معرفی
Pierre Marion is a Researcher at INRIA Paris, working within the Sierra research team since September 2025. His work focuses on the theoretical foundations of deep learning and he is beginning to explore applications of AI in mathematics. Marion has established collaborations across multiple institutions including EPFL, Sorbonne Université, and Google DeepMind.
His educational background includes:
- Engineering degree from École polytechnique (2015-2018) with specialization in Applied Mathematics
- Master's degree from Sorbonne Université (2019-2020)
- PhD from Sorbonne Université (2020-2023) under the supervision of Gérard Biau and Jean-Philippe Vert
- Postdoctoral research at EPFL (2024) supervised by Lénaïc Chizat
Marion's research interests primarily focus on the theory of deep learning, where he investigates the optimization and statistical properties of various neural network architectures. His work spans from shallow networks to complex generative models, with a particular emphasis on understanding the mathematical foundations that govern deep learning performance. Recently, he has begun exploring applications of AI in mathematical research, aiming to bridge the gap between theoretical machine learning and mathematical discovery. His research often combines rigorous theoretical analysis with practical implications for training deep neural networks.
Analysis of Marion's recent publications reveals several key trends in his research. He has made significant contributions to understanding the role of large learning rates in optimization dynamics, demonstrating how they can accelerate convergence in logistic regression and prevent memorization in score-based generative models. His work on attention mechanisms has provided theoretical guarantees for their effectiveness in specific tasks like single-location regression and clustering. Additionally, Marion has extensively studied the connections between residual networks and neural ordinary differential equations, establishing generalization bounds and exploring scaling properties in the large-depth regime. His earlier work included contributions to natural language processing and quasi-Monte Carlo methods, reflecting a broad mathematical foundation that informs his current deep learning research.
Marion has received several notable scientific awards:
- Runner-up PhD Award of AFIA (French Association for Artificial Intelligence) in 2024
- Google PhD Fellowship in 2022
- Ecole polytechnique Grand Prize of Research Internships in 2018
As an advisor, Marion currently supervises PhD student Yu-Han Wu (since 2024), with whom he has co-authored multiple publications on large learning rates and denoising score matching. Previously, he co-supervised several Master's students including Seorim Park, Yerkin Yesbay, and Nathan Doumèche. Marion has been actively involved in the machine learning community through conference organization (NeurIPS@Paris meetups), session chairing (ICSDS 2022), and extensive reviewing activities. He has served as a reviewer for top journals including JASA and Annals of Statistics, and conferences including NeurIPS and ICLR, where he was recognized as a top reviewer at NeurIPS 2023.
Marion is a member of the Sierra research team at INRIA Paris, which focuses on machine learning theory and applications. He has also collaborated with researchers at CREST (Center for Research in Economics and Statistics), as evidenced by his participation in seminars organized by Anna Korba and Karim Lounici. His work often bridges theoretical computer science, statistics, and applied mathematics, reflecting the interdisciplinary nature of modern machine learning research.
