Guillaume Braun | Riken AIP

About Me

Since April 2024, I have been a postdoctoral researcher in the High-Dimensional Structure Theory Team at RIKEN, under the supervision of Masaaki Imaizumi, working on theoretical machine learning and high-dimensional statistics.

Before that, I was a postdoctoral researcher in the Imperfect Information Learning Team from 2023 to 2024. Prior to that, I completed my Ph.D. at Inria Lille - Nord Europe, as part of the MODAL team, supervised by Christophe Biernacki and Hemant Tyagi. My thesis focused on clustering and matching problems on graphs with efficient first-order methods.

Research Interests

My research lies in high-dimensional statistics and optimization, with applications to the theoretical understanding of modern machine learning algorithms and architectures. My current work studies the dynamics of neural networks trained by gradient-based methods, focusing on feature learning and how data geometry shapes their behavior.

Publications

Spectral Gradient Descent Mitigates Anisotropy-Driven Misalignment: A Case Study in Phase Retrieval
G. Braun, H. Bao, W. Huang, M. Imaizumi, ICML (2026) [Arxiv]
Fast Escape, Slow Convergence: Learning Dynamics of Phase Retrieval under Power-Law Data
G. Braun, B. Loureiro, H.Q. Minh, M. Imaizumi, ICLR (2026), Oral Presentation [Arxiv]
Neuron Block Dynamics for XOR Classification with Zero-Margin
G. Braun and M. Imaizumi, AISTATS (2026)[Arxiv]
Learning a Single Index Model from Anisotropic Data with Vanilla Stochastic Gradient Descent
G. Braun, H.Q. Minh and M. Imaizumi, AISTATS (2025) [Arxiv]
VEC-SBM: Optimal Community Detection with Vectorial Edge Covariates
G. Braun and M. Sugiyama, AISTATS (2024) [PMLR]
Strong Consistency Guarantees for Clustering High-Dimensional Bipartite Graphs with the Spectral Method
G. Braun, Electronic Journal of Statistics (2024) [EJS]
Minimax Optimal Clustering of Bipartite Graphs with a Generalized Power Method
G. Braun and H. Tyagi, Information and Inference: A Journal of the IMA (2023). [ArXiv]
Seeded graph matching for the correlated Wigner model via the projected power method
E. Araya, G. Braun and H. Tyagi, Journal of Machine Learning Research (2024) [JMLR]
An iterative clustering algorithm for the Contextual Stochastic Block Model with optimality guarantees
G. Braun, H. Tyagi, and C. Biernacki, ICML (2022). [ArXiv]
Clustering multilayer graphs with missing nodes
G. Braun, H. Tyagi, and C. Biernacki, AISTATS (2021). [PMLR] [Code][Poster]

Invited Talks

Learning Dynamics of Phase Retrieval under Power-Law Data IMS-APRM, Hong-Kong, June 2026
Spectral Gradient Descent Mitigates Anisotropy-Driven Misalignment: A Case Study in Phase Retrieval
MyJP-SIM² 2026, Tokyo
How Anisotropy Shapes Learning Dynamics and the Role of Optimization Geometry
FIMI 2026, Tokyo.
異方性データを用いたシングルインデックスモデルの学習：バニラSGDの解析
FIT 2025, Sapporo.
Clustering de graphes bipartis en grande dimension : limites statistiques et algorithmes efficaces
Journées MAS 2024, Poitiers.
Clustering Bipartite graphs with the Generalized Power Method
SIAM Workshop on Network Science 2022, lightning talk
Clustering graphs with side information
SPSR Workshop, Bucharest (online), November 19th, 2021

Reviewing

Conferences: AISTATS, ICML, ICLR, NeurIPS.
Journals: Journal of Machine Learning Research, Annals of Statistics.