Congrats to John Hopfield and Geoffrey Hinton! well deserved recognition that some important foundations of AI rest on physics! Physics departments take note: understanding and improving AI systems is a new frontier topic for physics, just as biophysics was earlier. Time to hire!

Probing the Latent Hierarchical Structure of Data via Diffusion Models ift.tt/e5I0Kr7

I’m NeurIPS Conference until Sunday! Join my talk (Sun, 1:20 PM, at Scientific Methods for Understanding Deep Learning) on our work on data structure and diffusion models (neurips.cc/virtual/2024/1…), with Alessandro Favero, N. Levi, Physics of Complex Systems Lab. Happy to chat/connect with anyone interested in the science of deep learning! 🤖

I would like to see a talk that engages with a notion of prescriptive mathematical theory closer to something like the spirit of communication theory in its orientation, rather than "theory" as ML theory and CS theory use the term, which I think is quite narrow-minded.

Check it out & see you in Vancouver (again)! arxiv.org/abs/2502.12089 Huge thanks to my amazing coauthors Antonio Sclocchi (*) Francesco Cagnetta Pascal Frossard and Matthieu Wyart from EPFL Physics of Complex Systems Lab LTS4 Johns Hopkins University SISSA Gatsby Computational Neuroscience Unit

Yep. Physics is what physicists do! And a ton of them do AI well!

This is how our workshop on creativity & generative AI got underway this morning ! Organised with our colleagues Elisa Gremmo and Gabriele Tomasi from the philosophy department Università di Padova, we will discuss generative AI models and whether they are, in some sense, "creative".

Giorgio Parisi opening StatPhys29: AI resembles the heat engine in that the technology arrived before the theory.

Encouraging to see calls for rigour in the science of deep learning—robust phenomena, predictive theories, and all that. But if we need a "position paper" to lay this out, maybe it’s a sign the bar in the field has gotten worryingly low. Isn’t this just what doing science means?

And so it begins ICML Conference ... Tuesday 11 am (east hall) -> We dive into the training process of diffusion models to understand how generalization and combinatorial creativity emerge, with Alessandro Favero. Wednesday 11 am (east hall) -> What Really Drives Neural Scaling Laws?

Starting now, east hall 3508 ICML Conference

🔥 Excited to announce the Workshop on the Principles of Generative Models at #EurIPS2025 (the European #NeurIPS2025) 🇩🇰 Dec 6–7, Copenhagen! 📄CfS due Oct 17 🌐sites.google.com/view/prigm-eur… 👥Pierre Marion Bruno Loureiro Valentin De Bortoli Antonio Orvieto Elisabetta Cornacchia Francesco Cagnetta