New notebook! To get to grips with discrete flow matching paper, I've put together a little notebook to test it on MNIST. I kept the article's notations so it's easy to follow alongside the article! 🐍 notebookv1: github.com/gle-bellier/di…

I had a great time collaborating with Valerio Marsocci and many others on the Pangaea project! You want to know if your earth observation encoder learned useful representations for dense tasks? Use the Pangaea benchmark! 🛰️

Maybe we just need to wait for better SSL models to have better FID with diffusion/flow matching models... (alignement model linear probing accuracy vs aligned diffusion/fm model FID) from REPA article arxiv.org/abs/2410.06940)

🚨 New paper alert! [NeurIPS 2024 spotlight] 🚨 Trajectory Flow Matching with Applications to Clinical Time Series Modeling ⏳📈 With: Yuan Pu , Yuki Kawamura 川村祐貴 , Andrew Loza, Yoshua Bengio, Dennis Shung, Alex Tong 💻: github.com/nZhangx/Trajec… 📄: arxiv.org/abs/2410.21154 🧵👇

New paper out! We introduce “Generator Matching” (GM), a method to build GenAI models for any data type (incl. multimodal) with any Markov process. GM unifies a range of state-of-the-art models and enables new designs of generative models. arxiv.org/abs/2410.20587 (1/5)

Does equivariance matter at scale? Should a model rather learn equi- and invariances from data or should the architecture have the equiv property? This work provides some insights, and SCALING LAWS for each. P: arxiv.org/abs/2410.23179

Very nice paper! I like the decomposition between the encoder learning the deterministic components+spatial/modality alignment and the flow matching/diffusion learning the stochastic small-scale physics. (+noise scaling to avoid overfitting on the deterministic dynamics)

What are the good practices to set up the number of epochs when using a cosine learning rate scheduler? Set it to a large number + early stopping to determine the needed epochs and then rerun the experiment?

I've been on twitter for a long time and really liked it and still appreciate the relevance of so many research posts on this platform. So I will not quit but start cross-posting on 🎆🛤️🏙️ where you can find me @ lebellig

I'll present Stem-JEPA this afternoon at ISMIR Conference! Hope to see you there!

I made a first scan to find people from here in the other place 🦋, maybe some of them got through: don't hesitate to ping me! (my handle is simply lebellig) The scrolling experience is so much better there...

Do you work on Deep Learning for remote-sensing or Earth Observation (SSL, generative models, foundation models...) and want to move to the other place 🦋? I've initiated a pack with researchers in the field so we can find each other again! go.bsky.app/EKVwvde

Are geospatial foundation models really impactful? 🚀🚀🌏 Check it in our new pre-print! Welcome to **PANGAEA: a global and inclusive benchmark for GFMs** arxiv.org/abs/2412.04204 a short thread 🧵

A new (and comprehensive) Flow Matching guide and codebase released! Join us tomorrow at 9:30AM NeurIPS Conference for the FM tutorial to hear more... arxiv.org/abs/2412.06264 github.com/facebookresear…

From Diffusion Models to Schrödinger Bridges - A shame not to see this NeurIPS keynote live, by the incredible Arnaud Doucet - SB naturally extends flow/ bridge matching and diffusion models - Particularly useful for data to data - Links to OT neurips.cc/virtual/2024/i…

It was mostly a lie. I'm spending most of my social network time in the other place. I'm impressed with how fast the community moved there. My handle is lebellig (same but without the underscore)