Check out the amazing work by my collaborators Lukas and René, who trained a ViT-based masked autoencoder to learn patterns in mel-spectrograms of bird vocalizations without annotations, and then cleverly aggregated the learned features to solve downstream bird species classification tasks. 🐦

[1/10] Introducing CroVCA ✨ A simple, unified framework for supervised and unsupervised hashing that converts foundation model embeddings into compact binary codes. ✅ Preserves semantic structure ✅ Trains in just a few iterations

BioDCASE workshop - registration closes next week Oct 10th https://biodcase.github.io/workshop2025/ - Hope to see you there! #bioacoustics

A website to visually browse and explore the ImageNet-1k dataset (there are other supported datasets: IN-12M, WikiMedia, ETH Images, Pixabay, Fashion) navigu.net#imagenet (Maybe this is already known, but I was happy to discover it this morning)

Learning Deep Representations of Data Distributions Sam Buchanan · Druv Pai · Peng Wang · Yi Ma ma-lab-berkeley.github.io/deep-represe... The best Deep Learning book is out, I've been waiting for its release for more than a year. Let's learn how to build intelligent systems via compression.

1/ Can open-data models beat DINOv2? Today we release Franca, a fully open-sourced vision foundation model. Franca with ViT-G backbone matches (and often beats) proprietary models like SigLIPv2, CLIP, DINOv2 on various benchmarks setting a new standard for open-source research.

🏹 Job alert: Research Scientist at Prior Labs 📍Freiburg or Berlin 🇩🇪 📅 Apply by Dec 31 - preferably earlier 🔗 More info: https://bit.ly/4kqn5rY

Our computer vision textbook is now available for free online here: visionbook.mit.edu We are working on adding some interactive components like search and (beta) integration with LLMs. Hope this is useful and feel free to submit Github issues to help us improve the text!

⚠️❗Open PhD and Postdoc positions in Prague with Lukas Neumann! ❗⚠️ We rank #5 in computer vision in Europe and Lukas is a great supervisor, so this is a great opportunity! If you are interested, contact him, he will also be at CVPR with his group :)

We will be presenting the 🍄 FungiTastic 🍄, a multimodal, highly challenging dataset and benchmark covering many ML problems at @fgvcworkshop.bsky.social CVPR-W on Wednesday! ⏱️ 16:15 📍104 E, Level 1 📸 www.kaggle.com/datasets/pic... 📃 arxiv.org/abs/2408.13632 @cvprconference.bsky.social

Want stronger Vision Transformers? Use octic-equivariant layers (arxiv.org/abs/2505.15441). TLDR; We extend @bokmangeorg.bsky.social's reflection-equivariant ViTs to the (octic) group of 90-degree rotations and reflections and... it just works... (DINOv2+DeiT) Code: github.com/davnords/octic-vits

Interesting paper on getting representations from unconditional diffusion models: arxiv.org/abs/2506.01912 (Was excepting more theoretical insights, but nice nonetheless)

Want to use SOTA Self Supervised Learning (SSL) methods on noisy data? We provide a novel training curriculum that significantly improves test performance on clean and noisy samples! The approach is fully SSL and works on any method (DINOv2, MoCo, ...) arxiv.org/abs/2505.12191

Convolution theorem: Fourier transform of the convolution of two functions (under suitable assumptions) is the product of the Fourier transforms of these two functions. buff.ly/aOamDMF

J'ai le plaisir d'annoncer la soutenance de mon habilitation à diriger des recherches de l'Université Gustave Eiffel, intitulée « Apprentissage de représentations à partir d'observations » le mardi 20 mai 2025 à 14h00. 😀 ℹ️ Informations pratiques : nicolas.audebert.at/files/HDR

Perception Encoder: The best visual embeddings are not at the output of the network Daniel Bolya et 17 al. tl;dr: really in title + they propose a new model. arxiv.org/abs/2504.13181

2 jobs in our group at Tilburg University: "Postdoctoral researcher in AI: Learning from Sparse Examples" #nlp postdoc https://tiu.nu/22749 AND #computervision postdoc https://tiu.nu/22748 #postdoc #academicjobs #netherlands #TilburgU

🚀🚀 Reconstruct-Anything-Model: A single model to rule all imaging tasks! We challenge current beliefs and show that a single U-Net can obtain impressive performance across a wide variety of tasks, **without** relying on expensive iterative schemes such as unrolling, PnP, diffusion

🚢🚢 deepinv v0.3.0 is here, with many new features! 🚢 🚢 Our passionate team of contributors keeps shipping more exciting tools! Deepinverse (deepinv.github.io) is a library for solving imaging inverse problems with deep learning.

Applications are 📣OPEN📣 for #PAISS2025 THE AI summer school in #Grenoble 1-5 Sept! Speakers so far @yann-lecun.bsky.social @dimadamen.bsky.social @arthurgretton.bsky.social @gabrielpeyre.bsky.social @science4all.org A. Cristia J. Revaud M. Caron J. Carpentier M. Vladimirova ➡️ paiss.inria.fr

Just launched: the BioDCASE challenge! Monitor whales, birds and other animals, through their sounds? We're publishing new datasets, and new evaluations, to help you do so! Please join this new challenge, and please spread the word: biodcase.github.io #bioacoustics #machinelistening #ai4good

Just to share a bit of academic content… have you heard of VoRA? arxiv.org/abs/2503.20680 🙃

Come and work with us! Be a Postdoctoral Fellow in AI for Ultrasonic Bioacoustic Monitoring! All the details: https://www.naturalis.nl/en/about-us/job-opportunities/postdoctoral-fellow-in-ai-for-ultrasonic-bioacoustic-monitoring #academicjobs #postdoc #pdra #netherlands #biodiversity #ai […]

In case it's not clear! You can register even if you don't have a paper. The event is designed for you 🤗

Tomorrow in our TUM AI - Lecture Series we'll have none other than Saining Xie (@saining.bsky.social), NYU He'll talk about "𝐓𝐡𝐞 𝐦𝐮𝐥𝐭𝐢𝐦𝐨𝐝𝐚𝐥 𝐟𝐮𝐭𝐮𝐫𝐞: Why visual representation still matters". Live stream: www.youtube.com/live/hnu-mRL... 5pm GMT+1 / 9am PST (Mon Mar 17th)

We present Thera🔥: The new SOTA arbitrary-scale super-resolution method with built-in anti-aliasing. Our approach introduces Neural Heat Fields, which guarantee exact Gaussian filtering at any scale, enabling continuous image reconstruction without extra computational cost.

Long-audio understanding: Audio Flamingo 2 (AF2), using a custom CLAP model, synthetic data, and multi-stage curriculum learning, achieved state-of-the-art performance on over 20 benchmarks, including a new long-audio dataset (LongAudio).

Interesting talk by Yi Ma about the nature of intelligence, what have we done so far in AI, and what to do next: scds1001.dirk.hk/L-2.html

🎉 Celebrating 100,000 Modeled Taxa in the iNaturalist Open Range Map Dataset! To mark this milestone, we're making model-generated distribution data even more accessible. Explore, analyze, and use this data to power biodiversity research! 🌍🔍 www.inaturalist.org/posts/106918

Kernel Audio Distance (KAD), a new audio generation evaluation metric, was proposed, showing faster convergence, lower computational cost, and better alignment with human perception than Fréchet Audio Distance (FAD). It leverages MMD and advanced embeddings. GPU acceleration was used.

Are you still using LoRA to fine-tune your LLM? 2024 has seen an explosion of new parameter-efficient fine tuning technique (PEFT), thanks to clever uses of the singular value decomposition (SVD). Let's dive into the alphabet soup: SVF, SVFT, MiLoRA, PiSSA, LoRA-XS 🤯...

Launching: BioDCASE - the Bioacoustics Data Challenge! https://biodcase.github.io/ #DCASE #dcase2025 #dclde #bioacoustics #ai4good

Postdoc job at Naturalis: "Postdoctoral Fellow in Machine Learning & Butterfly Ecology" https://www.naturalis.nl/en/about-us/job-opportunities/postdoctoral-fellow-in-machine-learning-butterfly-ecology (Beautiful museum, great work environment, plus The Netherlands :) #academicjobs #postdoc

Yi Ma & colleagues managed to simplify DINO & DINOv2 by removing many ingredients and adding a robust regularization term from information theory (coding rate) that learn informative decorrelated features. Happy to see principled approaches advance deep representation learning!

Want strong SSL, but not the complexity of DINOv2? CAPI: Cluster and Predict Latents Patches for Improved Masked Image Modeling.

🚀 "UNSURE: self-supervised learning with Unknown Noise level and Stein's Unbiased Risk Estimate" is accepted at #ICLR2025 A thread! 📜 Paper: arxiv.org/abs/2409.01985 🖥️ Code: github.com/tachella/uns...

"Compositional Entailment Learning for Hyperbolic Vision-Language Models" #ICLR25 With: Avik Pal, Max van Spengler, Guido Maria D'Amely di Melendugno, Alessandro Flaborea, and @pascalmettes.bsky.social Paper: arxiv.org/abs/2410.06912

Short introduction to optimal transport with a simple 2D discrete example. The video was done in Manim a few years ago and I have sadly lost the original code. youtu.be/Os1xkUlwjjo

📢 The short description of the tasks is now available on the website 👇 dcase.community/challenge2025/

Great to see how all the CV pioneers have thought about various CV problems back then and how 20 years of research have changed the view on most of these problems. There is still much left to do. It would be great to repeat this series to look back to in 20 years from today.

The Nadaraya-Watson estimator is linear local averaging estimator relying on a pointwise nonnegative kernel. Most of the time, a box or Gaussian kernel is used. https://www.jstor.org/stable/25049340?seq=2

When I was a kid I was fascinated by SETI, the Search for Extraterrestrial Intellitence. Now we live in an era when it is becoming meaningful to search for "extraterrestrial life" not just in our universe but in simulated universes as well. This project provides new tools toward that dream:

Schrödinger Bridge Flow for Unpaired Data Translation (by @vdebortoli.bsky.social et al.) It will take me some time to digest this article fully, but it's important to follow the authors' advice and read the appendices, as the examples are helpful and well-illustrated. 📄 arxiv.org/abs/2409.09347

My book is (at last) out, just in time for Christmas! A blog post to celebrate and present it: francisbach.com/my-book-is-o...

General structure of a paper: - general ideas - general case - general case - general case - what we actually do how it should be: - what we actually do - why we think it's great as one method of a general class - how we got there - how we got there - how we got there

Hyperbolic learning is growing rapidly by the day. From weekly alerts in 2023 to daily digests in 2024! From our current research, it is clear that 2025 will be a huge year for hyperbolic learning research. I had an interview to elaborate our research: shorturl.at/CQD53

Brilliant talk by Ilya, but he's wrong on one point. We are NOT running out of data. We are running out of human-written text. We have more videos than we know what to do with. We just haven't solved pre-training in vision. Just go out and sense the world. Data is easy.

🚀 Introducing the Byte Latent Transformer (BLT) – A LLM architecture that scales better than Llama 3 using patches instead of tokens 🤯 Paper 📄 dl.fbaipublicfiles.com/blt/BLT__Pat... Code 🛠️ github.com/facebookrese...

We @imagineenpc.bsky.social are slowly but surely entering our proposals for master's degree internships here: docs.google.com/document/d/1... These are 6 months projects that typically correspond to the end-of-study project in the French curriculum. Probably more offers to come, check it regularly.

I'm pleased to share that our recent paper with @2ptmvd has been accepted to the Philoshophical Transactions of the Royal Society. Here's the ‘Accepted Author Version’: drive.google.com/file/d/1jdtr... And here it is on arxiv without the fancy formatting: arxiv.org/abs/2409.06219 1/3