Aristide Baratin

Aristide Baratin

AI Research Scientist

Samsung - SAIT AI Lab Montreal

About me

I’m a research scientist at Samsung SAIT AI Lab Montreal, an academic-style research lab situated within Mila - Quebec AI Institute. My research revolves around foundational aspects of deep learning algorithms, with a focus on understanding and improving their robustness and generalization performance. I recently earned a PhD in machine learning at Mila, University of Montreal, advised by Simon Lacoste-Julien.

Prior to my time at Mila, I’ve worked as a theorerical physicist, first as a junior scientist at the Max Planck Institute for Gravitational Physics in Potsdam, then as a Humboldt Fellow hosted by the University of Waterloo and McGill. My work was mainly about quantum gravity and associated mathematical structures from higher dimensional algebra and matrix models. I received a PhD in physics from Perimeter Institute in Waterloo and Ecole Normale Supérieure de Lyon, under the supervision of Laurent Freidel.

I also love teaching. I’ve taught numerous classes in mathematics, physics and computer science as a teacher’s assistant at ENS Lyon and UdeM and as a course lecturer at UW and McGill.

See my curriculum vitae for more details.

Interests
  • Machine learning
  • Deep learning theory
  • Representation learning
  • Optimization
  • Generalization, robustness
Education
  • PhD, Machine Learning, 2022

    Mila, Université de Montréal

  • PhD, Theoretical Physics, 2009

    ENS Lyon and Perimeter Institute, Waterloo

  • Master's degrees, Mathematics & Physics, 2002-2004

    University Paris-Saclay and ENS Paris

  • Stipendary student, 2002-2004

    ENS Paris-Saclay, Mathematics Department

Publications

(2023). CrossSplit: Mitigating Label Noise Memorization through Data Splitting. ICML, 2023.

arXiv

(2023). Promoting Exploration in Memory-Augmented Adam using Critical Momenta. arXiv 2023.

arXiv

(2022). Expressiveness and Learnability: A Unifying View for Evaluating Self-Supervised Learning. arxiv, 2022.

PDF

(2021). Implicit Regularization via Neural Feature Alignment. AISTATS 2021.

Code Video arXiv

(2021). On the Regularity of Attention. arXiv, 2021.

arXiv press

(2020). A Mathematical Theory of Attention. arXiv, 2020.

arXiv

(2019). On the Spectral Bias of Neural Networks. ICML 2019.

Code Slides arXiv

Contact