Computer Vision/Machine Learning research
@ImagineEnpc
/ LIGM, École nationale des Ponts et Chaussées. Music & overall happiness 🌳🪻 Born well below 350ppm
🎨We updated the demo, you can try your favorite text2image prompts:
Be gentle: it's a tiny model by image generation standards (300M params) trained from scratch on a ridiculously small dataset (20M img+txt pairs), so it doesn't have high def capabilities
Research in computer vision at the moment is crazy fast. You discuss a problem with your colleagues, find shortcomings in the literature and a way to address them, flesh out a plan, and 5 days later there are already 2 papers appearing on arxiv that do exactly what you wanted.
"torch.manual_seed(3407) is all you need"!
draft 📜:
Sorry for the title. I promise it's not (entirely) just for trolling. It's my little spare time project of this summer to investigate unaccounted randomness in
#ComputerVision
and
#DeepLearning
.
🧵👇 1/n
Please
@OpenAI
change your name ASAP. It's an insult to our intelligence to call yourself "open" and release that kind of "technical report" that contains no technical information whatsoever.
Folks, I'm furious🤬and need to vent. I have this paper that has been in rejection hell for a year and it needs some love🥰
If you're in retail/image search, it could have major impact on your business.
I'll explain it👇and then rant on bad reviews.
📜:
Just put online my
#MachineLearning
course at
@EcoledesPonts
!
It's not very different from the bazillion that already exist, but it's mine! 🥰
Also, it's all python notebooks cause I want to show actual implementations and not just fancy figures. 🤖
Wow! Looking at other reviews, I'm impressed by the nastiness of my fellow reviewers for
#ICCV2021
. So many "strong rejects" that are "very confident" and particularly aggressive. Chill out, it's not the end of the world if you don't like that paper!
Wondering how to detect when your neural network is about to predict pure non-sense in a safety critical scenario?
We answer your questions in our
#ICCV2021
@ICCV_2021
paper!
Thursday 1am (CET) or Friday 6pm (CET), Session 12, ID: 3734
📜
Thread 🧵👇
Stop believing that research is about coming up with the most convoluted system. It's not.
Research is about asking and answering questions. Practical or theoretical questions.
Go back to your local science fair to present your Frankenstein engineering shit, and stop reviewing.
So happy to finally be able to talk about this paper "Unveiling the Latent Space Geometry of Push-Forward Generative Models" with
@TIssenhuth
,
@ugotanielian
and
@JeremieMary
To me, it's an important paper, because we show that analysis and synthesis are linked.
🧵⬇️ (ICML'23)
Imagine you want to compare to a SotA method published at the latest major computer vision conference, but there is not implementation of their code. So you email the authors for tips to reproduce their results and they answer "Read the paper carefully".
What do you do? 🤬
@ykilcher
It's not a even a paper. It doesn't describe the model, it doesn't describe de data, it doesn't even mention the compute used...
What knowledge do researchers even gain from this? Nothing. It's super frustrating.
"7.7M GPU hours of computation on hardware of type H100-80GB (TDP of 700W)"
That's 5.4GWh. That's 6 hours of an entire nuclear reactor. AN ENTIRE NUCLEAR REACTOR.
And you know what, this is just the beginning. So long for frugality...
What's annoying with CVPR rejects, is that you're guaranteed to find a dozen accepted papers on the same subject that are just really uninteresting at all, with nothing exciting, limited results and mediocre intellectual qualities. But hey, the viz are nice and the names famous!
Long overdue extension of our ICCV'21 paper on out-of-distribution detection for semantic segmentation to objects. Work by
@victorbesnier1
with
@abursuc
.
We extend OOD to panoptic segmentation!
📜
mini 🧵1/4
Few bits of advice if it's your first time in Paris for
#ICCV2023
:
1. Take the metro, it's the fastest transport mode and the weekly card is cheap. How far is it? Count the number of stations, multiply by ~1.5-2, that's the number of minutes it will take.
architecture nobody did before, you're not doing research, you're doing shitty-ass undergrad engineering.
Research is about gaining knowledge, not frankenstein-ing whatever combination of hot tools that nobody has done before and tinkering with is until is gains 1% by pure luck.
This coffee shop uses AI to track the productivity of baristas and how much time customers are spending in the shop.
The NeuroSpot Barista Staff Control and Customer Monitoring Video Analytics Module, are tools designed to enhance the efficiency and
How is it that nobody is interested in out-of-distribution detection for semantic segmentation?
- It has important practical applications
- The SotA is kinda low
Come on folks! It's a low hanging fruit. Download our code and beat us, it's easy:
#CVPR2022
Today is SVM and Kernels lecture day, one of my favorite days!
Kernels don't get enough love, they deserve much more. Even if you don't use them, they give excellent insights to the properties of any learning machine.
Just put online my
#MachineLearning
course at
@EcoledesPonts
!
It's not very different from the bazillion that already exist, but it's mine! 🥰
Also, it's all python notebooks cause I want to show actual implementations and not just fancy figures. 🤖
So the Gemini paper () has 941 authors.
We are now well into the "CERN era" of AI development. Labs from public universities will need to either organize big networks, or focus on tabletop experiments...
Let's do a "collaborative science experiment" for
@icmlconf
!
I have a project replacing MHA with high order moments:
I'm training a ViT-like (⬇️), but I want to do more than that.
Who wants to help turn it into an ICML sub? (DM, gh issues, email and RT🙏)
I took a 2h30 long walk in Seattle (I'm European, we like to walk in cities), and it's truly an enchanting place. Very nice and pleasant to wander. One simple trick could make it even better: remove all the cars. Imagine what a paradise it would be.
#CVPR2024
@CVPR
@mtutek
Well, and I kid you not, I ran into this sort of code while reviewing a paper that had provided source code in the supplementary to reproduce the results.
To be fair, it did reproduce the scores written in the paper 🤣🙄
🚨Early preprint worth sharing!
If you're using diffusion models with classifier free guidance, we have a 1 line modification to improve your results
📜:
🤖: No git yet, but It's 1 line. I expect someone will add it to
@huggingface
diffusers at some point
Classifier-free guidance has been crucial in recent advancements in content generation, trading diversity for fidelity. However, can you get a better deal in this trade-off?
With a simple modification, you can get more diversity while keeping a higher level of details!
So apparently, I'm an outstanding reviewer for
#NeurIPS2021
which I should be happy and proud about, but I'm a bit worried too. I couldn't put in the effort I think they require (burnt out probably) and had I received such reviews myself, I would have thought "OK, but not great".
This work has seen 15 reviewers in a year, and 90% of them rejected the paper for the same reason: lack of novelty.
"BuT YoU hAvE OnLY a SimPLe ViT wiTh ConTrAsTiVE LeArNinG!"
Guys, you are just stupid, and toxic. If you think research is coming up with a complex ...
🎉Super happy for "Triggering Failures: Out-Of-Distribution detection by learning from local adversarial attacks in Semantic Segmentation" with Victor Besnier and
@abursuc
to be accepted
@ICCV_2021
! One step towards safer autonomous driving 🤖🚗 More details/ paper/code soon! 😎
Hot take: Constrained scientific writing (like a 1 page rebuttal) introduces a massive disadvantage for non-native English speakers - unfair IMHO. LLMs like ChatGPT are finally getting rid of that and put everybody on the same level w.r.t. writing skills.
Bonjour
@Interieur_Gouv
, notre système de visa marche tellement bien qu'une scientifique en IA de renommée internationale, récemment primée par l'ERC, ne peut même pas venir 10 jours pour présenter ses travaux à une conférence prestigieuse.
Une idée pour améliorer notre image ?
I still don’t have a visa for ICCV. let’s keep telling ourselves how inclusive we are.
the summer I got 3 papers and an ERC and still wasn’t allowed in Europe.
if you read this and think I deserve this because I want to live in the same country as my family, just unfollow me.
Can we do the same for CV? Authors of TPAMI papers get to present their work at CVPR/ICCV and authors of IJCV papers get spots at ECCV?
What can I do to make that happen? I'm volunteering to put this system in place.
We (
@BeEngelhardt
,
@NailaMurray
and I) are proud to announce the creation of a Journal-to-Conference track, in collaboration with JMLR and conferences NeurIPS 2022, ICLR 2023 and ICML 2023!
Tomorrow is the last day at
#ICCV2023
, if you're looking for opportunities in France with me or with the amazing people of my lab - PhD level, postdoc, young researcher - try to catch me and I'll find time to discuss.
I'm easy to recognize, if you can find a recent enough photo.
🚨 We've released a large number (15+ 😲) of pretrained models for this paper! Both for creating the dataset and for whole-body 3D pose estimation.
Just in time for your
@CVPR
#CVPR2024
submission! Boost your results with 3D whole-body!
Links on github:
🚨What about doing 3D pose estimation, but with the 133 keypoints of the COCO whole-body layout?
We have a cool poster with Yue Zhu and
@nemka_
Friday PM at N018. I'll show the many good tricks in the long 🧵 below.
💻
📜
1/12
More stunning views! If you're looking for a place to do graduate studies in computer science, you should definitely add Grenoble on your map. The research environment is excellent, and I can recommend you many advisors who will give you the best experience you can ever hope for.
The idea is simple: after years of reviewing deep learning stuff, I am frustrated of never seeing a paragraph that shows how robust the results are w.r.t the randomness (initial weights, batch composition, etc). 2/n
@BertozziRoger
@colinlebedev
Tout le monde a le droit d'exprimer son avis, aussi crétin soit-il. Ce qui est pénible, c'est l'utilisation de l'étiquette "universitaires et chercheurs" pour donner un argument d'autorité. C'est d'autant plus pénible quand on est soit-même chercheur. C'est une mauvaise pratique.
@ykilcher
"Grassroots acceptance of a paper by the community" means "the one with the biggest marketing budget takes all" like in art. Science != popularity. Making something that appeal to a lot of people doesn't make it true or even relevant, or we would all still be studying the aether.
This session was the worst I have ever experienced both as an author and as a reviewer. I got 144 notifications from openreview in 15 days. 144. In the middle of August.
Reviewing currency idea:
- New accounts get 3 tokens
- Any submission requires 3 tokens (all authors can contribute any amount)
- Reviewing 1 paper earns 1 token, subject to quality approval from the AC
If you want to submit past your 1st, you have to produce quality reviews.
Things I discovered on Monday
@CVPR
: Now I really want David Forsyth to narrate my life with the voice-over of God coming from everywhere at the same time, as I do my daily activities. With the echo of course.
@jbhuang0604
Here is my only advice to early career researcher and students: Do something you like a lot with people you like a lot, and nothing else. Your chances of having a good career by doing something you don't like with people you don't like are tiny to the point of being negligible.
Please don't do that. I know you're excited, and I'm excited for you too. But this is something you can enjoy without screaming and annoying everybody. Instead, just raise your glass and press F for all the fallen brothers and sisters.
Everybody complains about deep neural networks being over-confident at rejecting good samples, but have you read
#NeurIPS2023
reviews made by humans?
Joke aside, I sense a strong correlation for reviews with a confidence of 5 being total garbage.
@CVPR
Are you good? Do you need help? I mean, my first thought was that someone broke into your house and forced you to write that under threat...
Let's use coded language: Post "
#CVPR2024
will not allow bold numbers in tables anymore" if you need us to send a SWAT team.
From now on, I'm going to strong reject any paper that doesn't have a clear scientific agenda (question, hypothesis, method, experimental validation), and I don't care if you gain X% compared to SOTA. It's made up by grinding your luck anyway.
I've been way to lenient on that.
If you think a CVPR publication is highly valuable because the acceptance rate is low*, think about this: there are only ~250 outstanding reviewers out of more than 7000 reviewers. It requires way more skills to be a excellent reviewer than an author.
* You shouldn't, btw.
We sincerely thank all reviewers, area chairs and senior area chairs who contributed their time to
#CVPR2023
! Reviewers who did an outstanding job are recognized here:
@CSProfKGD
This year's ICLR has so much drama. This one is special because the review contains nothing about the content of the paper (whether it's correct, novel or impactful), but only cares about the paper containing the word "byzantine". I think it's a first?
R1: "The paper is well structured and easy to follow."
R2: "The paper is poorly written and organized."
🤷♂️
What's your opinion, R3? Speak now or (we) forever hold your piece!
Folks, this is the project you want to contribute to! It's 🔥
Simple: you fork, add you stuff (models, training and evaluation scripts, notebook, whatever), send a PR and I merge.
And we all do good collaborative open science!
github:
@LBantigny
En Saxe? Sachsen? Pas les autres saxes, hein? Parce-que le résultat des élections européennes en Saxe ne me donnait pas exactement l'impression d'une baisse de l'extrême droite.
@fhuszar
Here's an ablation performed by on of my PhD student (N. Luka) about 2 years ago, unpublished, but gives you an idea about what components matter or not. Bottom line: the attention has to sum to one, and that's all.
@CSProfKGD
It's just allocating resources to rich kids. If the goal is a social one, we ought to do better. Past CV conferences had high schoolers visit the conference and that was great. Having a dedicated track is just giving a trophy that reinforces inequalities.
@jbhuang0604
Just like A+ conferences have a 99% acceptance rate and everybody moves to a big company with a 6 figures salary.
Isn't that a perfect world?
Here's a conjecture of mine: "If a deep neural model has a high enough number of parameters, it tends to memorize the training set, but for large training sets, memorization is very effective at solving the task."
Every day of work makes me a bit more convinced this is true.
I don't buy the "brains are more energy efficient than silicon computers, they only use 20W!!1!"
How many joules does a brain need to compute the SVD of a 4000x4000 matrix? Pretty confident it's way more than a computer.
They're different tools optimized for different purposes.
In this week's science news: computers made of human brain cells, galaxies that are too big to exist, what the Milky Way’s black hole is having for dinner, Google’s second milestone on the way to quantum computing, and more
👏 These are real heroes of the conference! If you see any on the job market, close your eyes, relax, and make an offer. These reviewer awards are a much less noisy signal then any publication track record.
Well done and congrats to this year's awardees!🥂
@ylecun
Ah! I really love this old '88 paper about the Lagrangian view of backprop. I've used its ideas many time for the final exam of my ML course, which most of the students find (surprisingly) easy (and hopefully insightful too). Great paper, underrated I think!
I strongly disagree with this analysis. My take: Science is progressing faster now, but progress is shared among a larger base of contributors. People complaining just don't like the idea there are no genius, and the scientists of the past were just bourgeois winning the lottery.
We see constant progress in the world every day, from better cars to faster phones to virtual reality and the internet of things. However, despite all the technological and engineering advances, science seems to be slowing down.
Memory is all you need.
I know I sound like a broken record, but k-NN *is* a valid generalization strategy. So far, nothing tells us that extra-big neural networks aren't just learning a clever hash function and memorizing their training set.
New paper out today, asking: What books has ChatGPT/GPT-4 *memorized*? A LOT. Harry Potter, Pride & Prejudice, 1984, LotR, Hunger Games, GoT, 50 Shades of Grey, Dune. Memorization is linked to web popularity--lots of old classics + new sci-fi/fantasy. 1/6
Question for diffusion models people: To me, it looks like a DM is an iterated function system that is trained such that the training set is its attractor. What exactly prevents it from just memorizing the training set and output trajectories that land on training samples?
@francoisfleuret
I think if you answer no to this question, you should probably look for another job. If there is no hunch to discover or to understand, then you'd probably be happier doing something less uncertain, like engineering.
If you're using k-means for anything else than compression, you may be interested in my JAX implementation.
It's reasonably fast and features codeword shifting to make sure no cluster is super bad. Works very well for building visual dictionaries!
💻:
I can craft a random table where you throw a 20-sided dice 3 times, and get your next paper: "Neural ODE on the Grassmann manifold for open-set recognition".
Is this a research project? Hell no. It's a random-ass combination of stuff.
There's no research question, no logic to it.
The distribution of accuracy of as follows. It's pointy, but has outliers nonetheless. The diff between min/max is just below 2%, which is huge for computer vision! 6/n
@m_j_kremer
It won't change the situation, but you can think of Michael Jordan, the great professor of machine learning at Berkeley, every time he is invited to do a keynote at a prestigious conference, as he checks in the five star hotel just to see the disappointed face of the attendant.
Une anecdote qui m'a profondément énervé: Un de mes anciens doctorants, chinois, a fait prépa en France, l'X, thèse aux Ponts. Son visa a été prolongé jusqu'au jour de la soutenance, pas 1 de +, avec l'argument "vous soutenez le matin, vous pouvez prendre l'avion l'après-midi".🤬
The optimal computation of gradients for the composition of functions is an optimal parenthesis problem. Forward and backward (backpropagation) are two extreme cases. Backward is optimal for scalar-valued functions.
Our paper "Learning Uncertainty for Safety-Oriented Semantic Segmentation in Autonomous Driving" was presented this week at
#ICIP2021
with V. Besnier and A. Briot.
(the pdf is free for limited time so download it now)
A thread 🧵👇 1/6