevant for classification methods in the context of temporal analysis appears to be more difficult than expected and potentially time-consuming. The objective of this project, in strong link with the ANR HIATUS and HERELLES projects, is to define and validate in the context of high acquisition frequency remote sensing, an innovative method of interactive collaborative learning. The aim is to enable the expert to add “on the fly” information (labels, classes, constraints, etc.) that can be used to guide the learning process in order to produce clusters and models that are closer to the expert’s “intuition”, i.e. potential thematic classes . To do this, the expert will be actively assisted by the system, which will offer advice or proposals for new constraints or object labelling, for example.

Profil du candidat :
Master’s Degree in Computer Science or equivalent.

Formation et compétences requises :
The candidate must have good skills in data analysis and more particularly in supervised or unsupervised classification of time series. Skills in remote sensing image analysis is required. Good knowledge of English (French is not mandatory)

Adresse d’emploi :
Paris Campus Saclay

Document attaché : 202107060833_Sujet_HERELLES_2021.pdf

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

Laboratoire/Entreprise : LITIS laboratory (INSA ROUEN) and IRISA Vannes (Un
Durée : 3 ans
Contact : laetitia.chapel@irisa.fr
Date limite de publication : 2021-07-25

Contexte :
To be safe, a decision device learned from data requires a mechanism that adapts the decision according to whether or not there is a discrepancy between the distribution ptrain(Xtrain; Ytrain) of the training samples and the ones of test samples ptest(Xtest; Ytest). In case of distribution shift, deep-basedapproaches may be overconfident and tend to treat the given inputs as one of the previously seen situations leading to mislabelling. This brings to the scientific challenges of detecting out-of-distribution (OOD)
samples (the test point x0 is marginally sampled from ptest(x0)= ptrain(x0)) or of recognizing that point x0 belongs to an unseen class (new type of object occurs in the scenes). Moreover due to the multimodal nature of the inputs and sensors availability, the samples may not be embedded into the same space, and hence compromising the success of the detection task. We envision to leverage on the optimal transport theory to implement algorithms dealing with out-of-distribution detection, with specific applications on road scene.
Optimal transport (OT) has emerged as a powerful tool to compute distances (a.k.a. Wasserstein or earth mover’s distances) between empirical distribution of data, thanks to new computational schemes
that make the transport computation tractable. It has wide applications in computer vision, statistics, imaging and has been recently introduced in the machine learning community to efficiently solve
classification or transfer learning problems. The advantage of OT is that it can compare possibly high dimensional empirical probability measures, taking into account the geometry of the underlying metric spaces and dealing with discrete measures. Classical optimal transport problem seeks a transportation map that preserves the total mass between two probability distributions, requiring their mass to be the same. This may be too restrictive in certain applications such as color or shape matching, since the distributions may have arbitrary masses and/or
that only a fraction of the total mass has to be transported. This happens also when datasets Xtrain and/or Xtest are contaminated by outliers, in which we may want to discard them from the tranportation
plan: this is the unbalanced[5] or the partial OT problem. Several algorithms have been devised to solve the problem, among them solve the exact partial problem when given as input the total mass that has to be transported between the two empirical distributions. More recently, the team has been developped to solve the unbalanced problem, providing the first regularization path for unbalanced OT.

Sujet :
The objective of the thesis is to study and implement OT-based strategies for dealing with OOD samples or when the datasets are contaminated by outliers.
In many cases, the number of such samples are unknown and should be estimated from the data. To do so, one can rely on two-sample tests and their Wasserstein counterparts [9]; when there is a shift between ptrain and ptest, or even when the 2 distributions do not lie on the same space, one can rather build on the Gromov-Wasserstein based tests.

In more details, the aim is to study how the partial/unbalanced formulation of OT can be used in the OOD and outliers scenarii. Integration of two-sample tests within the OT formulation as a regularization term will be considered first. As such, we aim at estimating from the data the proportion of contaminated samples in the datasets, together with the optimal transport plan in a unified formulation, even when the 2 distributions live in incomparable spaces. One can also rely on the regularization path to select the “best” regularization parameter in a given context. Integration of partial-OT-based loss in deep-based approaches will serve as a playground to evaluate the proposed methods. The scalability
should be an important feature of the methods to be developed.
From an application point of view, a particular attention will be given on OOD detection for road scene. The intended methods will be evaluated on real-world datasets comprising of automotive images (such as nuScenes, KITTI) or on autonomous car scene benchmark https://github.com/OATML/
oatomobile in order to build robust system for road scene analysis. The developed methods will be challenged with some current position approaches and their applications.

Profil du candidat :
Applicants are expected to be graduated in computer science and/or machine learning and/or signal & image processing and/or applied mathematics/statistics, and show an excellent academic
profile. Beyond, good programming skills are expected.

Formation et compétences requises :
computer science and/or machine learning and/or signal & image processing and/or applied mathematics/statistics

Adresse d’emploi :
Vannes or Rouen

Document attaché : 202106301333_Unbalanced optimal transport for OOD.pdf

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

Laboratoire/Entreprise : PRISME (INSA CVL) ; LIFAT (Université de Tours)
Durée : 36 mois
Contact : adel.hafiane@insa-cvl.fr
Date limite de publication : 2021-07-15

Contexte :
Les adventices sont en concurrence directe avec les plantes cultivées dans la recherche d’humidité, des nutriments et de la lumière du soleil. Elles ont ainsi un important impact négatif sur le rendement agricole si leur présence n’est pas suffisamment contrôlée. D’après l’organisation de recherche environnementale de la Nouvelle Zélande, Land Care, les adventices ont été à l’origine d’environ 95 milliards de dollars de pertes sur la production vivrière à l’échelle mondiale. L’Organisation des Nations Unies pour l’Alimentation et l’Agriculture (FAO) considère que les adventices devraient être reconnues comme l’ennemi naturel numéro 1 pour la production agricole (FAO 2009). La lutte contre les adventices a toujours été considérée comme un défi majeur pour la production agricole. Les approches et tendances actuelles en perception (technologies des capteurs d’images et de localisation, méthodologies d’acquisition et de traitement des images), en robotique et en général en intelligence artificielle (IA), ouvrent la voie à de nouvelles avancées prometteuses. Il existe aujourd’hui une tendance vers l’agriculture numérique et robotisée pour résoudre les différents problèmes et améliorer les conditions du travail dans les domaines agricoles.

Sujet :
La détection et la reconnaissance des plantes à partir de capteurs optiques sont parmi les défis majeurs pour le désherbage automatique. Malgré les avancées considérables enregistrées ces dernières années concernant les méthodes de reconnaissance visuelle automatique, notamment grâce au deep learning, les capacités de discriminations entre végétations restent limitées. C’est le cas lorsque l’on observe des variations dans les conditions d’éclairage, de l’occultation entre plantes, des changements de terrain, ainsi qu’en fonction des stades de développement des cultures… L’un des points clés pour la réussite d’un algorithme de type deep learning est l’abondance de données étiquetées, qui est une limitation dans le domaine de l’agriculture numérique. Le processus d’étiquetage nécessite généralement l’intervention d’experts, ce qui constitue l’une des principales limites de construction de bons modèles robustes et généralisables avec les réseaux neuronaux profonds. Le travail de la thèse consistera à pallier ce type de problèmes en développant de nouvelles méthodes pour cartographier des adventices à partir d’images de drone à haute résolution, avec des algorithmes d’analyse d’images et d’apprentissage automatique (semi-supervised learning, weak learning, generative learning, attention mechanism, transformers,…). Cette cartographie sera améliorée en prenant en compte des données de natures différentes comme la biologie végétale, la météo, … En particulier, la prise en compte d’un modèle prédictif alimenté par des données hétérogènes multisources (météo régionale, capteurs locaux, historique de la cartographie…) devrait permettre de prédire la probabilité de présence et de croissance des plantes localement et améliorer ainsi la détection. Que ce soit sur la reconnaissance de végétaux ou la prédiction, de nouvelles contributions en machine learning seront étudiées.

Modalité de candidature
Transmettre par courriel aux contacts ci-dessous : un CV, lettre de motivation et relevés de notes avant le 08 juillet 2021.

Profil du candidat :
– Master 2 et/ou école d’ingénieur

Formation et compétences requises :
– Compétences en vision par ordinateur et en machine learning
– Des connaissances en robotique seront appréciées mais pas indispensable.
– Compétences en développement Python, C/C++, …
– Bon niveau en anglais

Adresse d’emploi :
L’encadrement et la direction de la thèse seront assurés par des chercheurs des laboratoires PRISME (INSA CVL-Université d’Orléans) et LIFAT (Université de Tours). La thèse est financée par un projet de recherche régionale. Elle se déroulera à l’INSA CVL sur le campus de Bourges.

Document attaché : 202106281536_sujetThèseDesherbrob.pdf

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

Laboratoire/Entreprise : Euromov D.H.M
Durée : 3 years
Contact : patrice.guyot@mines-ales.fr
Date limite de publication : 2021-10-01

Contexte :
A 3-year fully funded PhD scholarship is proposed by the PhD school (ED I2S) in Alès / Montpellier under the supervision of Patrice Guyot (PhD, sound analysis), Pierre Slangen (Pr, motion capture) and Benoît Bardy (Pr, embodied cognition).
The successful applicant will become part of a dynamic research environment within the newly multidisciplinary joint research center EUROMOV Digital Health in Motion (landing theme: Perception in Action & Synchronization – PIAS).

See this offer on the Euromov D.H.M website:
https://dhm.euromov.eu/wp-content/uploads/2021/06/Ph.D_MovementMusicSync.pdf

As a PhD student, you will be responsible for:
– Independently carrying out research and completing a PhD dissertation within three years,
– Recruiting participants and organize experiments in our labs,
– Collecting and synthesizing motion data and music,
– Developing algorithms and methods to analyze motion and music data,
– Reporting the results in international peer-reviewed scientific journals and conferences.

Start date: October 1st, 2021 (to September 2024).
Net remuneration around 1400€ monthly (including social security and health benefits).

Sujet :
Synchronized group activities, such as dancing, singing or certain sports, strengthen human attachment and improve individual well-being [Lau16]. In a physical activity such as tai chi chuan, synchronization within the group is based on the a priori knowledge of individuals and their perception of movements of other participants. In the context of a fitness or capoeira class, synchronization is also based on the common perception of the rhythms of the music.

In general, synchronies between individuals are based on predictive abilities that are fed by visual and auditory perception [Bar20, Tra18]. However, the way in which sound and visual information interacts in the perception and production of synchronization, intentional or spontaneous, is still poorly understood [Ips17].

The ability to synchronize movements and music is primarily analyzed through very simple tasks such as tapping. In more complex situations, such as walking, synchronization can be analyzed through the impact of the steps, and their correspondence with the strong beats of the piece [Dec18]. However, human movements, similarly to music, are composed of cycles at multiple levels, presenting complex rhythmic relationships (expiration and foot impact for running) or hierarchical structures (binary or ternary alternation of strong and weak beats for music).

Beyond classical approaches to detect simple cycles such as the downbeats of pieces (e.g., with neural networks [Jia19]), recent work on automatic analysis converges toward multi-scale modeling of musical content. In this context, conditional random fields have been proposed for leveraging multi-scale information for computational rhythm analysis [Fue19].

In the context of sports practice, coaches are required to perceive these complex multi-scale synchronization patterns. Research has shown that humans synchronize better through auditory or multimodal stimuli than through visual-only stimuli [El10]. These results can be exploited within the framework of synchronization perception to produce visualization and sonification tools. These tools could facilitate the task of the coach when practicing online sport, by enhancing group synchronization for instance, and by allowing rapid identification of people in difficulty in order to offer individualized coaching. Applied to motion capture data, they could also be used in a medical setting to illustrate stability loss in movement polyrhythms.

In this thesis, we propose to analyze multi-scale synchronization patterns between individual movements, group movements, and musical rhythms. We will produce data from motion capture of individuals and groups in the laboratory as well as in more natural settings, and sound synthesis of multi-scale rhythmic content. This data will be analyzed using different approaches from Artificial Intelligence, including neural networks and probabilistic graphical models. Experiments will also be carried out on the perception of synchronization via the representation and / or sonication of the results of these analyzes, with the aim of developing computer bricks facilitating human evaluation of synchronization.

A better understanding of synchronization mechanisms, and their inclusion in IT, may improve collaborative virtual as well as rehabilitation of patients with social disorders [Slo17] or Parkinson’s disease [Dec18].

References

– [Lau16] Launay, Jacques, Bronwyn Tarr, and Robin IM Dunbar. “Synchrony as an adaptive mechanism for large‐scale human social bonding.” Ethology 122.10 (2016): 779-789.
– [Bar20] Bardy, Benoît G., et al. “Moving in unison after perceptual interruption.” Scientific reports 10.1 (2020): 1-13.
– [Tra18] Tranchant, Pauline. “Synchronisation rythmique déficiente chez l’humain: bases comportementales.” Diss. Université de Montréal (2018).
– [Ips17] Ipser, Alberta, et al. “Sight and sound persistently out of synch: stable individual differences in audiovisual synchronisation revealed by implicit measures of lip-voice integration.” Scientific Reports 7.1 (2017): 1-12.
– [Dec18] De Cock, V. Cochen, et al. “Rhythmic abilities and musical training in Parkinson’s disease: do they help?.” NPJ Parkinson’s disease 4.1 (2018): 1-8.
– [Jia19] Jia, Bijue, Jiancheng Lv, and Dayiheng Liu. “Deep learning-based automatic downbeat tracking: a brief review.” Multimedia Systems 25.6 (2019): 617-638.
– [Fue19] Fuentes, Magdalena. “Multi-scale computational rhythm analysis: a framework for sections, downbeats, beats, and microtiming”. Diss. Université Paris-Saclay, 2019.
– [Chu16] Chung, Junyoung, Sungjin Ahn, and Yoshua Bengio. “Hierarchical multiscale recurrent neural networks.” arXiv preprint arXiv:1609.01704 (2016).
– [Tav19] Tavanaei, Amirhossein, et al. “Deep learning in spiking neural networks.” Neural Networks 111 (2019): 47-63.
– [El10] Elliott, Mark T., Alan M. Wing, and Andrew E. Welchman. “Multisensory cues improve sensorimotor synchronisation.” European Journal of Neuroscience 31.10 (2010): 1828-1835.
– [Slo17] Słowiński, Piotr, et al. “Unravelling socio-motor biomarkers in schizophrenia.” npj Schizophrenia 3.1 (2017): 1-10.

Profil du candidat :
Applicants should have (or anticipate having) a MSc and research background related to computer science, audio/signal processing, or computational movement science. Knowledge in music (theoretical and practical) will be valued. French is not mandatory, but the candidate must be willing to learn French during their PhD and they must be able to communicate in English.

Applications should include a cover letter discussing your interest in the position, detailed CV, academic results (evaluation, average and ranking of the candidate during the initial course and Msc) and two reference letters. Deadline is July 5, 2021. Interviews will be conducted via zoom on Tuesday, July 13 and Thursday, July 15.

Formation et compétences requises :
Applicants should have (or anticipate having) a MSc and research background related to computer science, audio/signal processing, or computational movement science.

Adresse d’emploi :
Euromov D.H.M
IMT Mines Ales / Univ. of Montpellier

Document attaché : 202106221434_Ph.D_MovementMusicSync.pdf