Quentin Berthet @ AISTATS @qberthet.bsky.social · 26/09/2026We will be presenting this at NeurIPS in Paris in December! 🇫🇷 📉 0110
Quentin Berthet @ AISTATS @qberthet.bsky.social · 04/09/2026It's a team job! With Aymeric Dieuleveut, as well as the General Chairs Aaditya Ramdas and Arno Solin, we care deeply about the quality of reviews, and this is one of the approaches that we are taking. More details to follow soon! 220
Quentin Berthet @ AISTATS @qberthet.bsky.social · 09/08/2026That's really cool! I have always wondered if we know a bit about the optimal long run race strategy in general (10k, half, marathon+). Feels like it could be tackled with dynamic programming. I don't know if "fastest uniform speed possible for that distance" is optimal, my PBs are neg-split 110
Quentin Berthet @ AISTATS @qberthet.bsky.social · 11/05/2026📝 Read our paper with Yu-Han Wu, Romuald Elie, Klaus Greff, Clément Crepy, and Michaël E. Sander: [arxiv.org/abs/2605.06797] Try out MIND in your evals, let us know about it!arxiv.orgMIND: Monge Inception Distance for Generative Models EvaluationWe propose the Monge Inception Distance (MIND), a metric for evaluating generative models that addresses key limitations of the widely adopted Fréchet Inception Distance (FID). The MIND metric... 050
Quentin Berthet @ AISTATS @qberthet.bsky.social · 11/05/2026Best part? MIND is embedding-agnostic. Drop in CLIP, DINOv2, or even audio/video embeddings! 🎵 Don't give up on embedding-based distances. 110
Quentin Berthet @ AISTATS @qberthet.bsky.social · 11/05/2026Finally, FID isn't a true distance. You can "hack" FID to by matching moments without improving visual quality. We saw this in a recent paper x.com/i/status/20... and observed similar phenomena! MIND is a proper distance and WAY more robust. 🛡️ 110
Quentin Berthet @ AISTATS @qberthet.bsky.social · 11/05/2026Because MIND bypasses massive covariance matrices, it isn't just sample-efficient—it is blazing fast. ⚡ Computing MIND is much faster than FID and uses less peak memory. You can more easily evaluate your models on the fly during training! 🚀 120
Quentin Berthet @ AISTATS @qberthet.bsky.social · 11/05/2026A good metric should correlate well with improvement during training! You want a metric that reliably tells you if step 200k is actually better than 100k. 📈 MIND correctly orders models with far fewer samples than FID, outperforming other baselines, same for perturbations 🏆 110
Quentin Berthet @ AISTATS @qberthet.bsky.social · 11/05/2026Our benchmarks: a good sanity check to evaluate metrics is distinguishing real images from generated ones. 🕵️♂️ FID requires over 10k samples to reliably separate the two distributions. MIND sees the gap clearly with less than 5k! 📉 120
Quentin Berthet @ AISTATS @qberthet.bsky.social · 11/05/2026It is also incredibly easy to implement, and include in your eval pipeline. Because it relies on simple 1D projections and sorting, MIND is very easy to code. Here is the entire metric in just a few lines of PyTorch and JAX! 💻👇 120
Quentin Berthet @ AISTATS @qberthet.bsky.social · 11/05/2026Why are 1D so important? Because in 1D, Optimal Transport is solved perfectly just by sorting! 🤯 We recover the full high-dimensional picture by averaging across many 1D projections. No huge matrices, just fast, parallelizable sorting. ⚡ Some eye-candy equations to illustrate 110
Quentin Berthet @ AISTATS @qberthet.bsky.social · 11/05/2026MIND takes the exact opposite approach. 🔄 We keep the true complexity (no Gaussian approximations!) but reduce the dimension. We project the high-dimensional embeddings onto random 1D lines to capture the true distance, with sliced Wasserstein. 110
Quentin Berthet @ AISTATS @qberthet.bsky.social · 11/05/2026This is due to FID's "trick" to simplify optimal transport: by approximating embeddings as Gaussians (parametric), it ditches complexity and keeps the high dimension. Doing so, it loses crucial distributional info and isn't sample-efficient. 🛑 110
Quentin Berthet @ AISTATS @qberthet.bsky.social · 11/05/2026We often evaluate generated images in high-dimensional embedding spaces (e.g., 2048D for Inception) We then compare the distributions for data / generated To do so, FID estimates a massive 2048x2048 covariance matrix. This is why you need 10-50k samples for a stable eval! 📉 100
Quentin Berthet @ AISTATS @qberthet.bsky.social · 11/05/2026🚨 New paper: Introducing MIND (Monge Inception Distance) Everyone agrees that FID is broken, requires too many samples, slowing down evals. MIND requires 10x fewer samples, is more robust, faster to compute. Our new drop-in replacement for evaluating generative models. 🧵👇 2294
Quentin Berthet @ AISTATS @qberthet.bsky.social · 01/05/2026Very happy to arrive in Tanger for AISTATS 2026, the conference with a palm tree in the logo 🌴 🌊 Let me know if you want to chat! 190
Reposted by Quentin Berthet @ AISTATSNeurIPS Europe @neuripseurope.bsky.social · 23/03/2026Today 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. 18933
Reposted by Quentin Berthet @ AISTATSNeurIPS Conference @neuripsconf.bsky.social · 23/03/2026Following the success of the EurIPS and NeurIPS-Mexico City pilots in 2025, we are thrilled to announce two official NeurIPS 2026 satellite events for this year! These will be held in Paris, France and Atlanta, USA, respectively, running alongside the main venue in Sydney, Australia.neurips.cc2025 ConferenceThe Thirty-Ninth Annual Conference on Neural Information Processing Systems 27017
Quentin Berthet @ AISTATS @qberthet.bsky.social · 04/03/2026I think that the purpose of science is to advance knowledge, not making sure that humans publish papers. I agree that an avalanche of low-quality papers is bad, but if the paper is good, and teaches us useful science, I don't mind that it was written by an AI 010
Quentin Berthet @ AISTATS @qberthet.bsky.social · 18/02/20263) "Causality in the Age of AI Scaling" causcale.github.io organized by David Inouye, Bryon Aragam, Murat Kocaoglucauscale.github.ioCausality in the Age of AI ScalingWorkshop on Causality in the Age of AI Scaling (5 May 2026, Tangier, Morocco) 020
Quentin Berthet @ AISTATS @qberthet.bsky.social · 18/02/20262) "OPTIMAL: (O)ptimisation and (P)os(T)-Bayesian (I)nference in (MA)chine (L)earning" optimal-workshop26.github.io organized by K. Daudel, J. Knoblauch, S. Filippi, V. Fortuin, T. Möllenhoff, J. Odgers, C. Tsay, C. Dellaporta, M. Altamirano, Y. McLatchieoptimal-workshop26.github.iooptimal - OPTIMAL: Optimisation and Post-Bayesian Inference in Machine Learning 120
Quentin Berthet @ AISTATS @qberthet.bsky.social · 18/02/20261) "Towards Trustworthy Predictions: Theory and Applications of Calibration for Modern AI" calibration-workshop.github.io organized by Eugène Berta, Sebastian G. Gruber, Francis Bach, Edgar Dobriban, Teodora Popordanoska, Yifan Wucalibration-workshop.github.ioCalibration workshop 110
Quentin Berthet @ AISTATS @qberthet.bsky.social · 18/02/2026🌴📊📈 - AISTATS 2026 workshops For the first time, AISTATS (May 2-5 in Tangier, Morocco) will have a day of workshops! Please submit your best work by Feb. 27th: - calibration-workshop.github.io - optimal-workshop26.github.io - causcale.github.io virtual.aistats.org/virtual/2026... 👇calibration-workshop.github.ioCalibration workshop 182
Quentin Berthet @ AISTATS @qberthet.bsky.social · 17/02/2026I don't know about the details or the situation you are referring to. The PhD students are Google employees and paid as such 100% during their PhDs. Some amount (overhead , support) is also transferred to the university. I don't know when 110
Quentin Berthet @ AISTATS @qberthet.bsky.social · 16/02/2026🚨 🔬 PhD positions at Google DeepMind in France 🇫🇷 We are advertising Master Level Intern positions at Google DeepMind within our Frontier AI Unit. These could lead to co-advised PhD positions with Google DeepMind and French academic institutions. job-boards.greenhouse.io/deepmind/job...job-boards.greenhouse.ioStagiaire de niveau Master, France / Master Level Intern, FranceGrenoble, France ; Paris, France 12917
Reposted by Quentin Berthet @ AISTATSNina Miolane @ninamiolane.bsky.social · 06/12/2025Arrived #NeurIPS2025 ☀️ If you're interested in the interplay of Geometry, Topology, Algebra w/ Neuroscience & AI, I'll give 2 talks on Sunday: 🌐11:30am: The Algebra of Spatial Navigation (Groups & grid cells) 🍩1:30pm: Topological Deep Learning (Complexes & Graphs) More👇 2183
Reposted by Quentin Berthet @ AISTATSOwensgrandpa @owensgrandpa.bsky.social · 06/12/2025Viola December 6th 2025 763352249
Quentin Berthet @ AISTATS @qberthet.bsky.social · 01/12/2025Intersection of topics: super interesting research and one of the best feel-good movies of all time 041
Reposted by Quentin Berthet @ AISTATSBen Recht @beenwrekt.bsky.social · 01/12/2025In honor of the 39th AI Winter, I’m going to spend the week disentangling the culture and code of reinforcement learning. There may be ranting...argmin.netReformist Reinforcement LearningWhat if we just begin and end with policy gradient? 3283
Quentin Berthet @ AISTATS @qberthet.bsky.social · 01/12/2025Sorry to miss it again! One of these days I'll be in running shape *and* at the same conference as you! Enjoy! 010
Reposted by Quentin Berthet @ AISTATSNina Miolane @ninamiolane.bsky.social · 01/12/2025Thanks @ucsantabarbara.bsky.social for highlighting our #TIMEBestInvention of the Year award! tinyurl.com/4939bxu5 This strengthens our commitment to innovating for women's brain health with @ai-ucsb.bsky.social & @bowers-wbhi.bsky.social 🧠 1105
Quentin Berthet @ AISTATS @qberthet.bsky.social · 01/12/2025Speaking of EurIPS: looking for coffee recommendations in Copenhagen (planning ahead of tomorrow morning) 000
Quentin Berthet @ AISTATS @qberthet.bsky.social · 01/12/202514hrs over four countries, didn't show my passport once, had a pretzel in Koln and a beer in Hamburg. As George Clooney says, it's the best! (Unlike Nespresso) 020
Quentin Berthet @ AISTATS @qberthet.bsky.social · 01/12/2025Today is a day about three very cool things: trains, Europe, and ML! I'm on my way from Paris to Copenhagen, to attend my first EurIPS! Looking forward to it, reach out if you want to chat! 4722
Quentin Berthet @ AISTATS @qberthet.bsky.social · 25/11/2025Nice to be at NeurIPS@Paris, and happy that Google DeepMind could support the organizers of this amazing event! 0171
Quentin Berthet @ AISTATS @qberthet.bsky.social · 24/11/2025Behold! (Best known solution for packing 17 squares) 0140
Quentin Berthet @ AISTATS @qberthet.bsky.social · 14/11/20251) des gens y travaillent 2) tout a fait d'accord ! Je ne justifiais pas la tâche kafkaïenne. Mon approche est plus de chercher un moyen de contourner que de protester, en général 000
Quentin Berthet @ AISTATS @qberthet.bsky.social · 14/11/2025Un des exemples de motivation derrière la recherche académique sur les LLMs je crois : "ça doit être plus facile d'apprendre a des modèles à écrire que de continuer à remplir ces formulaires 010
Quentin Berthet @ AISTATS @qberthet.bsky.social · 13/11/2025Si seulement il y avait une technologie permettant de faire ça automatiquement ;-) 220
Quentin Berthet @ AISTATS @qberthet.bsky.social · 20/09/2025Being a rejected paper seems like a good criteria (possible issues, but very easy to implement) 000
Quentin Berthet @ AISTATS @qberthet.bsky.social · 20/09/2025I'm surprised, there must have been an additional round of changes. All the emails I saw as an SAC were "help your ACs write new meta-reviews". 010
Quentin Berthet @ AISTATS @qberthet.bsky.social · 19/09/2025I think that they are hard to bootstrap. While I agree that general chairs, program chairs, local chairs etc do a lot of work (seeing a glimpse of that sometimes myself with ALT/AISTATS), once you start having a bit of money, it can get easier with using conference organising services 010
Quentin Berthet @ AISTATS @qberthet.bsky.social · 19/09/2025How do-able would it be to organize a conference with its deadline a week after NeurIPS, held a week before, that would always take place physically in Europe (but welcome contributions from anywhere)? 100
Quentin Berthet @ AISTATS @qberthet.bsky.social · 19/09/2025I'm very happy that EurIPS is happening, and planning to attend. But as a NeurIPS SAC I must say that the last definition should not exist - ACs are supposed to change the decisions and update their meta-review themselves. 210
Reposted by Quentin Berthet @ AISTATSClaire Vernade @claireve.bsky.social · 29/08/2025🚨AISTATS workshop proposals open! Big News: For the first time, there will be a day of workshops at #AISTATS 2026, in Tangier, Morocco 🌴🇲🇦 Quentin Berthet @qberthet.bsky.social and I are workshop chairs. virtual.aistats.org/Conferences/... Deadline: Oct 17, AOEvirtual.aistats.orgCall for Workshops 02912
Quentin Berthet @ AISTATS @qberthet.bsky.social · 17/04/2025At #AISTATS2025, I will be giving an "Oral" presentation of our work on "Implicit Diffusion" arxiv.org/abs/2402.05468 170
