cv
Concise overview of my academic background and research activities.
Basics
| Name | Zacharie Garnier-Cuchet |
| Label | Research Master Student |
| Url | https://zachariegarniercuchet.github.io/ |
| Summary | M.Sc. student in AI at Polytechnique Montréal / MILA |
Work
Education
-
2024.09 - Present Montréal, Canada
Master
Polytechnique Montréal / MILA - Québec AI Institute
Artificial Intelligence
- Natural Language Processing
- Deep Learning Dynamics
- Reinforcement Learning
- Representation Learning
- Stochastic Processes
- Data Science
-
2022.03 - 2022.08 Berlin, Germany
Academic exchange
Technische Universität Berlin (TUB) / Berlin University Alliance
Mathematics, Physics, mechanical engineering
- Linear algebra
- Statistics
- probability
- Intensive German Course
-
2020.09 - 2024.06 Compiègne, France
Engineering degree
Université Technologique de Compiègne (UTC) / Sorbonne University Alliance
Mathematics, Physics, mechanical engineering
- Thermodynamics
- Mechanics of fluids
- Mechanical vibrations
- Structure optimization
Publications
-
2025.06.09 Revisiting the Goldilocks Zone in Inhomogeneous Networks
HiLD at ICML 2025
It concerns how architectural components shape the loss landscape and how initialization geometry relates to trainability.
Languages
| French | |
| Native speaker |
| English | |
| Fluent |
| German | |
| Basic |
| Spanish | |
| Basic |
References
| Professor amal Zouaq | |
| Professor of AI and NLP at Polytechnique Montréal and MILA, specializing in knowledge extraction, semantic representations, and large-scale NLP systems. |
| Professor Philippe Langlais | |
| Professor of Computational Linguistics at Université de Montréal, renowned for his work in machine translation, text processing, and statistical NLP. |
Projects
- 2024.09 - 2024.12
Applying RLHF to Humor
As part of a Reinforcement Learning course, this project applied Reinforcement Learning from Human Feedback (RLHF) to align a language model toward generating humorous and engaging responses. Using Microsoft’s Phi-3-instruct, a BERT-based reward model, and PPO via the TRL library, we demonstrated how RLHF can transform a standard model into a more dynamic conversational assistant.
- RLHF Pipeline
- PPO with TRL
- BERT-based Reward Model
- 2024.03 - 2024.08
Dynamic Analysis of Taipei 101
The project investigates the dynamic behaviour of the Taipei 101 tower through a theoretical and model-driven approach. Using finite element abstractions and Rayleigh–Ritz formulations, it examines natural vibration modes, modal frequencies, and the impact of passive absorbers. The work emphasizes structural modelling, eigenvalue analysis, and the interpretation of complex dynamical responses using computational tools.
- Structural Modelling
- Eigenvalue & Modal Analysis
- Rayleigh–Ritz Method
- Dynamic Systems Interpretation
- Computational Simulation