Sid

Sid

@sidgairo18.bsky.social

πŸ”οΈπŸ“πŸ‡©πŸ‡ͺ πŸ‡ͺπŸ‡Ί PhD student at the Max Planck Institute for Informatics, and Institute of Science & Technology - Austria. πŸ’»πŸƒπŸ»β€β™‚οΈπŸš΄πŸ»πŸ‹πŸ»πŸŠβ›·οΈπŸŽΈπŸŽΉπŸ“š Webpage: https://sidgairo18.github.io/

63 followers142 following103 posts

Sid
@sidgairo18.bsky.social

Any plans for having a 2nd iteration of the CVPR@Paris meeting πŸ€”? Given that last year's event was great and very much appreciated it does call for another one πŸ˜‰ @davidpicard.eurosky.social ?

David Picard@davidpicard.eurosky.social

πŸ”₯πŸ”₯πŸ”₯ 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) πŸ₯πŸΎπŸ₯–πŸ· Registration is open (it's free) with priority given to authors of accepted papers: cvprinparis.github.io/CVPR2025InPa... Big πŸ§΅πŸ‘‡ with details!

Sid
@sidgairo18.bsky.social

πŸš€New preprint: DAVE β€” Distribution-aware Attribution via ViT Gradient DEcomposition. 1/11 πŸ” What’s new: We fix a persistent issue in ViT explainability: unstable, artifact-heavy pixel attributions. DAVE yields fine-grained pixel-level maps without patch-grid saliency.

View full thread
Sid
@sidgairo18.bsky.social

10/11 πŸ–ΌοΈ Qualitative comparison Vs prior methods, DAVE gives sharper, more object-centric attributions with less background + fewer patch-grid artifacts.

Sid
@sidgairo18.bsky.social

11/11 This is joint work with Adam WrΓ³bel (project lead), Jacek Tabor, Bernt Schiele, Bartosz ZieliΕ„ski, and Dawid Rymarczyk. πŸ“°Preprint: arxiv.org/abs/2602.06613 πŸ’»Code: github.com/a-vrobell/DAVE [to be released soon]

arxiv.orgDAVE: Distribution-aware Attribution via ViT Gradient DecompositionVision Transformers (ViTs) have become a dominant architecture in computer vision, yet producing stable and high-resolution attribution maps for these models remains challenging. Architectural compone...
Sid
@sidgairo18.bsky.social

πŸš€New preprint: DAVE β€” Distribution-aware Attribution via ViT Gradient DEcomposition. 1/11 πŸ” What’s new: We fix a persistent issue in ViT explainability: unstable, artifact-heavy pixel attributions. DAVE yields fine-grained pixel-level maps without patch-grid saliency.

View full thread
Sid
@sidgairo18.bsky.social

9/11 πŸ” Beyond standard ViTs On inherently interpretable B-cos ViTs, DAVE yields sharper, more object-aligned maps and improves localization vs built-in B-cos explanations. B-cos ViTs produce sharp attributions only when trained with a conv-stem, DAVE fixes this reliance.

Sid
@sidgairo18.bsky.social

πŸš€New preprint: DAVE β€” Distribution-aware Attribution via ViT Gradient DEcomposition. 1/11 πŸ” What’s new: We fix a persistent issue in ViT explainability: unstable, artifact-heavy pixel attributions. DAVE yields fine-grained pixel-level maps without patch-grid saliency.

View full thread
Sid
@sidgairo18.bsky.social

6/11 🌫️ Stabilize (low-pass) DAVE adds low-pass stabilization by averaging the equivariant effective transformation under small input perturbations (Gaussian smoothing in expectation). This removes components unstable to tiny input changes. (See above fig., last column.)

Sid
@sidgairo18.bsky.social

7/11 πŸ”§ The pipeline Sample small spatial transforms + noise, compute effective transformation (conditioned forward blocks gradients through conditioning), inverse-transform & average, then apply element-wise to the input.

Sid
@sidgairo18.bsky.social

πŸš€New preprint: DAVE β€” Distribution-aware Attribution via ViT Gradient DEcomposition. 1/11 πŸ” What’s new: We fix a persistent issue in ViT explainability: unstable, artifact-heavy pixel attributions. DAVE yields fine-grained pixel-level maps without patch-grid saliency.

View full thread
Sid
@sidgairo18.bsky.social

4/11 ⚠️ Why gradients break Operator variation can amplify tiny perturbations β†’ high-frequency junk in attributions. DAVE drops this term and keeps the effective transformation as a cleaner attribution operator.

Sid
@sidgairo18.bsky.social

5/11 🧼 Remove grid artifacts Even the effective transformation can carry architecture-induced grid patterns. DAVE filters them by enforcing local equivariance: under small spatial transforms, the attribution must transform consistently.

Sid
@sidgairo18.bsky.social

πŸš€New preprint: DAVE β€” Distribution-aware Attribution via ViT Gradient DEcomposition. 1/11 πŸ” What’s new: We fix a persistent issue in ViT explainability: unstable, artifact-heavy pixel attributions. DAVE yields fine-grained pixel-level maps without patch-grid saliency.

Sid
@sidgairo18.bsky.social

2/11 🧩 The problem ViT components (patch embedding, attention routing, etc.) inject structured artifacts into gradients β†’ explanations become noisy/unstable, or methods fall back to coarse patch-level maps.

Sid
@sidgairo18.bsky.social

3/11 πŸ’‘ Key idea Model each ViT layer as an input-dependent linear operator L(X) applied to X. Then the input-gradient decomposes into: (1) effective transformation L(X) (2) operator variation (how L changes w.r.t. X)

Sid
@sidgairo18.bsky.social

Heading to #ICCV2025? Don’t miss #CV4DC β€” Oct 19, Honolulu 🌺 An in-person workshop spotlighting research for/with developing regions. See exciting keynote talks from leading researches! See accepted papers & posters here: cv4dc.github.io/2025/ #AI4Good

Sid
@sidgairo18.bsky.social

🌟Here are some resources and notes about doing research that I’ve been compiling since I started my PhD.🌟 I rely on these often, and after sharing them with a few folks in my group who found them useful, I thought it could be of use to the broader community πŸ§΅πŸ‘‡ (1/n)

View full thread
Sid
@sidgairo18.bsky.social

3️⃣ How to Write Academic Papers A practical checklist-driven guide on writing with clarity, rigor, and reproducibility. Inspired by ICML's best practices and more. πŸ”—https://sidgairo18.github.io/how_to_write_academic_papers.html (4/n)

Sid
@sidgairo18.bsky.social

🌟Here are some resources and notes about doing research that I’ve been compiling since I started my PhD.🌟 I rely on these often, and after sharing them with a few folks in my group who found them useful, I thought it could be of use to the broader community πŸ§΅πŸ‘‡ (1/n)

Sid
@sidgairo18.bsky.social

1️⃣ How to Do Research Mindset, habits, tools, writing, productivity, and advice. πŸ”—https://sidgairo18.github.io/notes_and_resources_on_how_to_do_research.html (2/n)

Sid
@sidgairo18.bsky.social

2️⃣ How to Review Scientific Papers What makes a good review? This guide compiles best practices from ICML, ICLR, CVPR, and other leading conferences. πŸ”—https://sidgairo18.github.io/how_to_review_scientific_papers.html (3/n)

Sid
@sidgairo18.bsky.social

⏳Still need to wait for your last experiment results? πŸ“£ We're pleased to announce that the deadline for non-proceeding track #CV4DC at @iccv.bsky.social has been extended to August 15, 2025 Looking forward to your submissions! cv4dc.github.io/2025/

Sid@sidgairo18.bsky.social

πŸ“£ Proceeding track's results are out. πŸŽ‰ Congratulations to all the authors whose papers were accepted. We can't wait to meet you at @iccv.bsky.social in Hawaii on Oct 19th. ⏰ Our non-proceeding track is still accepting submissions until July 20th! Details in the comments

Sid
@sidgairo18.bsky.social

πŸ“£ Proceeding track's results are out. πŸŽ‰ Congratulations to all the authors whose papers were accepted. We can't wait to meet you at @iccv.bsky.social in Hawaii on Oct 19th. ⏰ Our non-proceeding track is still accepting submissions until July 20th! Details in the comments

Sid
@sidgairo18.bsky.social

🚨More tiny experiments that seem benign: linear probing on frozen features (but again not explicitly making BatchNorm params to .eval) can have big implications. Best test accuracy: 71.20% at Epoch 10, for Learning Rate = 0.01 Makes a hell of a lot of difference! Almost 10% 🚨

Sid@sidgairo18.bsky.social

Seemingly innocuous error that I see students making time and again: model.eval() or param.eval() for models with norm layers (like batch-norm) is critical. Especially when freezing ❄️ certain layers and training rest.

Sid
@sidgairo18.bsky.social

πŸ”—Try for yourself πŸ˜ƒ: github.com/sidgairo18/p...

github.comGitHub - sidgairo18/probing_pretrained_models: This is a repo for basic ImageNet1K running validation eval. Additionally, also to train a classifier on top of frozen ImageNet1k features on downstream ...This is a repo for basic ImageNet1K running validation eval. Additionally, also to train a classifier on top of frozen ImageNet1k features on downstream datasets - CIFAR100. - sidgairo18/probing_pr...
Sid
@sidgairo18.bsky.social

Seemingly innocuous error that I see students making time and again: model.eval() or param.eval() for models with norm layers (like batch-norm) is critical. Especially when freezing ❄️ certain layers and training rest.

Sid
@sidgairo18.bsky.social

Had my older MTB catching dust in the bike room, took 2 hours to replace both tyres and tubes with brand new ones, new brakes, and chain cleaned + greased. πŸ’₯ Pretty relaxing and a nice rejuvinating break from research✨ P.S.: Once you are used to driving road bikes, MTBs seem too slow πŸ€ͺ

Sid
@sidgairo18.bsky.social

Wow 🀯

Big Joel@bigjoel.bsky.social

So I have now read the Big Three Ayn Rand Books, atlas shrugged, the fountainhead, and anthem. And I have this to say: she’s bad at writing books, so bad, in fact, that I find her popularity genuinely strange

Sid
@sidgairo18.bsky.social

Had a great time presenting our work at ICLR 2025 in Singapore πŸ‡ΈπŸ‡¬. Lot of great interactions, meeting amazing people and a very well organised conference 🌟. It was great to catch up with old colleagues and meet new folks. For more detail see: arxiv.org/abs/2503.00641

Sid@sidgairo18.bsky.social

πŸš€ I will be @iclr_conf in Singapore next week presenting our recent work. We answer how exactly does "What you explain depend on how you train". πŸ€“ We will be @ Hall 3 + Hall 2B #492 (26th April, 3PM - 5:30 PM) Please reach out via dm if interestd in a chat or β˜•οΈ #ICLR2025

Sid
@sidgairo18.bsky.social

πŸš€ I will be @iclr_conf in Singapore next week presenting our recent work. We answer how exactly does "What you explain depend on how you train". πŸ€“ We will be @ Hall 3 + Hall 2B #492 (26th April, 3PM - 5:30 PM) Please reach out via dm if interestd in a chat or β˜•οΈ #ICLR2025

Sid reposted

Serge Belongie
@serge.belongie.com

So far, 2,135 people have responded to the poll SΓΈren and I posted a few days ago. Of those, 94.4% replied β€œYes” to being interested in officially presenting accepted @neuripsconf.bsky.social papers in Europe. (1/7)

Serge Belongie@serge.belongie.com

Would you present your next NeurIPS paper in Europe instead of traveling to San Diego (US) if this was an option? SΓΈren Hauberg (DTU) and I would love to hear the answer through this poll: (1/6)

docs.google.comNeurIPS participation in EuropeWe seek to understand if there is interest in being able to attend NeurIPS in Europe, i.e. without travelling to San Diego, US. In the following, assume that it is possible to present accepted papers ...
Sid
@sidgairo18.bsky.social

With systems like PaperBench -the reproducibility tracks at conferences will now also have AI participants πŸ˜ƒπŸ˜‰ Infact, (soonish) the reviewing process could be supported with pipelines that can authenticate claims and results made in papers. πŸ€“ openai.com/index/paperb...

openai.comPaperBench: Evaluating AI’s Ability to Replicate AI ResearchWe introduce PaperBench, a benchmark evaluating the ability of AI agents to replicate state-of-the-art AI research.

Sid reposted

Serge Belongie
@serge.belongie.com

Would you present your next NeurIPS paper in Europe instead of traveling to San Diego (US) if this was an option? SΓΈren Hauberg (DTU) and I would love to hear the answer through this poll: (1/6)

docs.google.comNeurIPS participation in EuropeWe seek to understand if there is interest in being able to attend NeurIPS in Europe, i.e. without travelling to San Diego, US. In the following, assume that it is possible to present accepted papers ...