PeopleofSuperintelligence

Researcher

Raphaël Berthier

Raphaël Berthier is a researcher in optimization and the theory of neural networks. His PhD work at Inria Paris with Francis Bach and Pierre Gaillard produced a 2021 thesis, "Analysis and acceleration of gradient descents and gossip algorithms", and he coauthored the NeurIPS 2021 Outstanding Paper "Continuized Accelerations of Deterministic and Stochastic Gradient Descents, and of Gossip Algorithms". In 2022, he was a postdoc at EPFL with Emmanuel Abbe and Andrea Montanari, and he went on to publish papers on the learning dynamics of diagonal linear networks, two-layer networks and attention layers. He is a tenure-track researcher at Inria Sorbonne Université and a member of the Inria project-team Megavolt.

Now at
Inria
Badges

Timeline

  1. 2017

    State evolution for message passing

    In August 2017, he posts a paper with Andrea Montanari and Phan-Minh Nguyen that generalizes state evolution for approximate message passing algorithms to Lipschitz continuous non-separable nonlinearities with Gaussian matrices.

  2. 2020

    Accelerated gossip with Jacobi polynomials

    In 2020, he publishes "Accelerated gossip in networks of given dimension using Jacobi polynomial iterations" with Francis Bach and Pierre Gaillard in the SIAM Journal on Mathematics of Data Science. The method solves the gossip problem, the distributed estimation of an average over a network, and depends only on the spectral dimension of the network.

  3. By 2021

    PhD thesis at Inria Paris

    His PhD at Inria Paris, supervised by Francis Bach and Pierre Gaillard, yields the 2021 thesis "Analysis and acceleration of gradient descents and gossip algorithms". The thesis is a parallel study of convex optimization and gossip algorithms.

  4. 2021

    A NeurIPS Outstanding Paper award

    In 2021, the paper "Continuized Accelerations of Deterministic and Stochastic Gradient Descents, and of Gossip Algorithms", which he writes with Mathieu Even, Francis Bach, Nicolas Flammarion, Pierre Gaillard, Hadrien Hendrikx, Laurent Massoulié and Adrien Taylor, receives a NeurIPS Outstanding Paper award. It gives the first rigorous acceleration of asynchronous gossip algorithms.

  5. 2022

    Postdoc at EPFL

    In 2022, he is a postdoc at EPFL, working with Emmanuel Abbe and Andrea Montanari. He joins the Simons Institute's Deep Learning Theory summer cluster as a visiting postdoc in summer 2022.

  6. 2023–2024

    Diagonal networks and two-layer dynamics

    In 2023, he publishes "Incremental learning in diagonal linear networks" in the Journal of Machine Learning Research; it shows that gradient flow with small initialization activates coordinates one after another. In 2024, "Learning time-scales in two-layers neural networks", written with Andrea Montanari and Kangjie Zhou, appears in Foundations of Computational Mathematics and studies the gradient flow of a wide two-layer network in high dimension.

  7. 2024–2025

    Attention and single-location regression

    In 2024, he coauthors "Attention layers provably solve single-location regression" with Pierre Marion, Gérard Biau and Claire Boyer; the paper is accepted to ICLR 2025. It proposes a predictor that is a simplified version of a non-linear self-attention layer and shows its asymptotic Bayes optimality.

  8. As of October 2026

    Tenure-track at Inria Sorbonne Université

    As of October 2026, he is a tenure-track researcher (chaire de professeur junior) at Inria Sorbonne Université and a member of the Inria project-team Megavolt. His 2026 preprints study gradient descent at the edge of stability and incremental learning in mirror flows.

Career

  1. NowInriaTenure-track researcher (chaire de professeur junior)
  2. From 2022EPFLPostdoc
  3. 2022Simons Institute for the Theory of ComputingVisiting postdoc

Education

  1. 2021InriaPhD

From the biography.

  1. People shaping AI work at a median of 3 organizations over a career: 3 for university starters and institute and lab starters, versus 2 for company startersIncludes Inria. Across 6,853 people shaping AI with a dated work role, we estimate a median of 3 parent organizations over a career (mean 3.6), comparing role groups, starting sectors and starting decades from the 1950s to 2020–2026.
  2. The company share of AI researchers’ top papers was 31% in 2020–2024 and 8% in 1970–1974Includes Inria. We estimate where the people shaping AI worked when their most-cited papers appeared: the company share was 31% in 2020–2024 and 8% in 1970–1974, and Google led all organizations in 2020–2024.
  3. Google has the highest combined h-index among AI organizationsIncludes Inria. We estimate that Google holds the highest combined h-index at 8,016, while universities account for 73% of the deduplicated total across AI organizations.
  4. Companies supply the most prior-work hires at 7 of the 9 frontier AI labsIncludes Inria. Our estimates of the organizations where 550 people worked directly before joining 9 frontier AI labs across 620 career transitions, comparing the main feeder organizations per lab.

Colleagues

People named in this biography, and people with dated positions at the same organization in the same years.

Sources and credits

Photo: We have no published photo of Raphaël Berthier yet. Send us a free-licensed photo.

  1. State Evolution for Approximate Message Passing with Non-Separable Functions arxiv.org. Checked 22 October 2026.
  2. Accelerated Gossip in Networks of Given Dimension using Jacobi Polynomial Iterations arxiv.org. Checked 22 October 2026.
  3. Raphaël Berthier raphael-berthier.github.io. Checked 22 October 2026.
  4. Raphaël Berthier | ISL Colloquium isl.stanford.edu. Checked 22 October 2026.
  5. Announcing the NeurIPS 2021 Award Recipients blog.neurips.cc. Checked 22 October 2026.
  6. Continuized Accelerations of Deterministic and Stochastic Gradient Descents, and of Gossip Algorithms papers.nips.cc. Checked 22 October 2026.
  7. Raphaël Berthier simons.berkeley.edu. Checked 22 October 2026.
  8. Incremental Learning in Diagonal Linear Networks jmlr.org. Checked 22 October 2026.
  9. Learning time-scales in two-layers neural networks arxiv.org. Checked 22 October 2026.
  10. Attention layers provably solve single-location regression arxiv.org. Checked 22 October 2026.