Vincent Dutordoir

Happy to share a major milestone: after years of development, we are officially launching Version 1.0 of the GeometricKernels library! To top it off, our accompanying paper has just been published in JMLR (MLOSS)! 🎉 github.com/geometric-kern…

Introducing AlphaGenome: an AI model to help scientists better understand our DNA – the instruction manual for life 🧬 Researchers can now quickly predict what impact genetic changes could have - helping to generate new hypotheses and drive biological discoveries. ↓

An opportunity to join our team. Feel free to reach out if you have any questions.

@edward_milsom Nice poster and work!

Come see my pretty poster at Neurips on Thursday 12 Dec 4:30pm - 7:30pm! (I can also tell you about our research I guess...) neurips.cc/virtual/2024/p… arxiv.org/abs/2410.06171

I’m really excited to be attending NeurIPS and presenting our work on efficient fine-tuning of pre-trained diffusion models for SOTA conditional generation. Come chat with us on 12th Dec (Thursday) at 11am! Thread below (🧵) -

@glouppe I also didn't realise how important this was until I recently started at DeepMind, where large parts of my time are spent optimising throughput.

Underrated knowledge in ML: the io stack is a performance killer. Students around me seem to have lost the knowledge of what is going on in the background. Few are aware of the hierarchy of memory and how it affects performance, even fewer understand how to optimize it. 🤨

@realGeorgeHotz Transformers have the ability to play chess without explicit search, reaching an ELO of ~2800 in blitz, as shown in arxiv.org/abs/2402.04494

@markvanderwilk @PTenigma Hey Mark! We've developed a python package for spherical harmonics in d-dimensions that might be of interest to you. github.com/vdutor/Spheric… Computing the fourier coefs in high dimensions is typically hard but for zonal functions it reduces to a 1D integral using Funk-Hecke thm.

@PTenigma Looks super interesting! Most packages that I have seen so far focus on harmonics on spheres in 3D space. Are there any good software options for using the same mathematical framework for higher-dimensional spheres?

For brain mappers, climate change modelers, and #AI enthusiasts working with spherical harmonics (which will help you implement AI or graph-CNNs on closed surfaces), a cool new toolbox (torch harmonics [1]) just came out. Having spent 4-5 years writing these functions to align cortical data using covariant PDEs, it's impressive that there's a tool box that will do this now. You can build vector fields, flows on surfaces, or really flows on any manifold homotopic to a sphere. For those if you doing latent diffusion or optimal transport in GenAI, or transforms like "waveStrap" [3] on cortexes, this will probably speed it up. [1] github.com/NVIDIA/torch-h… [2] pages.stat.wisc.edu/~mchung/teachi… [3] x.com/drbreaky/statu…

@louistiao @Meta Congrats Louis! Very happy for you!

🎉 I'm happy to share that I'm joining @Meta as a Research Scientist in New York City! 🇺🇸🗽

@VRLalchand Congrats Vidhi!

Throwback to the day I felt unstoppable and very Cambridge-esque 😇. Hard to describe how challenging and simultaneously growth inducing these last few years have been. But ultimately, it's been an utter privilege. So long, Cambridge. #PhD

@vabor112 @InfAtEd @EdinburghUni Congrats Slava! 👏

Excited to announce that I'll be joining @InfAtEd @EdinburghUni in early spring 2025 as a tenured lecturer (assistant professor) in machine learning! I will be hiring PhD student(s) to work on geometric learning and uncertainty quantification. Stay tuned!

A highlight of the #NeurIPS2023 meetup in Cambridge for me was Carl Rasmussen's keynote on "Halting Climate Change". Spoiler: it has nothing to do with AI. Worth a watch: youtube.com/watch?v=naFaQs…

Links: code: github.com/cambridge-mlg/… slides: drive.google.com/file/d/1k232KY… paper: arxiv.org/abs/2307.05431

Introducing a novel Langevin-based conditional sampler, our model scales to fit complex scalar/vector fields in both Euclidean & manifold codomains, exemplified by forecasting tropical hurricanes as per the illustrated figure.

Happy to share Geometric Neural Diffusion Processes at #NeurIPS. Join us for Poster Session 3 on Wednesday if you’re interested in diffusion models for modelling equivariant stochastic processes. With @MathieuEmile @MHutchinson141 @ValentinDeBort1 @yeewhye #richturner

as well as a keynote presentation by prof Carl Rasmussen on his latest research on climate change policy. Finally, prof @lawrennd will chair a panel discussion on "AI for Social Good".

Apart from posters, the day will feature two spotlight talk sessions...

There are still a couple of places left for the NeurIPS@Cambridge meetup next week, Dec 8th. Priority will be given to people presenting a poster. Registration & full schedule on neuripsmeetupcambridge.info