Sid

DAVE: Distribution-aware Attribution via ViT Gradient Decomposition Vision 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...

11/11

This is joint work with Adam Wróbel (project lead), Jacek Tabor, Bernt Schiele, Bartosz Zieliński, and Dawid Rymarczyk.

📰Preprint: arxiv.org/abs/2602.06613
💻Code: github.com/a-vrobell/DAVE [to be released soon]

10/11 🖼️ Qualitative comparison

Vs prior methods, DAVE gives sharper, more object-centric attributions with less background + fewer patch-grid artifacts.

9/11 🔁 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.

8/11 📈 Main results

Consistent improvements on localization metrics and perturbation evaluations.

7/11 🔧 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.

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.)

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.

4/11 ⚠️ Why gradients break

Operator variation can amplify tiny perturbations → high-frequency junk in attributions.

DAVE drops this term and keeps the effective transformation as a cleaner attribution operator.

3/11 💡 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)

2/11 🧩 The problem

ViT components (patch embedding, attention routing, etc.) inject structured artifacts into gradients → explanations become noisy/unstable, or methods fall back to coarse patch-level maps.

🚀New preprint: DAVE — Distribution-aware Attribution via ViT Gradient DEcomposition.

1/11 🔍 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.

ICCV 2025 🌺 Aloha from Hawaii! MPI-INF (D2) is presenting 4 papers this year (one Highlight). Thread 👇

Heading to #ICCV2025? Don’t miss #CV4DC — 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

The Great Depression, Recession and Stagnation in Full Historical Context

Source: www.cold-takes.com/the-great-st...

GitHub - sidgairo18/sidgairo18.github.io Contribute to sidgairo18/sidgairo18.github.io development by creating an account on GitHub.

Hope these help you as much as they’ve helped me.
📚Feel free to share, bookmark, and contribute here (github.com/sidgairo18/s...)!

P.S.: Please feel free to share relevant resources in the comments / thread and I'll add them as well 😀 (n/n)

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.
🔗https://sidgairo18.github.io/how_to_write_academic_papers.html (4/n)

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

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

🌟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 🧵👇 (1/n)

You can do more than you think: Practical principles for self-directed growth - Blog - Recurse Center How to develop your capacity for self-direction

www.recurse.com/blog/185-do-...

⏳Still need to wait for your last experiment results?
📣 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/

#ICML2025 Stats

Papers accepted at ICML 2025 from the Computer Vision and Machine Learning Department at the Max Planck Institute for Informatics.

Papers being presented from our group at #ICML2025!

Congratulations to all the authors! To know more, visit us in the poster sessions!

A 🧵with more details:

@icmlconf.bsky.social @mpi-inf.mpg.de

Submit your papers to our non-proceeding track here: cv4dc.github.io/2025/call-fo...

📣 Proceeding track's results are out.
🎉 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

Cc: @mpi-inf.mpg.de

Porto, Portugal 🇵🇹 June-July 2025

One of the prettiest and liveliest cities I’ve traveled to ✨ 🚤

Congratulations Matt! 🔥

GitHub - 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...

🔗Try for yourself 😃: github.com/sidgairo18/p...