A simple argument for equivariance at scale: 1) At scale, token-wise linear layers dominate compute. 2) Token-wise linear equivariant layers implemented in the Fourier domain are block-diagonal and hence fast.

😱😱😱 @jvgemert.bsky.social are you ok? You could have told us all you were a ghost from the beginning!👻

There are diminishing returns. Stop trying to improve review quality by excessive extra work. Remember the irreducible Bayes Error: If the task has inherent subjectivity, it will not help to get better quality measurements.

🚀 Today we're launching the 'Night Science Institute', a non-for-profit organization to lead a cultural shift in science! You may say we're dreamers 😉, but we think we're not the only ones. Perhaps today you will join us to make the Day Science and Night Science parts of the process live as one!

Guilty 😅

Great post on anti-bureaucracy, machine learning theory, and pull requests.

The biggest publishers pulled profit margins in 2023 that rival Big Tech. And they still charge you to publish and your library to read. When are we going to start seriously thinking of alternatives to reform #PeerReview? 🧪 #SciPub #AcademicPublishing

The NeurIPS paper checklist corroborates the bureaucratic theory of statistics.

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!

Instead of meeting in the home country, we do it in a country far away 🙂

"Academia is built upon a foundation of review. It is messy, decentralized collective decision making that selects a group of people and their writing into a canon of thought."

Worth your time to read this characteristically thoughtful blog post by Leon Bottou on LLM research. leon.bottou.org/news/two_les...

Amazing work on evaluating LLM evaluation.

Evaluating evaluation 🙂

Uhhh, hi! It's me, Bookshop.org, a place to buy books online where every purchase financially supports indie bookstores instead of some billionaire.

⏳Only 1 day left to submit your reviews! ⚠️Late or careless reviews won’t be taken lightly — authors of such reviews risk having their own submissions desk rejected. Please be fair and responsible!🙂

In today's episode of the Night Science Podcast we talk with Martin Schwartz from Yale about the importance of stupidity in science: while learning science makes you feel smart, true scientific discovery often involves feeling stupid, because it means venturing into the unknown.

Don’t be that person—take your #ICCV2025 reviewer responsibilities seriously.

After a short era in which people questioned the value of academia in ML, its value is more obvious than ever. Big labs stopped publishing the minute commercial incentives showed up and are relentlessly focused on a singular vision of scaling. Academia is a meaningful complement, bringing... 1/2

🔥🔥🔥 CV Folks, I have some news! We're organizing a 1-day meeting in center Paris on June 6th before CVPR called CVPR@Paris (similar as NeurIPS@Paris) 🥐🍾🥖🍷 Registration is open (it's free) with priority given to authors of accepted papers: cvprinparis.github.io/CVPR2025InPa... Big 🧵👇 with details!

arxiv.org/abs/2503.15156 This is a very interesting paper about locally anchored implicit neural representation. To me, it looks more like a joint keypoints detector and descriptor trained for image reconstruction. I wonder if these approaches could be tuned for geometric tasks or retrieval tasks.

*Please repost* @sjgreenwood.bsky.social and I just launched a new personalized feed (*please pin*) that we hope will become a "must use" for #academicsky. The feed shows posts about papers filtered by *your* follower network. It's become my default Bluesky experience bsky.app/profile/pape...

Get to know @gulvarol.bsky.social, Senior Researcher at École des Ponts ParisTech 🇫🇷 She's an ELLIS Scholar, member of ELLIS Unit Paris, former winner of the ELLIS PhD Award. Her research covers computer vision, vision & language, human motion generation & sign languages. #WomenInELLIS

I'm proud to have been awarded a prestigieus VICI grant (€1.5m) for data efficient AI foundation models. AI with less data breaks our dependency on third party companies, and, allow curatable datasets for bias, privacy, and transparency. www.tudelft.nl/en/2025/tu-d...

I love this type of exploratory, elegant, research! What would it take to generalize these findings to a testable hypothesis?

Do large language models develop "emergent" models of the world? My latest Substack posts explore this claim and more generally the nature of "world models": LLMs and World Models, Part 1: aiguide.substack.com/p/llms-and-w... LLMs and World Models, Part 2: aiguide.substack.com/p/llms-and-w...

As someone who has reported on AI for 7 years and covered China tech as well, I think the biggest lesson to be drawn from DeepSeek is the huge cracks it illustrates with the current dominant paradigm of AI development. A long thread. 1/

Language seems not needed to make concepts; but language is needed for shared concepts (at least it seems like that for color) www.pnas.org/doi/10.1073/...

The origin of color categories | PNAS www.pnas.org/doi/10.1073/...

75% of the reviews arrived. The deadline is in 2 hours 😬 When you know you're late, message your AC to avoid unnecessary stress.

Interested in doing a PhD in Computer Vision Super-Tools in my group at UCL? If you count as a "UK Home" resident, see my PhD student vacancies here: www0.cs.ucl.ac.uk/staff/G.Bros....

Don’t forget that in 2025, the days of the week will be determined by your favourite LLM, trained on day/month pairs from 2004-2023. So Jan 2 will have odds of: 1/10 of being Tue or Thu 3/20 of being Wed, Fri, Sat, Sun 1/5 of being Monday

62 years later, the book that changed everything is still a must read. Kuhn distinguished between 'normal science' and 'revolutionary science', where in the former we work within the paradigm but if anomalies add up, a new paradigm emerges in a period of revolutionary science.

Great reading, as usual, by @melaniemitchell.bsky.social

Some of my thoughts on OpenAI's o3 and the ARC-AGI benchmark aiguide.substack.com/p/did-openai...

It's possible to do good machine learning research, even without impossibly huge data, without enormous compute clusters, without architecture hacking, and without making unrealistic assumptions of convexity, Gaussianity, etc. Intriguing Properties of Robust Classification arxiv.org/abs/2412.04245

Arxiv sharing reminder pdf ❌ abs ✅

I'm slowly putting my intro to ML course material on github, starting with the lab sessions: github.com/davidpicard/... These are self-contained notebooks in which you have to implement famous algorithms from the literature (k-NN, SVM, DT, etc), with a custom dataset that I (painstakingly) made!

I made the following logo for my "Machine and Deep Learning" course. ('MDL' cameo intended ;) ) Can you spot the two? There's "Machine Learning" and there's "Deep Learning" ? 😊