🤖 AI

DAVE: Distribution-aware Attribution via ViT Gradient Decomposition

The paper introduces DAVE, a mathematically grounded attribution method for Vision Transformers that decomposes input gradients to isolate stable, locally equivariant components and eliminate architectural artifacts, thereby enabling the generation of high-resolution and stable pixel-level explanations.

Original authors: Adam Wróbel, Siddhartha Gairola, Jacek Tabor, Bernt Schiele, Bartosz Zieliński, Dawid Rymarczyk

Published 2026-02-09

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Adam Wróbel, Siddhartha Gairola, Jacek Tabor, Bernt Schiele, Bartosz Zieliński, Dawid Rymarczyk

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you have a very smart robot (a Vision Transformer) that looks at a picture and says, "That's a soccer ball!" But if you ask the robot, "Why did you think that?" it gives you a confusing answer. It might point to the whole picture vaguely, or worse, it might point to a weird grid pattern that only exists because of how the robot was built, not because of the ball itself.

This paper introduces a new tool called DAVE (Distribution-aware Attribution via ViT Gradient Decomposition) to fix this confusion. Think of DAVE as a high-resolution, noise-canceling flashlight that helps us see exactly what the robot is actually looking at.

Here is how it works, broken down into simple concepts:

1. The Problem: The "Static" on the Radio

When current AI models try to explain their decisions, they often produce "static" or "artifacts."

The Analogy: Imagine listening to a radio station, but there's a constant hum and a weird grid pattern overlaying the music. You can hear the song (the prediction), but the static (the explanation) makes it hard to tell which instrument is playing which note.
The Reality: In Vision Transformers, the way the model processes images (breaking them into patches and using "attention") creates these artificial patterns. Existing methods often get distracted by this static, pointing to the grid lines instead of the actual object.

2. The Solution: Separating the Signal from the Noise

DAVE works by mathematically taking apart the robot's "thought process" (the gradient) into two distinct parts:

Part A: The Real Signal (Effective Transformation): This is the part of the robot's brain that actually changes based on what is in the picture. If you move the soccer ball, this part moves with it. This is the useful information.
Part B: The Robot's Quirks (Operator Variation): This is the part that changes just because the robot's internal gears shifted slightly, even if the picture didn't. This is the "noise" or the "grid pattern."

The Metaphor: Imagine a chef tasting a soup.

The Signal: The chef says, "It tastes salty because of the salt." (This is the real reason).
The Noise: The chef says, "It tastes salty because the spoon I'm holding is vibrating." (This is an artifact of the tool, not the soup).
DAVE's Job: It filters out the vibrating spoon and tells you, "It's the salt."

3. The Magic Trick: The "Steady Hand" Filter

Once DAVE separates the real signal from the robot's quirks, it does something clever called Equivariant Filtering.

The Analogy: Imagine you are trying to find a specific face in a crowd. If you tilt your head slightly, the face is still the same face. If you shift your eyes slightly, the face is still there.
How DAVE uses this: DAVE asks the model, "If I rotate this picture a tiny bit or shift it a few pixels, does your explanation change?"
- If the explanation jumps around wildly, DAVE knows that's unstable "noise" and throws it away.
- If the explanation stays consistent (like the face staying the same), DAVE keeps it.
The Result: This creates a smooth, stable map that highlights the actual object (like the soccer ball) without the jagged, grid-like artifacts.

4. The Final Polish: Blurring the Rough Edges

Finally, DAVE applies a gentle "blur" (low-pass filtering) to the result.

The Analogy: Think of looking at a high-definition photo through a slightly foggy window. The fog removes the tiny, distracting specks of dust (high-frequency noise) but keeps the main picture clear.
The Result: The final map is smooth, precise, and highlights exactly the pixels the model used to make its decision.

What Did They Find?

The authors tested DAVE on many different AI models (like DeiT and DINO) using a massive database of images (ImageNet).

Better Accuracy: When they checked if the AI's explanation actually pointed to the object, DAVE was much better than previous methods. It found the object more often.
Less Noise: The visual maps produced by DAVE didn't have those annoying grid patterns. They looked like natural, human-understandable highlights.
Stability: If you slightly changed the picture (rotated it or shifted it), DAVE's explanation stayed the same, proving it was looking at the object, not the pixels.

In short: DAVE is a new way to listen to AI models that filters out the "static" caused by the model's own architecture, giving us a clear, stable, and accurate view of what the AI is actually seeing.

Technical Summary: DAVE – Distribution-aware Attribution via ViT Gradient Decomposition

1. Problem Statement

Vision Transformers (ViTs) have achieved state-of-the-art performance in computer vision but present significant challenges for Explainable AI (XAI). Unlike Convolutional Neural Networks (CNNs), ViTs process images as sequences of patch tokens using attention mechanisms. This architecture introduces specific artifacts in attribution maps:

Structured Artifacts: Gradient- and attention-based explanations often exhibit grid-like patterns or "patch-grid" artifacts induced by tokenization and attention routing.
Instability: Existing methods frequently produce unstable pixel-level attributions or are forced to rely on coarse patch-level explanations that lack fine-grained visual evidence.
Sensitivity: Standard gradient-based methods capture both the effective input-output transformation and the variation of the operator with respect to the input. The latter often amplifies high-frequency, locally unstable components, leading to noisy explanations.

The paper posits that obtaining stable, high-resolution, and class-consistent attribution maps for ViTs remains an open challenge, necessitating a method that can disentangle true semantic features from architecture-induced noise.

2. Methodology: DAVE

The authors propose DAVE (Distribution-aware Attribution via ViT Gradient Decomposition), a mathematically grounded framework that decomposes the input gradient to isolate the stable, effective transformation of the model.

2.1. Structured Gradient Decomposition

The core theoretical contribution is the decomposition of the layer derivative. The authors model ViT layers as differentiable operator-valued maps $L(X)$ followed by a bias. The derivative of a layer $F(X)$ is shown to consist of two distinct terms (Equation 3):

Effective Transformation ( $L(X)$ ): The direct, input-conditioned action of the layer on its input. This represents the stable mapping used for prediction.
Operator Variation: The derivative of the operator $L(X)$ with respect to the input. This term captures how the internal mechanisms change under small perturbations and is identified as the source of high-frequency instability and artifacts.

DAVE discards the operator variation term, retaining only the effective transformation as the baseline attribution signal.

2.2. Equivariant Filtering

To address architecture-induced patterns (such as grid artifacts from patch embeddings) that are stable across inputs but uninformative for specific predictions, DAVE applies a Reynolds-inspired operator.

Equivariance Criterion: The method assumes that attribution maps should transform consistently under small spatial transformations (e.g., rotations, shifts) of the input.
Filtering: By averaging the effective transformation over a local neighborhood of spatial transformations (sampled from a probability measure $\nu$ ), DAVE suppresses components that violate local equivariance. This isolates attribution signals that vary consistently with the input's semantic content.

2.3. Low-Pass Stabilization

To further suppress high-frequency artifacts, DAVE applies a low-pass filtering operation. This is achieved by taking the expectation of the equivariant effective transformation under small Gaussian perturbations ( $\epsilon$ ).

Mathematically, this is equivalent to convolving the equivariant transformation with a smoothing kernel $K$ (Equation 7).
Unlike standard SmoothGrad (which averages raw gradients), DAVE averages the effective transformation, thereby stabilizing the attribution without reintroducing the discarded operator-variation artifacts.

2.4. Final Attribution Map

The final DAVE attribution map is computed by applying the distribution-aware effective transformation to the input via element-wise multiplication:
$\text{DAVE}(X) = \tilde{W}^{eq}_L(X) \odot X$
where $\tilde{W}^{eq}_L(X)$ is the expectation of the transformed effective operator over spatial transformations and noise.

3. Key Contributions

Mathematical Framework: A structured decomposition of the input gradient that separates the effective input-output transformation from operator variation, providing a theoretical basis for stable ViT attribution.
DAVE Algorithm: A practical implementation for Vision Transformers that leverages architectural properties to suppress patch- and attention-induced artifacts, producing stable, high-resolution pixel-level attributions.
Versatility: Demonstration that the method works effectively on both fully supervised (e.g., DeiT, ViT) and self-supervised (e.g., DINO) foundation models, as well as inherently interpretable architectures (B-cos networks).

4. Experimental Results

The authors evaluated DAVE on ImageNet-1k using standard ViT backbones (ViT-B/16, DeiT-B/16, DeiT-III-B/16, DINO ViT-B/16) and B-cos models.

4.1. Localization Performance

Grid Pointing Game (GridPG): DAVE outperformed all baselines (including I×G, Integrated Gradients, SmoothGrad, LeGrad, AttnLRP, and Chefer-LRP).
- On DeiT-III-B/16, DAVE achieved 65.76%, an improvement of +11.92 percentage points over the best competing method.
- On DeiT-B/16, it achieved 63.52% (+8.05 p.p. improvement).
Energy Pointing Game (EnergyPG): DAVE achieved the highest scores on DeiT-B/16 (82.23%) and DeiT-III-B/16 (82.43%), and improved significantly on DINO-B/16 (83.38%).
B-cos Models: DAVE improved upon the inherent explanations of B-cos networks, achieving +9.14 p.p. improvement in EnergyPG on B-cos-ViT.

4.2. Faithfulness (Pixel Deletion)

Using pixel deletion curves (removing pixels from least to most important), DAVE exhibited the flattest curves, indicating that the model's prediction confidence remained stable longer as pixels were removed. This suggests DAVE attributions are more consistent with the model's actual decision-making process compared to prior methods.

4.3. Qualitative Analysis

Visual comparisons showed that DAVE produces sharper, object-aligned maps with reduced background noise.
Unlike baselines (e.g., SmoothGrad, IntGrad) which often displayed diffuse responses or patch-grid artifacts, DAVE consistently highlighted semantically meaningful features (e.g., contours, distinctive parts) across various ViT variants.

5. Significance and Claims

The paper claims that DAVE addresses the critical gap in ViT explainability by providing a mechanism to produce stable, high-resolution, and class-consistent attribution maps.

Reliability: By isolating the effective transformation and suppressing operator variation and architecture-induced artifacts, DAVE offers a more reliable mechanism for auditing high-capacity vision models.
Applicability: The method is presented as a tool for practitioners to identify model bias, debug misclassifications, and enhance human-AI collaboration by providing visual evidence that aligns with human perception.
Modesty: The authors acknowledge that the averaging operator introduces higher computational overhead compared to raw gradient methods. They suggest future work could focus on more efficient sampling or automatic selection of transformation groups to reduce this cost.

In summary, DAVE provides a mathematically rigorous approach to disentangle the "signal" of a ViT's prediction from the "noise" of its architectural implementation, resulting in explanations that are both spatially precise and robust.