On a few occasions, I've been asked by fellow researchers with backgrounds in computational statistics how to get started with learning about statistical physics. I have a necessarily narrow view of the field, but I do have default references which I share. I document them here.

Yes! If you want to know more about Hopfield networks and Boltzmann machines (Hinton's extension of Hopfield), and how they relate to Ising models, PGMs, and EBMs, see sec 4.3 of my book (probml.github.io/pml-book/book2…)

Yet another super nice paper by Hyvarinen, how to implement Langevin when only having access to the score of a noised version of the target arxiv.org/abs/2410.05837

I am very happy to share: " Manifolds, Random Matrices and Spectral Gaps: The geometric phases of generative diffusion" We study the evolving latent geometry of diffusion by analyzing gaps in the Jacobian spectrum with statistical physics methods Paper: arxiv.org/pdf/2410.05898

Coming next week, a two hour war of attrition with Tamler Sommers where we argue about the problem of induction, justification, prediction, and belief. Tamler looking suspicious, Vaden ranting, and Ben looking away in horror sums things up pretty well

To me this is where GFlowNets are most powerful: arbitrary constraints defined by real chemistry & you still get normalising constants / nice probabilistic interpretations / etc… and it seems to work well!

I often find standard causal attention masks boring or not sensible in many settings. Therefore I wrote a small package to generate attention masks with **block structure** using a DAG-like language. github.com/xidulu/LE-ATTE…

Pleased to announce that two new papers are up on arXiv! 1. arxiv.org/abs/2410.18929 2. arxiv.org/abs/2410.03630 (uploaded a few weeks ago, but thought I'd share it with the new paper above)

In addition to the ELLIS PhD (closing this week), I'm looking for two more PhDs funded through UKRI's AI CDT program: Multimodal representation learning w.@MagnusRattray findaphd.com/phds/project/u… Casual abstraction of dynamical systems w. Mauricio Alvarez findaphd.com/phds/project/c…

UBC Statistics Profs. Alexandre Bouchard-Côté & Trevor Campbell, along with collaborators from UdeM & SFU, are leading a CANSSI CRT project on distributed MCMC methods. Funded grad student positions available in Bayesian stats & computational methods: bit.ly/3YOaztn

Exciting opportunity to intern with me 🙋‍♀️ & the entire research unit lead by Jason Hartford 🧠 on single-cell representation challenges 🦠 in gene perturbation image data 🧬 Come join us in our @Valence_AI London office 🇬🇧 Don't self-reject & apply at: job-boards.greenhouse.io/valencelabs/jo… ✍️

There’s likely no other place that offers such immense intellectual freedom and computational resources. I had the flexibility to explore multiple projects, collaborate with an incredible team, and experiment random ideas with 50 H100s 🔥! Don't hesitate to apply!

(OT/Gradient Flow/DRO, slides) A great time visiting EPFL Math EPFL EPFL CDM & one of my fav cities Lausanne CH. Stimulating exchanges with scholars such as Xue-Mei Li, Daniel Kuhn, Victor Panaretos, Lénaïc Chizat... Talk: jj-zhu.github.io/file/epfl-nov-… Some insights 👇

Come to talk to us if you are interested in finding the optimal joint embeddings of multiple modalities!

Definitely a career highlight to be awarded a best paper award 🤩🤩 But even more exciting is that the 2nd and 3rd awards are given to colleagues Kian and Mohit, a podium sweep from Valence Labs / Recursion 🥇🥈🥉

A year ago, I would think this has to be an outlier. But now I got NeurIPS reviewer asking about "what do you mean by almost surely?"