Ouverture d’une Chaire de Professeur Junior en Sciences des Données appliquées à la bio-informatique: CAIMAN [Cancer : Artificial Intelligence for Multiomics Analyses]

Le centre Inria de Lyon, en partenariat avec le Centre Léon Bérard et l’Université Claude Bernard Lyon 1, propose une Chaire de Professeur Junior en Sciences des Données appliquées a la bio-informatique: CAIMAN [Cancer: Artificial Intelligence for Multiomics Analyses]

Description du poste et candidature via le lien : https://www.inria.fr/fr/chaires-professeur-junior-recrutement-emploi

Pour toute information, vous pouvez contacter Alain Viari (alain.viari@inria.fr), Stéphane Ubeda (stephane.ubeda@inria.fr) ou Paulo Gonçalves (paulo.goncalves@inria.fr).

Keywords: Incremental learning; Deep learning; Frugal AI; Action recognition; Geometrical transformation invariance; Real conditions; Robustness.

Description
This subject aims at developing incremental learning algorithms for action recognition, with an emphasis on robustness and frugality. Incremental learning is a category of machine learning allowing to update the parameters of the decision model as the data is acquired. This type of approach proves particularly relevant in order to dynamically adapt the model to patterns not encountered in the data. With regard to the application studied, these patterns may correspond to new scenes or unexpected actions [1, 2], e.g. a dangerous act for a pedestrian for a road monitoring use case. The developments will focus on three complementary aspects: theoretical guarantees, algorithmic considerations and implementation for experimental validation. A particular attention will be paid to the memory and energy footprint of the techniques, as well as their robustness to different variations in the data and in the acquisition devices.

• Development of frugal incremental learning algorithms. Methodological developments will have to respond to two specificities. First, the relatively low proportion of unexpected actions as well as their diverse nature complicates their identification. It will then be necessary to consider, in an online and adaptive manner, the unbalanced and heterogeneous nature of the data within a machine learning framework. Second, in the perspective of a more frugal AI, we will study how to adapt this class of incremental algorithms for sparse models [3].

• Invariance and robustness of incremental algorithms. The same actions in the scenes can be seen from different perspectives, hence modifying the angles and changing the scales of the captured images. It will therefore be necessary to develop a decision model invariant to certain transformation groups, e.g. translations, scaling, rotation, etc. [4]. Departing from these standard transformations, one will also investigate how to devise a model invariant to the physical parameters of the acquisition device, such as the calibration settings. Last but not least, the decision model should also be robust to variations in the data stream due to its ever-evolving environment. Therefore, it should notably perform equally well under different lighting and weather conditions.

Host laboratory
Hubert Curien Lab UMR CNRS 5516, Saint-Etienne, France https://laboratoirehubertcurien.univ-st-etienne.fr.

Candidate profile
• Master or Engineer school in computer science, applied mathematics or related.
• Good Python programming skills. PyTorch experience is welcomed.
• Good knowledge of neural networks. Additional knowledge in probabilities, statistics and physical models would also be appreciated.
• High proficiency in English.

Application
Candidate must send the following documents to both olivier.alata@univ-st-etienne.fr and jordan.frecon.deloire@univ-st-etienne.fr as soon as possible:
• Cover letter with justification of your skills for the topic
• A complete Curriculum Vitae
• Transcript of your bachelor and master’s/Engineer school’s grades.
• CEFR level in English (except if university courses were taught in English)
• Any additional document: letter(s) of recommendation, publications, master thesis, etc.
Please feel free to contact us beforehand for any further pieces of information.

Funding
The selected candidate will obtain a 36-month funding.The net salary will be around 1700€. Additional paid teaching activities can be envisaged on demand.

References
[1] J. Ma, X. Tao, J. Ma, X. Hong and Y. Gong, “Class Incremental Learning for Video Action Classi fication”, ICIP 2021.
[2] J. Park, M. Kang and B. Han, “Class-Incremental Learning for Action Recognition in Videos”, ICCV 2021.
[3] M. Barlaud and F. Guyard, “Learning Sparse deep neural networks using efficient structured projec tions on convex constraints for green AI”, ICPR 2020

Description
During conversation, many studies have shown that interlocutors tend to progressively converge by adopting similar verbal and non-verbal behaviors. We also know that the level of alignment between participants can be correlated with the success of the interaction, the quality of information transfer and inter-mutual comprehension. The question is to know whether alignment is a consequence of our perception of other’s behavior, with a certain control and intention in doing that, or whether it is a deeper spontaneous mechanism, based on the fact that both participants form a unique system and behave synchronously in an equal manner during the interaction. In the first case, each participant has his/her own language model used in production and perception and tries to decode the other production from this model. In the second case, the hypothesis is that both participants share the same model and generate the same predictions at the same time, the speaker producing the message and the hearer inhibiting it. This second options is based on the predictive coding hypothesis, which can be observed at the brain level.

This Phd will address this question by investigating the different hypothesis. In a first step, it will consist in evaluating during conversation whether one unique model performs better than two separate. These models will be built in two different manners: 1/ on the basis of pre-trained models, fine-tuned thanks to existing conversational datasets, 2/ with multimodal models, taking into account all verbal and non-verbal features. A unique model based both speakers production will be compared with separate model by participant. In a second step, a model of alignment between participants will be built, based on the different features extracted from our datasets. Multimodal learning methods will be applied, focusing in particular on the question of feature interaction and modalities fusion. A correlation between the level of alignment and the behavioral prediction will be analyzed. In collaboration with neuroscientists, we will finally analyze the dynamics in the brain signal, looking for traces of synchronization predicted by our models.

Selection Criteria
– A master’s degree in Computer Science, Artificial Intelligence, Computational Linguistics, or a related field
– Knowledge of deep learning architectures, specifically Transformer models
– Skills in natural language processing, multimodality and interaction
– Interest in interdisciplinary work
– Good research skills
– Good writing and presentation skills
– Good programming skills

Information Enquiries can be made to Philippe Blache blache@ilcb.fr.

Applications
To apply for this vacancy, please send an email to blache@ilcb.fr. Please ensure that you upload the following additional documents:
– Resume
– Application letter, including your applicability to aforementioned selection criteria
– Grade list (preferably for both bachelor’s degree and master’s degree)
– Examples of Scientific Writing (e.g., scientific publication or master’s thesis)
– Reference letters

Only applications received before May 15th, 23:59 can be considered.

Un poste de Chaire de Professeur Junior (CPJ) est à pourvoir pour la rentrée 2023 à l’INSA Rouen Normandie, laboratoire LITIS, sur la thématique « Intelligence Artificielle pour une Mobilité Intelligente Sûre ».

Le/la candidat.e intègrera le département Informatique, Traitement de l’Information (http://iti.insa-rouen.fr) pour ses enseignements et au sein du LITIS l’équipe
– App (https://www.litislab.fr/equipe/app) ou,
– MIND (https://www.litislab.fr/equipe/mind) ou,
– STI (https://www.litislab.fr/equipe/sti),
selon l’adéquation de son profil aux thèmes de recherche de ces équipes.

Le profil complet est disponible sur Galaxie: https://www.galaxie.enseignementsup-recherche.gouv.fr/ensup/ListesPostesPublies/FIDIS/0760165S/FOPC_0760165S_4108.pdf

Date limite de candidature sur Galaxie : 27 avril 2023

Contact recherche : laurent.vercouter@insa-rouen.fr
Contact enseignement : geraldine.del_mondo@insa-rouen.fr

Advisors: Amaury Habrard (Professor) and Jordan Frecon-Deloire (Associate Professor)
Team: Data Intelligence
Host laboratory: Hubert Curien Lab UMR CNRS 5516, Saint-Etienne, France
Application deadline: May 1st, 2023
Keywords: Physics-guided models; Neural networks; Sparsity; Transfer learning; Optimization

Context:
In many physical systems, the governing partial differentiation equations (PDEs) are known with high confidence, but simulating a numerical solution can be prohibitively expensive. In other contexts, the PDEs are unknown (or partly known to some extent) and unveiling them from experimental data is the central goal since they could help in shedding some lights on the underlying physical process. Recently, physics-guided machine learning models have shown to be a promising tool in both above-mentioned scenarios. They rely on neural networks in order to simulate the physical quantities of interest at various temporal and spatial positions. Training such neural networks entails to incorporate physical constraints, usually in the form of a PDE and boundary conditions, and/or to be able to generate plausible simulated data reproducing the experimental data at hand.

Description:
The originality of the present thesis proposal is to embrace the extreme setting encountered in surface engineering, that is limited prior physical knowledge and few experimental data. In order to overcome both limitations, the thesis will develop a unified end-to-end framework from the physics modeling to the algorithms used for training physics-guided models:
• Develop novel regularization techniques, possibly on latent representations, to incorporate partial physical constraints (e.g., conservation laws or decreasing lypuanov exponents along the trajectories) and incertitude about prior knowledge.
• Promote sparse methods to circumvent the lack of data. Investigate neural network sparse by design preventing redundancy and over-parametrization of trivial mappings. Explore optimization strategies to promote sparse weights by leveraging the properties of the compositional form of neural networks.
• Discover the semantic of PDEs from few data with the dual objective of adapting to new physical environments (e.g., different surface properties or laser characteristics). The advances that will be carried out in machine learning will allow to better understand the physics underlying the mechanisms of laser/radiation-matter interaction, enabling to address numerous societal challenges in the fields of space, nuclear, defense, energy or health.

Candidate profile:
• Master in computer science, machine learning, applied mathematics or related. Outstanding applications from physicists will also be considered
• Good Python programming skills. PyTorch experience is welcomed
• Good knowledge of neural networks
• Basic knowledge on optimization and partial differential equations
• High proficiency in English

Application:
Candidate must send the following documents to both amaury.habrard@univ-st-etienne.fr and jordan.frecon.deloire@univ-st-etienne.fr as soon as possible:
• Cover letter with justification of your skills for the topic
• A complete Curriculum Vitae
• Transcript of your bachelor and master’s grades (Semester 1 and 2, Semester 3 if available)
• CEFR level in English (except if university courses were taught in English)
• Any additional document: letter(s) of recommendation, publications, master thesis, etc. Please feel free to contact us beforehand for any further pieces of information.

Host laboratory:
The Hubert Curien Lab combines internationally recognized experts in both machine learning and laser-matter interaction. The present thesis is in line with the Hubert Curien Lab commitment to foster the development of new joint methodological contributions at the interface between machine learning and surface engineering.
More information: https://laboratoirehubertcurien.univ-st-etienne.fr.

Optimal transport has recently attracted much interest in the machine learning community. Its inherent ability to compare probability measures is now at the heart of several recent machine learning successes. We aim to use an optimal transport distance to create a kernel-based graph-level prediction model. In this model, the new node representation will be learned by a graph neural network (GNN). Then, the distance of the new representation will be calculated by an optimal transport to create a better kernel for the prediction model. The prediction will be performed by support vector machines (SVM) or Gaussian random process (GPP) which are fully differentiable.

Position :
• CDD (18 mois)
• Rémunération entre 2604.47 € et 2832.42 € bruts/mois selon expérience
• À pourvoir dès que possible et au plus tard au 1er juin 2023
• Localisation : INRAE, centre Occitanie Toulouse, Castanet Tolosan, unité MIAT https://miat.inrae.fr/

Les missions du (de la) postdoc recruté·e se dérouleront dans le cadre du projet SubtilNet fédérant les compétences informatiques / mathématiques / statistiques de l’équipe SaAB sur l’inférence de réseaux et la biologie des systèmes. Ce projet se positionne sur l’étude des méthodes mathématiques permettant de reconstruire des réseaux biologiques. Il a pour ambition, en se basant sur un réseau réel exhaustif de la bactérie Bacillus subtilis, de mieux évaluer les méthodes d’inférence actuelles et leurs caractéristiques. Le but final est d’améliorer l’état de l’art en termes de méthodes d’inférence en se rapprochant de la réalité biologique et en intégrant, dans les modèles, des informations biologiques pertinentes.

Profil recherché :
Doctorat en informatique, bioinformatique, machine learning ou statistique (thèse soutenue depuis moins de 3 ans). Nous recherchons un·e candidat·e ayant une expérience avérée en analyses de données omiques, biologie cellulaire et/ou biologie des systèmes. Des compétences sur l’inférence ou l’analyse de réseaux sont également souhaitables ou, à défaut, des compétences en apprentissage automatique ou statistique. Enfin, un bon niveau en programmation, de préférence avec le langage de programmation R, est requis. Une connaissance de Python, Matlab, … serait un plus. Compte tenu des nécessaires interactions entre les divers membres du projet, une aptitude au travail en groupe serait appréciée. Le (la) candidat·e doit également posséder un très bon niveau d’anglais scientifique.

Présentation de la structure d’accueil :
L’Unité de Mathématiques et Informatique Appliquées de Toulouse https://mia.toulouse.inra.fr est une unité propre (UR875) d’INRAE https://www.inrae.fr. MIAT a pour mission scientifique de développer et mettre en œuvre des méthodes mathématiques et/ou informatiques pertinentes pour résoudre des problèmes identifiés avec nos collaborateurs qui sont issus principalement d’autres départements d’INRAE. L’unité comporte actuellement deux équipes de recherche (SciDyn et SaAB) et trois équipes de service (Plateformes BIOINFO, RECORD et SIGENAE).

Comment postuler ?
• envoyer CV et lettre de motivation à nathalie.vialaneix@inrae.fr et anne.goelzer@inrae.fr. On attend que la lettre de motivation mette en avant les compétences et expériences du (de la) candidat⋅e en lien avec les attentes du poste. Une lettre de recommandation issue des encadrants de thèse et/ou précédents postdoc peut compléter le dossier.

Un poste de professeur.e des universités en intelligence artificielle sera publié lors de la session synchronisée au Conservatoire National des Arts et Métiers (Paris).
Le laboratoire d’accueil est le CEDRIC (EA4629) : https://cedric.cnam.fr/
Le profil détaillé et les différents contacts se trouvent dans ce document : http://cedric.cnam.fr/lab/wp-content/uploads/2023/01/CnamParisPUIA2023.pdf

Dear all,
we have an open 24-month post-doc position on Computer Vision and semi-supervised representation learning at the LIRIS lab in Lyon.
This is part of a French-German (ANR) research project on resource-efficient tunnel construction involving several academic and industrial partners from both countries in the domains Computer Science (AI) and Geology/Geophysics. The post-doc will be supervised by Stefan Duffner (LIRIS, Imagine team, INSA Lyon) and Catherine Pothier (LIRIS, Imagine team, INSA Lyon).
Requirements :
– Master in computer science, CV, ML, applied mathematics or related
– Good python programming skills
– Autonomy and interdisciplinary curiosity
More details can be found here: https://partage.liris.cnrs.fr/index.php/s/6RjzwjfYJ6tnJ3H
Best regards,
Stefan Duffner

Autonomous agents require reasoning and planning strategies for performing tasks.

Natural language can serve for building and clarifying the planning strategy, and therefore the actions done by a robot. Several works have addressed instruction identification as abstract representation [2, 4, 5] or natural language expression, but the limited data supervision is often a challenge [3, 5]. To tackle this issue, we propose to develop interactive training processes, which imply asking humans to label situations with sentences, with strong care on limiting interactions to a few relevant situations, to reduce human effort. The underlying assumption is that the compositionality of language is correlated to compositionality in the agent’s world. In this internship, we envision working on the generation of natural language instructions and improve currents model. Our objective is to enhance the semantics behind objects to identify the most relevant actions/sub-actions.

More info: https://mlia.isir.upmc.fr/jobs.html

In this internship, we assume that the deployment of virtual assistants can be done step by step over different countries in the world and, thus, that virtual assistants will face different languages at different timestamp. We propose a continual learning setting in which the task is fixed, but the stream is based on different languages. The model learn the knowledge of language peculiarities. Therefore, to satisfy the initial condition of virtual assistants to address different languages, we therefore need to ensure that our task-based model does not forget previous languages while training on new ones.

Two preliminary works have been done: 1) [Coria et al., 2022], investigating BERT’s cross-lingual transfer capabilities in two continual sequence labeling tasks. 2) [Gerald and Soulier, 2022] designing continual learning streams for information retrieval. In practice, we will focus on the Massively Multilingual NLU 2022 data [FitzGerald et al., 2022], which includes slot-filling and NER tasks for 51 languages in parallel. The objective of the internship will be to 1) build a stream of languages for a given task, 2) run baseline models in the stream, and 3) design a continual learning model for cross-lingual transfer.

More info: https://mlia.isir.upmc.fr/jobs.html

In this internship we are interested in a special case of text generation, which is data-to-text generation. In this setting, the task is to generate sentences in natural language based on structured or semi-structured data. To provide a data to text example, a famous academic dataset is made of statistics of baseball games paired with human written summary of the game, that we ultimately want to the system to learn to generate. Beyond this toy example, data to text is of the utmost practical interests in many scenarios such as finance, . . . This task is a special case in text generation and comes with its own specific challenges. The data to text models are prone to hallucinations, that is generating grammatically correct but irrelevant and out of the blue sentences . Moreover, the inputs being structured or semi-structured data, this calls for alternative solution to encode w.r.t. standard texts inputs made of sequence of tokens arranged as sentences.

The objective of the internship is to develop a neural data to text system able to personalize the text generation. Based on a previous work, we will first focus on a movie dataset in which we dispose of movie tabular description (the data), and reviews. The objective will be to personalize review for a given user. Secondly, while there exists studies on how to evaluate data to text generator, to the best of our knowledge none consider style transfer/text personalization for text generation besides. As such, finding means to perform style transfer evaluation for data to text generators is fully part of the internship, on top of finding neural solution to perform style transfer aware data to text. The evaluation we seek has to be automatic or semi automatic.

More info: https://mlia.isir.upmc.fr/jobs.html

Contexte et problématique
L’apprentissage autosupervisé de représentation est un domaine actuellement actif et qui traite de problématiques telles que:
– L’apprentissage sans supervision humaine, et l’utilisation de données non annotées;
– L’apprentissage des représentations générales applicables à des tâches en aval;
– Le pré-entraînement de réseaux de neurones pour le transfert d’apprentissage.

Ces problématiques sont clef en intelligence artificielle et en traitement d’images, et une avancée dans leur résolution permettrait l’utilisation de l’apprentissage profond dans des domaines d’applications souffrant d’un manque de données. Cela aiderait aussi à réduire les coûts de calcul pour l’entraînement de tâches en aval comparativement à la méthode habituelle bout en bout.

Les approches récentes d’apprentissage autosupervisé de représentation visuelle, reposent sur des réseaux de neurones siamois effectuant une tâche contrastive de discrimination d’instance. Et bien que ces méthodes diffèrent légèrement les unes des autres en termes d’architecture et de tâche apprise, le principe sous-jacent reste le même: l’objectif est d’apprendre un espace latent où deux vues d’une même image (des augmentations de données), ont des représentations similaires, le tout en utilisant différentes astuces pour éviter un effondrement vers des solutions simples. L’intuition est que des entrées synthétiques similaires représentent des concepts similaires, et donc doivent avoir des représentations similaires.

Cependant, bien que ces méthodes évitent un effondrement total des représentations, elles souffrent quand même d’un effondrement partiel de leurs dimensions, c’est-à-dire que même si les représentations sont différentes les unes des autres, une partie de leurs dimensions sont fortement corrélées entre elles. D’un point de vue théorique, cela montre que les représentations n’exploitent pas pleinement les dimensions mises à disposition et que les diverses contraintes imposées par ces modèles pour éviter un effondrement ne fonctionnent que partiellement. D’un point de vue pratique cela tend à produire des représentations dont la qualité est dégradée.

D’autre part, avant l’apparition de ces méthodes dites contrastives, l’état de l’art en apprentissage de représentation reposait sur l’utilisation d’auto-encodeurs variationnels. Ces derniers sont une amélioration des auto-encodeurs où les représentations prennent la forme de distributions statistiques au lieu de points dans un espace latent. L’ajout de cet aspect probabiliste permet de lisser l’espace de représentations, mais aussi de démêler les différentes dimensions des représentations apprises, menant ainsi à des représentations de meilleure qualité comparées à celles apprises par un simple auto-encodeur.

Description du sujet
Bien que l’avancée obtenue avec les auto-encodeurs variationnels ait eu un impact important dans le domaine de l’apprentissage de représentation, les méthodes contrastives de l’état de l’art ne capitalisent pas sur ses principes. Sachant que les auto-encodeurs variationnels permettent d’apprendre des caractéristiques décorrélées et que la décorrélation de caractéristiques a un impact positif sur l’effondrement dans les représentations.

Le sujet du stage consiste à étudier si les représentations apprises par ces méthodes contrastives récentes peuvent bénéficier d’un aspect variationnel avec les tâches suivantes:
– proposer une modification des méthodes actuelles pour que les représentations produites soient des distributions et non des points dans l’espace de représentation
– étudier comment régulariser les distributions produites, comme c’est le cas dans les auto-encodeurs variationnels, afin d’éviter toute forme d’effondrement. Cela nécessitera un travail d’adaptation étant donné que les méthodes d’apprentissage contrastif récentes utilisent un espace latent contraint à une hypersphère, là où les auto-encodeurs variationnels demandent un espace latent sans contrainte