Serge Belongie

One week left to apply! Deadline: Oct. 31

Serge Belongie@serge.belongie.com

It’s ELLIS PhD application season! 🎓 @belongielab.org has openings in CV/ML; fine-grained analysis of multimodal data; generative models for images and video; narrative analysis and misinformation detection, self-supervised learning; human-in-the-loop computing. Due date: 31-Oct-2026 🎃

Serge Belongie reposted

Big news for ELLIS! The network is strengthening its presence in Brussels with a new outpost at the European Tech Embassy (ETE). ELLIS President @serge.belongie.com visited the site ahead of its official opening in early 2027. Read the full story: https://bit.ly/4ixdf9N

ellis-unconference.github.ioELLIS UnConference

Registration is open for the @ellis.eu UnConference! 🎉 This is an outstanding community driven event for the machine learning community, taking place the day before @neuripsconf.bsky.social Paris 🗓️ ellis-unconference.github.io

In Oslo today for the GBIF Science Symposium 🇳🇴🌱

GBIF@gbif.org

🎥Tune in today to livestream the GBIF Science Symposium hosted by @gbifnorway.bsky.social! 🌍 The event will be streamed here: 🔗 https://www.gbif.no/ View the full program: 🔗www.gbif.no/events/2... 📸 Maylinn Hovengen Byrknes

ELLIS@ellis.eu

🚨 Apply now for the #ELLISPhD Program! 🎓 Work with leading AI/ML researchers across Europe, with exchange visits & joint supervision. The application deadline is 31 Oct. Get all info & apply now at ellis.eu/news/ellis-p...

10 Sept

Panel discussion at @eccv.bsky.social ECCV 2026 on The Age of LLMs: Who Is the Researcher Now? How LLMs Are Changing Search, Ideation, and Scientific Judgment, featuring @annarogers.bsky.social, @iaugenstein.bsky.social, and @robysinatra.bsky.social, moderated by Anna Rohrbach (1/2)

10 Sept

A timely topic in light of the proliferation of AI-based tools in mathematical and scientific discovery, with insights shared by thought leaders from CAISA - Det Nationale Center for AI i Samfundet, @aicentre.dk, and @nerdsitu.bsky.social in Denmark 🇩🇰 (2/2)

curateddata.github.ioCurated Data for Efficient Learning: ECCV 2026ECCV 2026 Workshop on data curation for efficient machine learning.

9 Sept

Next stop: Curated Data for Efficient Learning workshop at @eccv.bsky.social ECCV 2026 in Malmö

xreason-workshop.github.ioX-REASON | ECCV 2026 WorkshopECCV 2026 Workshop on Visual Perception and Reasoning in the Interactable World.

8 Sept

Next stop at @eccv.bsky.social: X-REASON workshop on Visual Perception and Reasoning in the Interactable World, to talk about challenges in fine-grained image analysis xreason-workshop.github.io

8 Sept

ECCV 2026 @eccv.bsky.social is in full swing! 🇸🇪 First stop: Visual Persuasion workshop sites.google.com/view/eccv-20...

Serge Belongie reposted

1 Sept

REMINDER – Call for Sessions! The ELLIS UnConference comes to Paris on 8 Dec, kicking off the @neuripseurope.bsky.social satellite event. Submit proposals for 2-hour sessions on ML and its applications—any format welcome! ⏰ Due: 15 Sept 🔗 https://ellis-unconference.github.io/ #NeurIPS2026

Silpa Vadakkeeveetil Sreelatha

19 Aug

Inspiring and thought provoking talk by @carlbergstrom.com at @aicentre.dk today on high-risk, high-return research, and why rational scientists avoid it in favor of safe, incremental projects

Serge Belongie reposted

@silpavs.bsky.social

2 July

Presenting RAIGen at #ICML2026 in Seoul! Can we automatically surface rare, underrepresented attributes in text-to-image diffusion models, with no prior knowledge of what they are? Come meet me in poster session! 📅 Wed Jul 8, 10:30–12:15 KST 📍 Hall A, #3308

Serge Belongie reposted

NeurIPS Europe

@neuripseurope.bsky.social

24 June

NeurIPS Europe is launching a Start-Up Sponsorship Package for early-stage teams who want a seat at one of the world's premier AI events. This year, NeurIPS happens simultaneously in Atlanta, Sydney, and Paris (where we expect 4,000, researchers, engineers, founders, and students). 1/2

Serge Belongie reposted

24 June

🎓 G-Research is supporting the #ELLISPhD Program with scholarships for 36 students across Paris 🇫🇷, Tübingen 🇩🇪, Munich 🇩🇪 & Amsterdam 🇳🇱 covering fees and living expenses from 2026–2028. Get more details: https://bit.ly/3QwAFBb

23 June

Thank you for the invitation @munichcenterml.bsky.social! I’m grateful for the opportunity to spread the word about @ellis.eu and present recent research on fine grained image analysis 🙌

Munich Center for Machine Learning@munichcenterml.bsky.social

Our first keynotes of the #MCML Munich AI Day 2026 by Serge Belongie (University of Copenhagen) and Klaus-Robert Müller (BIFOLD) explored how AI is transforming the way we understand the world – from images to scientific discovery.

Serge Belongie reposted

GBIF

@gbif.org

22 May

🌍 25 years of enabling open access to biodiversity data. 🤝 25 years of global collaboration. 🌱 25 years of supporting research and policy worldwide. 🥳 🎂 We’re celebrating 25 years of GBIF on #InternationalDayForBiologicalDiversity, made possible by our incredible network. 💚 #BiodiversityDay

20 May

Fine Grained Analysis 🤝 Robotic Manipulation Evaluation MetaFine, developed in collaboration with Xiu-Shen Wei and team, provides a diagnostic meta-evaluation framework for fine-grained robotic manipulation. (1/4)

The work aims to move embodied AI evaluation from ranking to diagnosis, and to provide a more extensible and reproducible foundation for assessing physical intelligence. (3/4)

metafine.github.ioMetaFine — Diagnosing Fine-Grained Robotic Manipulation Beyond Binary SuccessA compositional meta-evaluation framework that disentangles robotic manipulation into understanding, perception, and behavior, revealing hidden bottlenecks in vision-language-action models.

20 May

Project: metafine.github.io Code: github.com/Hiangx-robot... Preprint: arxiv.org/abs/2605.19986 (4/4)

Instead of relying on binary success rates, MetaFine disentangles manipulation competency into understanding, perception, and controlled behavior, revealing hidden failure modes in current VLA and robotic policies. (2/4)

Serge Belongie reposted

@neuripseurope.bsky.social

19 May

🎉 7 new ELLIS Units join our network Units Franconia 🇩🇪, Grenoble 🇫🇷, and NRW 🇩🇪 strengthen existing hubs; Units Slovenia 🇸🇮and Sweden 🇸🇪 add new countries; Units Czechia 🇨🇿& Denmark 🇩🇰 expand from city to national Units. Learn more: https://bit.ly/4nN3vcD

Serge Belongie reposted

NeurIPS Europe

17 May

FAQ on NeurIPS Europe: NeurIPS Europe is an official NeurIPS 2026 satellite event taking place in Paris, France, alongside the main conference in Sydney and the other satellite event in Atlanta. NeurIPS authors can present their papers at any of the three locations, subject to space availability.

Serge Belongie reposted

Oisin Mac Aodha

@oisinmacaodha.bsky.social

15 May

🎓 Announcing the 1st Bath Summer School in Machine Learning 📅 5–7 August 2026 | 📍 University of Bath Topics: ODEs/SDEs, Bayesian methods, Self-Supervised Learning, World Models Open to PhDs, postdocs & staff at EU/UK universities www.bathmlsummerschool.org

1 May

This interview provides an overview of my thoughts on how AI is developing in Europe vs. other parts of the world:

open.spotify.comEp. 01 - Serge Belongie: AI Is Ordinary Technology

1 May

Spotify link:

Challenges in Fine-Grained Image Analysis Look at this horse. The bushy tail. The big teeth. The hooves.

References: Wei et al., "Fine-Grained Image Analysis with Deep Learning: A Survey," TPAMI 2022 arxiv.org/abs/2111.06119 Yu et al., "Benchmarking Large Vision-Language Models on Fine-Grained Image Tasks: A Comprehensive Evaluation," ICLR 2026 arxiv.org/abs/2504.14988

youtube.comHome Shopping Blooper (FUNNY!)YouTube video by l3igballer23

30 Apr

p.s. Here's the original video clip: www.youtube.com/watch?v=Sv5w...

Challenges in Fine-Grained Image Analysis Look at this horse. The bushy tail. The big teeth. The hooves.

Where do we go from here? I argue that we need some "grand unified" style thinking, e.g., to marry Out of Distribution (OOD) studies with FGVC. GenAI tools perform well within-distribution, but famously confabulate in OOD settings.

If we want to provide researchers tools for scientific discovery, we must focus on OOD, as that is exactly where one finds the loose threads to tug that lead to old theories toppling and new theories emerging.

Challenges in Fine-Grained Image Analysis Look at this horse. The bushy tail. The big teeth. The hooves.

The presenter rehearsed his segment with a large poster of a horse, but at show time, someone replaced it with a poster of a moth. While it's evident the host spotted the mismatch, he nonetheless read the text on the TelePrompTer, describing the attributes of a horse while pointing to the moth.

Is this just a matter of scale? Can we — or should we — simply shove all the expert knowledge we can collect into ever larger VLMs? Can we approach it as a mixture-of-experts and/or a matter of agentic coordination? Time will tell.

Challenges in Fine-Grained Image Analysis Look at this horse. The bushy tail. The big teeth. The hooves.

It turns out there's a large performance gap between the above tailored methods and leading VLMs such as Qwen and LLaVa; see the bar at the far left for iNat.

Does this mean the VLMs are dumb? Not necessarily. In many cases, the vision encoder has the capability to discriminate the classes, but it gets overruled in the process of aligning with the language modality. The following example from the Home Shopping Network provides a useful analogy.

Challenges in Fine-Grained Image Analysis Look at this horse. The bushy tail. The big teeth. The hooves.

State-of-the-art tailored methods for multimodal fine grained categorization, as deployed for example in the Merlin Bird ID and iNaturalist apps, achieved excellent performance using various CNN or ViT backbones circa 2021.

As the Generative AI era emerged, with CLIP-style encoders and open-world chat-based interfaces, one can ask how well off-the-shelf VLMs can solve the fine grained classification problem.

Challenges in Fine-Grained Image Analysis Look at this horse. The bushy tail. The big teeth. The hooves.

In these domains, we often face the additional challenge of long-tailed distributions, meaning that some categories have abundant examples, while many categories have scant coverage.

Naming is capture. Labels supplied by humans — with all the bias that entails — confer meaning to instances that otherwise would merely be points in a high dimensional space with various measures of pairwise similarity.

Challenges in Fine-Grained Image Analysis Look at this horse. The bushy tail. The big teeth. The hooves.

Here is a brief overview of the talk I gave yesterday on Fine-Grained Image Analysis at the @ellis.eu Computer Vision Workshop hosted by INSAIT in Sofia, Bulgaria. Fine-grained categories sit on a spectrum with basic level categories on one end and biometrics and re-identification on the other.

Fine-grained problems are characterized by (a) tightly related categories that require highly discriminative learned representations, and (b) reliance on human expertise to obtain the labels.

candidate.hr-manager.netSoftware Engineer - Department of Computer ScienceWe are seeking a highly capable and motivated Software Engineer to design and develop a state-of-the-art scalable infrastructure for digital narratives. This wo

27 Apr

Software Engineer position open at U. Copenhagen and @aicentre.dk on the Narratives project, co-advised by @ysiglidis.bsky.social, @mariaa.bsky.social (CU Boulder), and me: candidate.hr-manager.net/ApplicationI...

24 Mar

Congratulation Edwige Cyffers, @pierremarion.bsky.social, and all of team Paris on this milestone for the machine learning community! 🎉🇫🇷

NeurIPS Europe@neuripseurope.bsky.social

Today NeurIPS is announcing our official satellite event in Paris. After responding to the call from Ellis following the success of EurIPS in December, we are pleased to reach a new milestone by joining forces with the NeurIPS organizing committee for the 2026 edition.

20 Mar

Feeling overwhelmed by all the recent developments in video understanding? What used to require dozens of modular computational workflows involving SLAM, feature tracking, optical flow, camera calibration, multiview geometric constraints, and resnet backbones is now... (1/3)

20 Mar

— whether one likes it or not — transformer archtectures all the way down. This new survey article by @aicentre.dk PhD students @zhaochongan.bsky.social, @jiaangli.bsky.social, et al. provides an overview of video understanding techniques spanning geometry and semantics circa 2026. (2/3)

arxiv.orgVideo Understanding: From Geometry and Semantics to Unified ModelsVideo understanding aims to enable models to perceive, reason about, and interact with the dynamic visual world. In contrast to image understanding, video understanding inherently requires modeling te...

20 Mar

We hope you find it useful, and we welcome your feedback on it: arxiv.org/abs/2603.17840 (3/3)

17 Mar

@luciagordon.bsky.social (Harvard / P1) presents “MMEarth-Bench: Global Model Adaptation via Multimodal Test-Time Training” at the @aicentre.dk Fine-Grained Analysis Collaboratory Meetup at @sdu.dk Vejle (1/2)

mmearth-bench.comMMEarth-Bench: Global Model Adaptation via Multimodal Test-Time TrainingGlobal, multimodal environmental benchmarking datasets for pretrained models and a multimodal test-time training framework for global model adaptation

17 Mar

Learn more about MMEarth-Bench at mmearth-bench.com Thank you @jbhaurum.bsky.social & team for organizing this great event! www.aicentre.dk/events/20260... (2/2)