Romain Menegaux

I am a Starting Research Fellow at Inria Grenoble in the Thoth team team, supervised by Julien Mairal.

I earned my PhD in machine learning and bioinformatics from Mines Paristech as part of the CBIO team, under the supervision of Jean-Philippe Vert.
Prior to my PhD, I obtained master's degrees from Ecole Polytechnique and Columbia. I then spent 3 years as a research scientist at Bloomberg in New York before returning to France to pursue academic research.

Email  /  CV  /  Scholar  /  Github

Research

I'm interested in applications of machine learning to healthcare and biology, with a focus on Transformers, graph neural networks and optimization. I also have ongoing projects in flow matching, 3D-scene modelling and geoscience.
Below is a list of my recent publications.

Lifting 2D visual features from pretrained models to 3D Gaussian Splatting scenes with a simple aggregation scheme and no optimization.

A global convergence result for Gauss-Newton in over-parameterized one-hidden layer neural networks, and an empirical study showing hidden learning with small initialization and small step-size.

Encoding graph structure in Transformers for molecule property prediction, and introducing a novel attention mechanism.
State of the art on the ZINC benchmark dataset.

Vocabulary reduction in DNA textual embeddings by mapping DNA "tokens" to their corresponding nodes in a de Bruijn graph.

DNA sequence classification with textual embeddings, paper for the fastDNA software.

Selected Teaching

Teaching Assistant with Jean-Philippe Vert

Website template graciously lifted from source code.