PhD Position: Toward Frugal Machine Learning with Physics-Aware Models
8 Novembre 2023
Catégorie : Doctorant
This PhD thesis aims to develop an action recognition algorithm with a minimal ecological footprint. It leverages physics-based machine learning models to simulate complex systems, focuses on geometric and physical invariance, and implements incremental learning for long-term sustainability.
Advisors: Olivier Alata (Professor) and Jordan Frecon-Deloire (Associate Professor)
Host laboratory: Hubert Curien Lab UMR CNRS 5516, Saint-Étienne, France
Starting date: Early 2024 - at your earliest convenience
Keywords: Physics-based machine learning; Frugal AI; Incremental learning; Action recognition
Context: The present thesis proposal is part of the GreenAI research project resulting from the collaboration between academic (Hubert Curien Lab) and industrial partners (Asygn, DRACULA Technologies). Its main objective is to create autonomous smart sensors benefiting from a low ecological footprint for road monitoring purposes. As such, this project is inherently a vector of cross-cutting developments in physics, machine learning and embedded systems design.
Description: This subject focuses on the development of a learning algorithm for action recognition  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.
- Physics-aware models. Recently, physics-guided machine learning models  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 . In this thesis, the candidate should contribute to the development of physic-based models in order to both learn and simulate the dynamics of complex systems. Methodological contributions will be applied to road traffic monitoring in order 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 physic-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.) . 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 .
- 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.
- 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.
Host Laboratory Created in 2006, the Hubert Curien laboratory is a joint research unit (UMR 5516) of the Jean Monnet University, Saint-Étienne, the National Research Centre "CNRS," and the Institut d’Optique Graduate School. It is composed of about 90 researchers, professors, and assistant professors, 20 engineers and administrative staff, and 130 PhD and post-PhD students. This makes the Hubert Curien laboratory with a total of about 240 staff the most important research structure of Saint-Étienne. More information at https://laboratoirehubertcurien.univ-st-etienne.fr.
 J. Park, M. Kang, and B. Han, "Class-Incremental Learning for Action Recognition in Videos," ICCV 2021.
 M. Raissi, P. Perdikaris, and G. Karniadakis, "Physics-Informed Neural Networks: A Deep Learning Framework for Solving Forward and Inverse Problems Involving Nonlinear PDE," Journal of Computational Physics 2019.
 M. Buisson-Fenet, V. Morgenthaler, S. Trimpe, and F. Di Meglio, "Recognition Models to Learn Dynamics from Partial Observations with Neural ODEs," TMLR 2023.
 J. Bruna and S. Mallat, "Invariant scattering convolution networks," IEEE TPAMI 2013.
 A. Dekhovich, M. Sluiter, D. Tax, and M. Bessa, "iPINNs: Incremental learning for Physics-informed neural networks," preprint ArXiv 2023.