David Heurtel-Depeiges
Guest Researcher (and former intern), Center for Computational Mathematics, Flatiron Institute, NY. Supervised by Bruno Regaldo-Saint Blancard and Ruben Ohana.
Master Student, Ecole Normale Supérieure Paris-Saclay, MVA program, Paris
Looking for a PhD position starting September 2024!
About me & Interests
Hi, I am David Heurtel-Depeiges, a final year Master Student at Ecole Normale Supérieure Paris-Saclay through the MVA program: Mathematics, Vision and Learning (Apprentissage in French). I completed my undergraduate studies at Ecole Polytechnique, where I studied Applied Mathematics and Computer Science. I also did my first year of Master’s there1. I am widely interested in Machine Learning, Deep Learning and their application to many interesting problem from environemental studies to cosmology or other fields. More specifically, I am interested in the two (wide) following topics:
- Deep Generative Models with a focus on helping solve complex inference problems (including inverse problems) for scientific or medical applications. Right now I am looking at the intersection of optimal transport and generative modeling. And Representation Learning as data in science and medicine come in increasingly large quantity and a multimodal fashion. Advances in both these fields seem key to solving many challenges in scientific and medical AI.
- Foundation Models and AI Safety. I am interested in foundation models at scale and the role of multimodality in building better models. I am also interested in AI safety and its derivatives (interpretability, bias/fairness). Ultimately, understanding how foundation models work can help us understand and detect their failure modes (jailbreaking, hallucinations, etc.) and build better systems (truthful, robust ones).
I am currently looking for a PhD/industry position starting september 2024! I am also looking for a 4 to 6 months research internship starting in March/April 2024.
News
10. February 2024 Started a project at MVA to extend results from this paper to non-linear or non-gaussian cases
4. February 2024 Finished a project on CLIP embeddings for molecules seen as graphs and their text descriptions.
2. February 2024 Paper submitted (just in time) to ICML, extending our previous work on colored noise diffusion model to allow for “double-blind” denoising and solving complicated inference problems in cosmology.
On 16. December 2023 Will be giving a brief talk at NeurIPS 2023 ML and the Physical Sciences workshop!
13. Decembre 2023 NeurIPS This is so overwhelming! So many people, so many talks, so many posters. On day 1, I have “saved” exactly 94 papers to read…it must be my enthusiasm. I am looking forward to the rest of the conference and the workshops, but at this pace, I will need a few months to comb through all the papers I want to read (and it would probably mean missing a few classes…).
30. November 2023 Google France Research Day Fascinating day at Google France in Paris with a series of exciting talks on AI safety and privacy. From Fully Homomorphic Encryption in ML to Differential Privacy and AI safety, it was a great opportunity to learn about the latest developments in these fields and hear some great speakers from Google and academia.
Main Research Projects and Contributions
A list of my contributions, publications and code can be found here.
Other projects
Towards Monosemanticity in Vision Models with dictionary learning
Code
PageFootnotes
-
The French higher education system is highly convoluted, trust me when I say what I did is not unusual. If you want more information: preparatory classes, structure of the programm at Polytechnique ↩