Florentin Guth
Florentin Guth studies how deep neural networks learn and generalize, with work on image generation and learned network weights. He defended a doctoral thesis on deep convolutional networks at École Normale Supérieure in August 2023 and served as a Faculty Fellow at NYU's Center for Data Science from 2023 to 2026. His papers examined phase collapse, wavelet-based score generation, deep-network weight distributions, and diffusion-model generalization; the diffusion paper received an ICLR 2024 Outstanding Paper Award. In 2025, he coauthored a method for learning normalized image densities. As of October 2026, he is a Research Fellow at Flatiron Institute and an incoming Assistant Professor at MIT EECS, and co-leads the Sci4DL workshop series.
- Now at
- Simons Foundation
Timeline
-
2022
Phase collapse in neural networks
In 2022, he coauthors “Phase Collapse in Neural Networks” with John Zarka and Stéphane Mallat. The paper argues that phase collapse explains how deep convolutional networks improve image classification.
-
2022
Wavelet score-based generation
In 2022, he coauthors “Wavelet Score-Based Generative Modeling” with Simon Coste, Valentin De Bortoli, and Stéphane Mallat. Its model generates wavelet coefficients with equal time steps across scales, so its time complexity grows linearly with image size.
-
August 2023
A thesis at École Normale
In August 2023, he defends his doctoral thesis at École Normale Supérieure in Paris on the mathematical understanding of deep convolutional neural networks. The work examines image generation, classification, and learned network weights.
-
2023–2026
Faculty Fellow at NYU
From 2023 to 2026, he serves as a Faculty Fellow at NYU's Center for Data Science.
-
2024
A rainbow model of network weights
In 2024, he coauthors the JMLR paper “A Rainbow in Deep Network Black Boxes.” The rainbow model samples deep networks that reach performance similar to trained networks.
-
2024
Diffusion models generalize
With Zahra Kadkhodaie, Eero P. Simoncelli, and Stéphane Mallat, he studies when image diffusion models generalize beyond training samples. Their paper finds that networks trained on non-overlapping image subsets generate identical images from the same noise; it wins an ICLR 2024 Outstanding Paper Award.
-
2025
Dual score matching
In 2025, he coauthors “Learning normalized image densities via dual score matching” with Zahra Kadkhodaie and Eero P. Simoncelli. The method uses two score-matching objectives to fit normalized energy models from data.
-
As of October 2026
Research Fellow at Flatiron
In October 2026, he is a Research Fellow in Flatiron Institute's Center for Computational Neuroscience and an incoming Assistant Professor at MIT EECS. He is a co-lead organizer of the Sci4DL workshop series on scientific methods for understanding deep learning; his research asks how architecture and optimization encode assumptions that help neural networks generalize.
Career
- NowSimons FoundationResearch Fellow
- 2023–2026New York UniversityFaculty Fellow
- Massachusetts Institute of TechnologyAssistant Professor
Education
- 2023École Normale SupérieureDoctorate
Colleagues
Sources and credits
Photo: We have no published photo of Florentin Guth yet. Send us a free-licensed photo.
- Phase Collapse in Neural Networks
- Wavelet Score-Based Generative Modeling
- DIENS seminars
- Florentin Guth
- A Rainbow in Deep Network Black Boxes
- Generalization in diffusion models arises from geometry-adaptive harmonic representations
- ICLR 2024 Outstanding Paper Awards
- Learning normalized image densities via dual score matching
- Florentin Guth - Oct - 2026 - Redwood Center for Theoretical Neuroscience