The COPHY team (Computation, Cognition and Neurophysiology) is led by Mathilde Bonnefond and Jérémie Mattout. It is composed of 5 researchers, 2 research engineers, 1 clinician and 11 national and international docs and postdocs. It is part of CRNL which is a large INSERM/CNRS/University Lyon 1 research center, gathering more than 450 people. It is located in the Lyon East Hospital Pole, a unique site combining medical care (4 hospitals respectively dedicated to neurology, psychiatry, cardiology and pediatrics), basic and clinical research (from genes and cells to cognition and behaviour; from bench to patient), as well as high-tech technological platforms (incl. Neuroimaging, MEG, SEEG, virtual reality).

Scientific context : In various sensory modalities, in health and disease, we mainly study human perception in the aim of unravelling its underlying cognitive, computational and neurophysiological mechanisms. Our team aims at innovating at the theoretical, methodological, experimental and clinical levels. In particular, we study neuronal encoding and communication (e.g., the emergence and functional role of oscillations; the computation behind the contextual modulations of cortical responses) in relation to brain disorders. This also guides the design of neurotechnologies (brain-computer interfaces) for research and clinical applications: namely ADHD, motor deficits and disorders of consciousness. Therefore, and beside behavioural and classical neuroimaging approaches (electrophysiology, fMRI), we use multimodal high-resolution neuroimaging (e.g., laminar MEG and fMRI) and computational modelling (e.g., Bayesian inference, Dynamic Causal Models of neuronal and cognitive processes; Spiking/Deep Neural Networks).
Methods and models lie at the heart of our research. With this call, we aim at strengthening this core axis that is key to translating our theoretical hypothesis into testable quantitative predictions. Our team has recently evolved so that our hypothesis driven approach now spans several spatial and temporal scales (e.g., from neurotransmitters to behaviour; from trial-by-trial dynamics to neurodevelopment). Instantiating these hypotheses requires to build on most recent advances in computational neurosciences, for developing biologically plausible models of cognitive processes and their underlying physiological mechanisms.
We thus develop models to be confronted with actual data. We actually use models to strongly guide the design of our experiments and provide new tools for data analysis (new algorithms that can be implemented in open software and shared with the whole team and beyond).
We attach particular importance to the valorisation of computational models and methods, be it through their publication and diffusion, or by virtue of their impact on translating our research into societal and clinical applications.
Finally, we also value methodological developments that support the emergence of new functional techniques (e.g. for the analysis of data from the new generation of MEG systems: optically pumped magnetometers (OPMs)).

We seek for a young researcher whose profile will strongly complement our team. We expect the candidate to lead her/his own research program in computational neuroscience, but in strong interaction with the topics and experimental projects in the team.
The important skills and topics we would like to further develop in the team include but are not limited to:
– Spiking neural networks
– Bayesian models (e.g. Active Inference)
– Dynamical systems
– Brain-Computer Interfaces
– Machine Learning
A previous experience in software development, teaching and human electrophysiology is not mandatory but will be a plus.

Application should contain a full resume (CV), a research and motivation statement (1 page max.) as well as the names of 2 references. Please send your application as a single document in pdf format,to both mathilde.bonnefond@inserm.fr and jeremie.mattout@inserm.fr

Advisors: Olivier Alata (Professor) and Jordan Frecon-Deloire (Associate Professor)
Host laboratory: Hubert Curien Lab UMR CNRS 5516, Saint-Etienne, France
Starting date: Early 2024 – at your earliest convenience

Description: This subject focuses on the development of a learning algorithm for action recognition [1] with a specific emphasis on its environmental impact. On one hand, we aim to minimize its memory usage by learning and imposing both physics and geometric constraints. On the other hand, we will employ incremental learning techniques to ensure its long-term sustainability. 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.

– Physics-aware models. Recently, physics-guided machine learning models [2] have shown to be a promising tool to incorporate known physical priors in the learning process of the model. Conversely, some approaches have been designed to discover the underlying physics from observations [3]. In this thesis, the candidate should contribute to the development of physics-based models in order to both learn and simulate the dynamics of complex systems. Methodological contributions will be applied to road traffic monitoring to learn the dynamics of « normal » baselines from which rare « dangerous » events can be detected.
– Equivariance and invariance: Another research direction will focus on the development of physics-aware models invariant/equivariant to geometric and physical parameters. For instance, 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 that is, by design, invariant to certain transformation groups (scaling, translation, etc.) [4]. Departing from these standard transformations, the model should also be robust to variations in the data stream due to its ever-evolving environment (e.g., different lighting and weather conditions).
– Incremental learning: In addition to the energy consumption of the machine learning model, another significant factor contributing to its ecological footprint is its long-term sustainability. In order to dynamically adapt the model to patterns not yet encountered in the data, the candidate should tailor the latest advances in incremental learning to update both the model parameters and the physical priors as the data is acquired. Contributions could explore the connections with incremental learning for physics-informed neural networks [5].

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 jordan.frecon.deloire@univ-st-etienne.fr and olivier.alata@univ-st-etienne.fr as soon as possible:
Cover letter + CV
Transcript of your bachelor and master’s/Engineer school’s grades.
CEFR level 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.

References
[1] J. Park et al. “Class-Incremental Learning for Action Recognition in Videos”, ICCV 2021.
[2] M. Raissi et al. “Physics-Informed Neural Networks:”, Journal of Computational Physics 2019.
[3] M. Buisson-Fenet et al., Recognition Models to Learn Dynamics from Partial Observations with Neural ODEs », TMLR 2023.
[4] J. Bruna and S. Mallat, “Invariant scattering convolution networks », IEEE TPAMI 2013.
[5] A. Dekhovich et al. “iPINNs: Incremental learning for Physics-informed neural networks », preprint ArXiv 2023.

Location: Laboratoire d’Informatique de Grenoble, Grenoble, France
Position Type: Full-Time (18 months)

About us:
The Grenoble computer science laboratory (LIG – Laboratoire d’Informatique de Grenoble) of the University Grenoble Alpes (UGA) is a leading laboratory that fosters and promotes innovative research in various fields of computer science. The machine learning team of LIG has put forth a project with the goal of developing an advanced chatbot that incorporates the analysis of physiological signals from audio and video data, primarily for stress detection, to improve its interactions. This initiative has been selected as a groundbreaking project by the university’s technology transfer company.

What’s intended? to take an amazing scientific idea and make it into a profitable business. This endeavor marks a significant step towards the future, as ultimately we aim to found a firm that will transform the human resources’ landscape.

Job Description and Responsibilities:
The development of the multimodal sentiment analysis module that will analyze audio and video data involves several steps shown below.

Data collection and annotation:
The first step will be to gather a diverse and well-annotated dataset consisting of audio and video recordings with associated sentiment labels.

Feature extraction:
This step will be followed by a step of extracting relevant features from the audio and video data. For audio, this includes acoustic characteristics such as pitch, intensity, and energy. For video, features will involve facial expressions, body language and motion analysis. We will examine the case where video and audio data cannot match certain time intervals. To do this, we consider data that does not correspond as missing data that we will cover or complete with the model that will be developed.

Labeling and Training Models:
This will be followed by a training pipeline in order to label audio and video data with sentiment information. To do this, convolutional neural networks (CNN), recurrent neural networks (RNN) and transformers on the labeled audio and video data obtained will be considered

Integration of audio and video analysis:
The final prediction will be based on a combination of the outputs of the audio and video sentiment analysis models.

Model Evaluation:
The performance of the multimodal sentiment analysis module will be carried out by conducting extensive testingon various audio and video inputs and implementing several metrics like precision, recall, and F1 score.

Adjustment and optimization:
The evaluation phase will refine the model and adjust the architecture, hyperparameters and training process in order to improve performance and resolve limitations encountered during the evaluation phase.

Real-time analysis and deployment:
The selected model will then be optimized in order to be deployed and analyze sentiments in real time conjugated with the development of a chatbot.

For this position, we are seeking highly skilled engineers or post-doctoral researchers with expertise in the domains of deep-learning and . Ideal candidates should possess experience in various areas, including but not limited to

• data processing for both pre-training and fine-tuning,
• optimization,
• training and inference of deep language models,
• expertise in multiview learning,
• sentiment analysis

Qualifications:

• Computer Engineering, or Master’s degree in Computer Science (Machine Learning and/or Natural Language Processing).
• PhD in AI, computer science, data science, or related technical fields
• Proficiency in machine learning frameworks at scale from inception to business impact
• Excellent problem-solving and analytical skills.
• Experience with distributed computing and cloud platforms is a plus.
• Strong communication and collaboration skills.
• A passion for staying updated with the latest developments in the field.

Benefits:

• Salary will be adapted with respect to the qualification and experience.
• Possibility to join the startup (which is expected to be launched in current 2025) following the University program.
• Opportunities for career growth and professional development.
• Collaborative and innovative work environment.
• Contribution to impactful projects and research in the field of natural language processing.
• Direct relationship with the founding team in a day-to-day basis.

How to Apply:
If you are passionate about advancing the field of large language models and eager to work with a team of experts, we encourage you to apply by sending your resume and a cover letter detailing your relevant experience to Massih-Reza.Amini@imag.fr.

Location: Laboratoire d’Informatique de Grenoble, Grenoble, France
Position Type: Full-Time (18 months)

About us:
The Grenoble computer science laboratory (LIG – Laboratoire d’Informatique de Grenoble) of the University Grenoble Alpes (UGA) is a leading laboratory that fosters and promotes innovative research in various fields of computer science. A project proposed by APTIKAL, the machine learning group of LIG, with the objective of developing a cutting-edge Large Language Model, has been selected as a pioneering initiative by the university’ transfer company.

What’s intended? to take an amazing scientific idea and make it into a profitable business. This endeavor marks a significant step towards the future, as ultimately we aim to found a firm that will transform the human resources’ landscape.

Job Description:
The specificity of this solution will be to learn from a highly specialized domain with scarce data but for which a complete knowledge graph exists. For this position, we are seeking highly skilled engineers or post-doctoral researchers with expertise in the domains of generative AI and Natural Language Processing (NLP). Ideal candidates should possess experience in various areas, including but not limited to

• data processing for both pre-training and fine-tuning,
• optimization,
• training and inference of deep language models,
• ensuring alignment of LLMs,

Responsibilities:

• Model Development: Play an active role in conducting experiments, encompassing tasks such as designing experiment specifications, developing reusable code, conducting evaluations, and arranging results.
• Innovation: Formulate innovative approaches, tools, and infrastructure to advance the frontiers of large language models.
• Scalability: Work on the scalability of language models, ensuring they can operate efficiently on distributed systems.
• Testing and Evaluation: Develop evaluation metrics and methodologies to assess the performance and robustness of language models, and contribute to their continuous improvement.
• Model Integration: Work closely with software engineers to integrate language models into various applications and services, ensuring seamless functionality.
• Research Collaboration: Collaborate with internal and external researchers and scientists to advance the state-of-the-art in language models.
• Documentation: Create clear and comprehensive documentation for model development, integration, and usage.

Qualifications:

• Computer Engineering, or Master’s degree in Computer Science (Machine Learning and/or Natural Language Processing).
• PhD in AI, computer science, data science, or related technical fields
• Proficiency in machine learning frameworks at scale from inception to business impact
• Excellent problem-solving and analytical skills.
• Experience with distributed computing and cloud platforms is a plus.
• Strong communication and collaboration skills.
• A passion for staying updated with the latest developments in the field.

Benefits:

• Salary will be adapted with respect to the qualification and experience.
• Possibility to join the startup (which is expected to be launched in current 2025) following the University program.
• Opportunities for career growth and professional development.
• Collaborative and innovative work environment.
• Contribution to impactful projects and research in the field of natural language processing.
• Direct relationship with the founding team in a day-to-day basis.

How to Apply:
If you are passionate about advancing the field of large language models and eager to work with a team of experts, we encourage you to apply by sending your resume and a cover letter detailing your relevant experience to Massih-Reza.Amini@imag.fr.

Le sujet de ce post-doctorat est mené par le L@bISEN et s’inscrit dans le cadre du projet HYPOCHIR en collaboration avec le LaTIM (Laboratoire de traitement de l’information médicale), une unité mixte de recherche (UMR1101) associant l’Inserm, l’Université de Bretagne Occidentale (UBO), l’IMT Atlantique et le CHRU de Brest. L’objectif de ce projet est de développer une approche novatrice de cartographie assistée par l’imagerie hyperspectrale pour la planification orthopédique en temps réel.

Fiche d’identité du poste :

• Établissement de rattachement : Yncréa Ouest, Établissement d’enseignement supérieur privé d’intérêt général (EESPIG), sous contrat avec le ministère de l’enseignement supérieur et de la recherche ;
• Unité de recherche : L@bISEN ;
• Équipe de recherche : LSL (Light – Scatter – Learning) ;
• Lieu de travail : campus de Brest ;
• Durée du contrat : 2 ans / 18 mois ;
• Salaire : 35k/an € voir selon expérience ;
• Possibilité de participer aux activités d’enseignements d’Yncréa Ouest.

Objectif : L’objectif principal de ce projet est de développer une approche rapide et efficace de segmentation des images hyperspectrales pour différencier et identifier automatiquement les différentes structures anatomiques et les instruments chirurgicaux. En exploitant les caractéristiques spectrales distinctives de chaque élément, il sera possible de cartographier précisément le bloc opératoire, facilitant ainsi la planification préopératoire, la navigation intraopératoire et l’assistance en temps réel lors des interventions chirurgicales orthopédiques.

Des expérimentations seront menées sur des objets anatomiques présentant des caractéristiques physiques bien définies afin de valider l’efficacité des différentes méthodes de segmentation proposées. L’objectif principal du projet HYPOCHIR est d’améliorer la précision et la sécurité des interventions chirurgicales orthopédiques en exploitant les avantages de l’imagerie hyperspectrale pour la cartographie du bloc opératoire. Grâce à la segmentation automatique basée sur les signatures spectrales, le projet vise à optimiser les résultats obtenus pour les patients dans le domaine de l’orthopédie. En offrant une meilleure précision chirurgicale, HYPOCHIR contribuera à améliorer les pratiques en chirurgie orthopédique et à optimiser les résultats pour les patients concernés.

Mots-clés : Imagerie hyperspectrale, chirurgie orthopédique, Segmentation d’images, apprentissage profond, Planification préopératoire, Navigation intraopératoire, Assistance en temps réel.

Le ou la candidat(e) doit détenir :

• Doctorat en informatique, en science des données, en apprentissage automatique ou dans un domaine connexe.
• Le.la candidat(e).e doit avoir effectué.e un séjour d’au moins 18 mois hors de la France à partir du 1er mai 2020 et le démarrage du projet (une personne venant de l’étranger est totalement éligible pour ce poste).
• Solides compétences en programmation, en particulier en Python et en langages de deep learning tels que TensorFlow ou PyTorch.
• Compréhension approfondie des réseaux de neurones profonds, de l’apprentissage automatique pour la segmentation d’images.
• Capacité à travailler de manière autonome et à résoudre des problèmes de manière créative.
• Excellentes compétences en communication orale et écrite, avec la capacité de présenter des résultats de recherche de manière claire et concise.
• Une expérience préalable dans le domaine de l’hyperspectral serait un atout majeur.
• Motivé.e et passionné.e par le domaine médical.

Pour candidater :
Faire parvenir les documents suivants :

• Curriculum vitae (CV) ;
• Lettre de motivation ;
• Tout autre document jugé utile pour enrichir le dossier de candidature (lettres de recommandation, articles scientifiques, rapport de thèse, etc.)

Si vous avez une expérience concernant les agents conversationnels, ou souhaitez approfondir ce thème, n’hésitez pas à candidater ici : https://emploi.cnrs.fr/Offres/CDD/UMR5191-MAXPEN-022/Default.aspx !
Les dates du contrat peuvent être aménagées (frederic.armetta@univ-lyon1.fr).

@ ICAR (InSitu Team), Ecole Normale Supérieure de Lyon
Durée: 12 mois a discuter – Début fin 2023 ou avant, à discuter
Projet: PepperMint financé par ASLAN Labex (https://peppermint.projet.liris.cnrs.fr/)
Partenaires: LIRIS (SyCoSMA, SAARA Teams), ICAR (InSitu Team), Université d’Oulu-Finlande (GenZ),
Encadrement : Dr Heike Baldauf-Quilliatre, Dr Frédéric Armetta, Dr Mathieu Lefort

Contexte

Le projet PepperMint (https://peppermint.projet.liris.cnrs.fr/) est financé par le Labex ASLAN (https://aslan.universite-lyon.fr/). Il propose une étude exploratoire des pratiques incarnées de prise de tours de parole dans l’Interaction Humain-Robot (IHR) permettant d’améliorer les capacités sociales des robots et rendre l’IHR plus naturelle pour les humains. Le projet initie une coopération entre des chercheurs en Intelligence Artificielle (IA) (LIRIS) et Analyse Conversationnelle (AC) (ICAR et GenZ Oulu – Finlande). Il étudie si et comment les résultats de l’AC sur l’interaction naturelle peuvent être utilisés pour développer des modèles d’IA innovants et efficaces pour l’IHR.

Le cas d’application choisi correspond à un robot Pepper ayant pour rôle d’informer et orienter les utilisateurs dans une bibliothèque. Afin d’accomplir sa tâche, le robot utilise actuellement pour sa politique de dialogue une machine à états finis, celle-ci étant conçue pour gérer la sémantique de la conversation (et fournir des réponses à des questions telles que où trouver un livre de biologie par exemple). Du point de vue du robot, le cours de la conversation passe d’un état à l’autre, grâce à des transitions déclenchées par des intentions uniquement orientées vers la tâche, reconnues dans les phrases prononcées par l’utilisateur du système. Bien entendu, une interaction aussi systématique ne rend pas l’interaction naturelle, car elle exclut le caractère émergent de l’interaction. En effet chaque étape de la conversation peut mener à différents types de réponses, suivant la manière dont les humains l’interprètent.

L’interaction doit ainsi suivre des règles spécifiques. Par exemple, une offre (« comment puis-je vous aider ? ») doit être suivie d’une acceptation (« oui, je voudrais … ») ou d’un rejet (« eh bien, hum … »). De nombreux motifs séquentiels de ce type sont imbriqués dans l’interaction et sont au moins aussi importants que le contenu sémantique porté par l’échange pour rendre l’interaction naturelle.

Le projet en cours a déjà recueilli quelque 30 heures de sessions d’enregistrement à partir desquelles de tels schémas sont actuellement analysés et étiquetés. Afin de proposer un meilleur schéma interactif, nous prévoyons de reconnaître et qualifier ces motifs afin de les exploiter pour rendre la gestion de la conversation plus naturelle et définir la « bonne » interprétation et la « meilleure » réponse pour l’utilisateur. Plus précisément, l’analyse du texte parlé nécessitera une reconnaissance des intentions émises (NLU) basée sur l’apprentissage (Few-Shot Learning) incluant ce niveau d’interaction. La politique de dialogue devra alors intégrer ce canal complémentaire (ceci pourra être fait grâce à une simple adaptation de la machine à états finis mais un cadre de politique de conversation plus générique et innovant dédié à notre approche pourrait être développé).

L’objectif de ce poste est ainsi d’utiliser l’ensemble des données annotées pour des méthodes d’apprentissage automatique afin de proposer un nouveau modèle d’IA pour l’HRI (Human Robot Interaction). L’apprentissage sera basé sur la modalité du langage, mais pourra également être étendu à d’autres modalités également collectées qui pourraient avantageusement aider à la détection.

Pour plus de détails et pour postuler merci de passer par notre page welcome to the jungle: https://www.welcometothejungle.com/fr/companies/yanport/jobs/data-scientist-r-d-jeune-docteur-h-f_paris

Quelques mots sur le contexte de ce recrutement:
Aujourd’hui nous cherchons à franchir un cap, développer plusieurs nouveaux produits et accompagner la croissance de Yanport. Pour cela, nous recherchons un (futur) Docteur (H/F) en Mathématiques/Statistiques/Econométrie brillant et engagé, qui au sein de l’équipe technique, intégrera le pôle Data/R&D.

Ton rôle en tant que Data Scientist R&D:
– Approfondir des sujets passionants : Tu participeras à la définition des problématiques de R&D et de les mettre en oeuvre afin de valoriser nos données et enrichir nos produits. Nos sujets de R&D sont variés et passionants; estimation du prix des biens immobiliers, extraction d’information des descriptions (NLP) ou des images (computer vision), prévision de l’évolution des prix à court/moyen/long termes…
– Garant de notre veille scientifique : La littérature scientifique évolue particulièrement vite et tout projet de recherche se base sur une analyse pointue de l’état de l’art. Tu seras donc l’un des garants de notre veille scientifique en matière de modélisation et des applications aux données immobilières.

Ce que nous attendons de toi:
– Doctorat (obtenu ou en cours de finalisation) en Mathématiques/Statistiques/Econométrie (ou autre sujet pertinent).
– Un connaissance de langages informatiques (en particulier Python) serait appréciée.
– Idéalement tu as porté des travaux portant sur des données spatio-temporelles.
– Une appétence pour les enjeux de l’immobilier sera un vrai plus.
– Tu as une bonne capacité d’écoute, d’échange et de communication.

Il est indispensable de ne jamais avoir été embauché en CDI après l’obtention de la thèse (contrainte impérative pour respecter les critères du CIR).

Thesis general information:
– Institutional affiliation: Yncréa Ouest, private institution of higher education of general interest (EESPIG), under contract with the French Ministry of higher education and research ;
– Research unit: L@bISEN ;
– Research team: LSL (Light – Scatter – Learning) ;
– Thesis location: campus of Brest or Caen, Yncréa Ouest (to be determined) ;
– Contract duration: 3 ans ;
– Salary: € 2044.12 (gross monthly) ;
– Possibility of taking part to the teaching activities of Yncréa Ouest.

Application context:
Manual detection of chromosomal anomalies in conventional cytogenetics can be a lengthy and tedious process, especially when it comes to identifying subtle or complex abnormalities involving small parts of chromosomes. The consequences of these anomalies on patients’ health are significant, making it crucial to detect them quickly and accurately.

Objective of the thesis:
This thesis aims to develop a deep learning model for the automatic detection of structural anomalies in chromosomes using both conventional and unconventional imaging techniques.

Keywords: metaphase images, karyotype images, chromosomal anomalies, deep learning models, automatic detection, data preprocessing, experimental validation.

Proposed methodology:
-Data preparation and preprocessing:
-Establishment of an image acquisition protocol for metaphase or karyotype chromosome images;
-Exploration and development of deep learning techniques to enhance the resolution and the contrast of obtained images;

Deep learning models for chromosome structure analysis:
-Design and training of deep learning models specifically tailored for the analysis of chromosome structure;

Experimental validation and comparison:
– Validation of the results obtained with the deep learning model using real datasets;
– Comparison of the deep learning model’s performance against existing methods.

Expected results:
This thesis will result in the development of a deep learning model capable of automatically detecting structural anomalies in chromosomes. The expected outcomes include a significant increase in the accuracy and efficiency of anomaly detection, the ability to identify subtle or complex abnormalities, and a reduction in the time and costs associated with manual detection.

The candidate must have:
-A Master’s degree in Computer Science or an Engineering degree with a specialization in Computer Science, Image Processing, or Artificial Intelligence. Having taken relevant courses or completed internships related to medical imaging will be advantageous;
-Strong knowledge in machine learning, particularly deep learning;
-Solid programming skills, especially in Python, and familiarity with common libraries in machine learning and deep learning;
– A proficient knowledge of the English language.

To apply: Please send the following documents:
-Curriculum vitae (CV);
-Cover letter;
-Master’s degree academic transcript (or equivalent);
-Any other document that you believe would enhance your application (letters of recommendation, scientific articles, Master’s thesis report, etc.)