Hugues Van Assel

Postdoctoral Fellow at Genentech, South San Francisco CA

photo_profil.jpeg

I am a postdoc at Genentech with Aviv Regev and Tommaso Biancalani.

I am interested in how machines learn rich and reliable representations of complex data. My work explores representation learning, self-supervised and multi-modal methods, optimal transport, and dimensionality reduction. I develop computational approaches that uncover structure in data, motivated by challenges in the life sciences.

I enjoy building and sharing open-source tools, including:

  • TorchDR : a modular, GPU-friendly toolbox for dimensionality reduction (DR) that offers a unified interface for state-of-the-art DR methods.
  • stable-pretraining : a PyTorch library for foundation model pretraining with real-time training monitoring.

I did my PhD in the math department of ENS Lyon on Optimal Transport and Probabilistic Modeling for Dimensionality Reduction. Prior to my PhD, I was a student at Ecole polytechnique and MVA.


selected publications

  1. NeurIPS
    Joint Embedding vs Reconstruction: Provable Benefits of Latent Space Prediction for Self-Supervised Learning
    Hugues Van Assel, Mark Ibrahim, Tommaso Biancalani, Aviv Regev, and Randall Balestriero
    Advances in Neural Information Processing Systems (Spotlight), 2025
    arXiv
  2. TMLR
    Distributional Reduction: Unifying Dimensionality Reduction and Clustering with Gromov-Wasserstein
    Hugues Van Assel, Cédric Vincent-Cuaz, Nicolas Courty, Rémi Flamary, Pascal Frossard, and Titouan Vayer
    Transactions on Machine Learning Research, 2024
    arXiv
  3. NeurIPS
    SNEkhorn: Dimension Reduction with Symmetric Entropic Affinities
    Hugues Van Assel, Titouan Vayer, Rémi Flamary, and Nicolas Courty
    Advances in Neural Information Processing Systems, 2023
    arXiv Poster
  4. NeurIPS
    A Probabilistic Graph Coupling View of Dimension Reduction
    Hugues Van Assel, Thibault Espinasse, Julien Chiquet, and Franck Picard
    Advances in Neural Information Processing Systems, 2022
    arXiv Poster