Jérôme Rony

Check out our new paper on adversarial attacks for semantic segmentation! arxiv.org/abs/2206.07179 Using a proximal splitting, we generate small perturbations (e.g. ‖δ‖_∞ < 1/255) for deep segmentation models, with high pixel success rates.

@YouTubeInsider @DeanoSauruz @earthklaans You are trying to be precise but you are mistaken in what GenAI means. Tranformers are not GenAI per se, they are a type of neural network building block that can be used in so called GenAI, but does not have to be.

Hi! I'm a tech nerd and I try to be precise about the terminology I use GenAI typically refers to technologies like transformers and large language models, which are relatively new Upscaling typically refers to taking one resolution (like SD/480p) and making it look good at a higher resolution (like HD/1080p) This isn't using GenAI or doing any upscaling It's using the kind of machine learning you experience with computational photography on smartphones, for example, and it's not changing the resolution And sincerely appreciate the feedback! 🙏

youtube is applying post-processing to all shorts so they look more like Al. this is intentional & malicious. when their generative Al shorts feature launches users will be less likely to discern what's Al-slop and what's not, and the audience will think Al video is just as good

@raymondh Second one: it implies that h(x) is the default case, and g(x) is a specific one conditioned on test. I think it is also dependent on what test is: are we expecting test(x) to be mostly true ?

#Python practices question: Do you prefer else-clauses in conditional where there is a return on both sides? def wordy(x): if test(x): return g(x) else: return h(x) or def terse(x): if test(x): return g(x) return h(x)

If you are at #CVPR2023 drop by to see our posters for tomorrow. AM: 289 - A strong baseline for generalized few-shot semantic segmentation. PM: Class adaptive network calibration. with @IsmailBenAyed1 @bing_bingyuan @jerome_rony @adrian_galdran Sina Hajimiri and Malik Boudiaf

@DrewLinsley That's the issue, even C&W is not strong enough to assess adversarial robustness

@jerome_rony Thanks for the feedback. We found the same of results on a very small scale study of C&W. We can try replicating with the others as well. Appreciate it!

The sensitivity of DNNs to adversarial attacks has long been thought to be an "Achilles heel" that will ultimately make them unsafe in real-world applications. Has this sensitivity changed as DNNs have scaled-up and rivaled or beat human-level accuracy? serre-lab.github.io/Adversarial-Al…

@DrewLinsley I can see a major issue here: \ell_2 PGD (and even C&W) is known to be weak to evaluate a model's robustness. There are much more reliable options (and less expensive than C&W) now: APGD, DDN, FMN, ALMA to name a few.

What we found was surprising. Across 283 DNNs, we saw that the size of successful attacks has been growing as DNNs have improved on ImageNet. In other words, the scaling trends designed for improving ImageNet acc have also led to state-of-the-art adversarial tolerance!

@Vicnent @MathisHammel @JacquesDelamar2 Sur les tests de Mathis, c'est environ 15% plus lent avec la condition (testé sur ma machine car CoderPad n'est pas très régulier sur les run-times)

@jerome_rony @MathisHammel @JacquesDelamar2 oui clairement, ça dépend, du temps d'éxécution de la boucle, du nombre de maisons, de la répartition aléatoire des maisons etc. mais je me demandais si on était sur du temps d'exécution supplémentaire de 0,0001% ou 0,01 % ou 1%. Tu testes ? (je peux pas copié collé ton code)

Exercice inspiré d'un vrai entretien Google ! Pouvez-vous optimiser ce code qui calcule la plus longue suite de maisons ne contenant pas de couleur en double ? Obtenez votre score ici : app.coderpad.io/sandbox?questi…

@Vicnent @MathisHammel @JacquesDelamar2 Ça va beaucoup dépendre de la longueur de la meilleure solution: si elle est significativement plus courte que la liste initiale, ça n'affecte pas la complexité

@MathisHammel @jerome_rony @JacquesDelamar2 est ce qu'une boucle qui optimiserait une condition finale serait intéressante ? Ex : tu as 200 maisons, tu as len_max = 40, et ton next start est sur 162 (et donc tu trouveras une longueur max de 39 quoi qu'il arrive). Donc avoir un break si new start > len(color)-len_max ?

@CSProfKGD Oh I did that in the first comment to start the discussion.

@jerome_rony May even want to summarize diverging views.

#CVPR2023 reviewer thoughts: Not helpful when reviewers skip the discussion and simply finalize their review. Please engage with your fellow reviewers. Even if no consensus is reached, the discussion will provide the AC a better understanding of the various views.

@MathisHammel IMO la façon la plus simple de s'assurer que le cas négatif soit clair.

Vous avez vu passer ce code source extrait d'un projet du gouvernement néerlandais ? Je vous propose un petit challenge : proposez une version refactorisée à votre sauce ! Sandbox : app.coderpad.io/sandbox?questi… La solution avec le plus de likes gagne un pack de goodies CodinGame😉

@colorado_reed We did! We worked on a problem that is not novel, but also somewhat understudied IMO.

@CSProfKGD No worries, just sharing my experience 🙂 Even though I sometimes end up disappointed for my own papers, I'm extremely grateful of the work ACs do!

@jerome_rony It’s tough, I get it. Especially, this time of year with conference paper submission & review deadlines. I also had new papers sent to me as an AE on Xmas to distribute. I held off till January to send reviewer requests. We all deserve a break. I was just venting a bit 😉

Note to self: Don't take it personally when people reject your review request.

@Ana_Kay31 @_davyben @CultureCrave Should have clarified: Tick-Tock is the long version of Mountains in the Interstellar expanded album

@jerome_rony @_davyben @CultureCrave For Interstellar or The Last Samurai?

Hans Zimmer says his favorite score that he's ever written is #Interstellar 🪐

@Ana_Kay31 @_davyben @CultureCrave You should check the extended version called "Tick-Tock" in the expanded version of the album!

@_davyben @CultureCrave Mountains is so good on the Interstellar soundtrack 🔥. Spectres in the Fog from The Last Samurai is pretty a amazing too.

@CVPR @mahmoudnafifi Is there going to be an additional week for review?

@mahmoudnafifi Assignments are a bit delayed — they are expected to be released around next Wednesday.

@CVPR when do we expect to get assigned papers for review?

@benjamin_warner @wightmanr The EMA step could be even faster using: >>> torch._foreach_mul_(ema_tensors, decay) in the first step.

Here is the code for my EMA update step. The standard EMA step is lightly modified from .@wightmanr's ModelEmaV2 in PyTorch Image Models. github.com/rwightman/pyto…

PyTorch's fused foreach methods, used in their multi-tensor optimizers, are surprisingly undocumented given how useful and fast they are. Rewrote my EMACallback and the EMA step is now significantly faster on tested models vs a standard for loop over both sets of weights.

@le_science4all @NeurIPSConf Ce n'est pas obligatoire d'écrire ce paragraphe pour NeurIPS.

Je félicite au passage @NeurIPSConf, qui exige des auteurs qu'ils écrivent un paragraphe sur les limites et les conséquences sociales des articles soumis, et des revieweurs qu'ils intègrent cela dans leurs reviews. La science est devenue trop dangereuse pour être une fin en soi.

Les applications médicales sont devenues l'#ethicswashing du machine learning académique : il y a tellement d'articles de recherche qui motivent leurs algorithmes avec ces applications, eu égard aux conséquences sociales néfastes de ces mêmes algorithmes. theverge.com/2022/3/17/2298…

@AMurderOfDucks @gabrielpeyre You can use it to minimize a sum of nonsmooth functions (e.g. infinity norm) by iteratively minimizing each using their respective proximity operator arxiv.org/abs/0912.3522

@gabrielpeyre What are some of the uses of this algorithm?

Oldies but goldies: R. Tyrrell Rockafellar, Monotone Operators and the Proximal Point Algorithm, 1976. The proximal point algorithm is a fundamental non-smooth optimization method. en.wikipedia.org/wiki/Proximal_…