Thomas George

pic.jpg

I am a researcher at Orange Research in the Responsible and Impactful AI and Machine learning foundations research groups.

My research interests include:

  • Generalization and explainability of deep neural networks
  • Weakly supervised learning
  • Causality in machine learning

I did my PhD at Mila in Québec, under the joint supervision of Pascal Vincent and Guillaume Lajoie, followed by a post-doc in Vincent Lemaire’s group at Orange Research. Previously, I was a multiple hats engineer at Eco-Adapt where I worked on time series from industrial sensors, in order to implement automated data analysis algorithms. Prior to that I studied at École des Mines.

Here is my academic CV.

news

Oct 14, 2025 At the Third Conference of the AI Fairness Cluster in Bruxelles, to talk about Fairness in intersectional setups, presented by Jeanne Monnier.
Oct 10, 2025 New preprint Training Feature Attribution for Vision Models available on arXiv for our work on explainability of vision models with Aziz Bacha.
Oct 09, 2025 I will present our ongoing work Statistical diagnostic measures for deep learning with Pierre Nodet at the Workshop on cautiousness and imprecision in machine learning.
Sep 16, 2025 I will attend the Trustworthy AI Summit with Jeanne Monnier to present our work Fairness in intersectional setups - aggregation choice and some paradoxes.
Mar 13, 2025 PhD position at Orange Research with Pierre Nodet and me: Explaining “black box” AI algorithms through their training examples (en, fr).

latest posts

selected publications

  1. EKFAC
    Fast approximate natural gradient descent in a kronecker factored eigenbasis
    Thomas George, César Laurent, Xavier Bouthillier, and 2 more authors
    Advances in Neural Information Processing Systems, 2018
  2. NTKAlign
    Implicit regularization via neural feature alignment
    Aristide Baratin, Thomas George, César Laurent, and 4 more authors
    In International Conference on Artificial Intelligence and Statistics, 2021
  3. LazyHasty
    Lazy vs hasty: linearization in deep networks impacts learning schedule based on example difficulty
    Thomas George, Guillaume Lajoie, and Aristide Baratin
    Transactions on Machine Learning Research, 2022
  4. Mislabeled
    Mislabeled examples detection viewed as probing machine learning models: concepts, survey and extensive benchmark
    Thomas George, Pierre Nodet, Alexis Bondu, and 1 more author
    Transactions on Machine Learning Research, 2024