En poursuivant votre navigation sur ce site, vous acceptez le dépôt de cookies dans votre navigateur. (En savoir plus)
Portail > Offres > Offre UMR7271-VIVROS-038 - Chercheur en Apprentissage symbolique et non symbolique dans un contexte fédéré H/F

Researcher Symbolic and non-symbolic learning in federated learning context H/F

This offer is available in the following languages:
Français - Anglais

Date Limite Candidature : jeudi 8 décembre 2022

Assurez-vous que votre profil candidat soit correctement renseigné avant de postuler. Les informations de votre profil complètent celles associées à chaque candidature. Afin d’augmenter votre visibilité sur notre Portail Emploi et ainsi permettre aux recruteurs de consulter votre profil candidat, vous avez la possibilité de déposer votre CV dans notre CVThèque en un clic !

General information

Reference : UMR7271-VIVROS-038
Date of publication : Thursday, November 17, 2022
Type of Contract : FTC Scientist
Contract Period : 24 months
Expected date of employment : 1 January 2023
Proportion of work : Full time
Remuneration : Between 2833,40 and 3257,06 € gross salary depending on experience
Desired level of education : PhD
Experience required : 1 to 4 years


The ANR MultiTrans project aims at developing new decision algorithms for autonomous vehicles.
The work concerns federated learning, the goal of which is to allow vehicles to exchange local information within a reduced geographical perimeter (e.g., vehicles that pass each other). However, this issue remains difficult for deep learning techniques due to the multitude of interactions between very diverse actors on the road (vehicles, pedestrians, etc.). Therefore, it is necessary to help neural networks to model these dependencies. This recruitment will have to answer this need by introducing a scene graph structure in the data processing chain. This graph structure naturally allows the articulation of possible interactions to improve model performance, as shown in a number of recent publications [1, 2]. Indeed, the nature of the information learned and exchanged can be symbolic (modification of a portion of road, of a roundabout) or non-symbolic (improvement of lane guidance in degraded conditions such as fog or snow, presence of walking or running pedestrians).
All this implies being able to combine symbolic learning (represented by the graph structure) with non-symbolic learning (perception-based).


The symbolic data of a visual scene is usually represented as a scene graph while the non-symbolic data is represented as the weights learned in a neural network. Different works already combine such learning.
In this post-doc, we want to:
• implement a deep learning system based on scene graphs,
• improve the definition and extraction of scene graphs,
• propose useful improvements to autonomous driving context:
◦ define the useful concepts that a scene graph node should represent
◦ would it be possible to learn new objects (this new vehicle driving in front of me is a skateboard with handlebars (scooter concept))?
◦ study the dynamics of symbolic data (the car that was in front of me and that I just passed ...)
• Study the possibility of introducing a priori knowledge into the models using this graph structure,
• Evaluate the usefulness of these techniques for the transferability of the learned systems between various domains (simulation, remote control car circuit, autonomous shuttles, real-life use case) involved in the MultiTrans project,
• Publish the findings of the work in conferences and journals in the related field.

[1] Graph-Based Spatial-Temporal Convolutional Network for Vehicle Trajectory Prediction in Autonomous Driving, Sheng et al, IEEE Transactions on Intelligent Transport Systems, 2022.
[2] Scene-Graph Augmented Data-Driven Risk Assessment of Autonomous Vehicle Decisions, Yu et al, IEEE Transactions on Intelligent Transport Systems, 2021.


The candidate should have a PhD related to machine learning techniques for computer vision, and have knowledge in:
- machine learning, and more particularly in deep learning,
- processing of weakly labeled situations,
- image and signal processing,
- emerging vision transformer architectures,
- graph processing.
In addition, the candidate should have the following expertise:
- Training of neural networks on supercomputers,
- Implementation of new deep learning methods,
- Programming in Python, with mastery of the two main learning libraries (pytorch, tensorflow).

Work Context

This post-doc researcher is part of the ANR MultiTrans project which concerns the transfer of learning between different platforms (fully virtualized, remote control cars, or real) for autonomous driving.

Constraints and risks


We talk about it on Twitter!