Antonio Sclocchi

Research Fellow, Gatsby Unit, UCL Theory of Learning Lab

prof_pic.jpg

Sainsbury Wellcome Centre, UCL

25 Howland St

London, UK

I am a theoretical physicist working at the intersection between statistical physics and deep learning. My current research interest revolves around neural network optimization, diffusion models, and data structure.

Currently, I work in the group of Andrew Saxe at UCL. Previously, I worked in the group of Matthieu Wyart at EPFL, building models of data structure and learning dynamics inspired by statistical physics. I earned a PhD in theoretical physics at LPTMS, Université Paris-Saclay under the supervision of Silvio Franz and Pierfrancesco Urbani, working on disordered and glassy systems.

recent publications

  1. unet_rhm.png
    How compositional generalization and creativity improve as diffusion models are trained
    Alessandro Favero , Antonio Sclocchi, Francesco Cagnetta , and 2 more authors
    2025
  2. diffusion_tree.png
    Probing the Latent Hierarchical Structure of Data via Diffusion Models
    Antonio Sclocchi, Alessandro Favero , Noam Itzhak Levi , and 1 more author
    In The Thirteenth International Conference on Learning Representations , 2025
  3. leo_to_butterfly.jpg
    A phase transition in diffusion models reveals the hierarchical nature of data
    Antonio Sclocchi, Alessandro Favero , and Matthieu Wyart
    Proceedings of the National Academy of Sciences, 2025
  4. sgd_phase_diagram.png
    On the different regimes of stochastic gradient descent
    Antonio Sclocchi, and Matthieu Wyart
    Proceedings of the National Academy of Sciences, 2024

recent news

Feb 17, 2025 Our work on compositional generalization and creativity just appeared on the arXiv
Jan 22, 2025 Our work on probing the data structure with diffusion models has been accepted at ICLR 2025 :tada:
Jan 02, 2025 Our work on diffusion models on hierarchical data has been finally published on PNAS :tada:
Feb 20, 2024 Our work on Stochastic Gradient Descent has been finally published on PNAS.