Research map / Architectures and efficiency / RNNs and normalization

Comprehensive Information Bottleneck for Unveiling Universal Attribution to Interpret Vision Transformers

CVPR 2025 · Architectures and efficiency › RNNs and normalization · approach: Representation geometry and collapse · 2 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes CoIBA, an information-bottleneck attribution method that shares one learnable universal damping ratio across multiple ViT layers and upper-bounds layer-wise mutual information with a single first-layer term, yielding more faithful attribution maps for vision transformers.

Problem

Layer-specific information-bottleneck attribution methods (IBA, InputIBA) capture only part of the distributed decision evidence across ViT layers, and no single layer gives the most faithful attribution, requiring heuristic layer search.

Approach

Insert bottleneck layers into multiple targeted layers sharing one universal damping ratio (uniform across channels per token), and replace the linear sum of layer-wise mutual information terms with a variational upper bound involving only the first layer's mutual information, removing per-layer trade-off tuning.

Key results

• CoIBA achieves the best insertion/deletion among compared methods across ViT, DeiT, and Swin variants; e.g., ViT-B-16/224 insertion 16.63 vs 20.99 for IBA (deletion lower is better; insertion column shown is deletion) and ViT-B insertion 13.01/62.58 vs 17.23/59.06 for IBA.
• On ROAD, CoIBA gives ViT-B-16/224 MoRF/LeRF of 16.63/73.68 vs 20.99/68.80 for IBA.
• Ablation shows that the universal damping ratio and variational upper bound each improve faithfulness over individual per-layer ratios and linear mutual information combination.

Measured effects reported in the paper

Surprising findings

• Iterating IBA across layers and linearly combining attributions (IBA*) improves performance only on high-confidence samples, not low-confidence ones.

Limitations

• Layer-wise trade-off tuning and heuristic search are replaced by a single hyperparameter beta, but beta and the start/end layers (4 and 12 for ViT-B) are still selected by the authors.
• Evaluations are approximations since no ground truth exists for most benchmarks; FunnyBirds is the only ground-truth-based test.

Experimental setup

Datasets: ImageNet-1k ImageNet-A ImageNet-R FunnyBirds

Benchmarks: FunnyBirds ImageNet-1k insertion/deletion ROAD ImageNet-A ImageNet-R

Metrics: insertion deletion ROAD MoRF ROAD LeRF completeness correctness contrastivity accuracy background independence CKA

Baselines: Chefer-LRP Generic IIA ViT-CX IBA Beyond InputIBA

Builds on: IBA InputIBA Deep Variational Information Bottleneck Chefer-LRP

Closest papers by meaning

• One Wave To Explain Them All: A Unifying Perspective On Feature Attribution (ICML 2025 · similarity 0.87)
• Benchmarking the Attribution Quality of Vision Models (NeurIPS 2024 · similarity 0.87)
• Diffusion Models for Attribution (AAAI 2025 · similarity 0.86)
• LibraGrad: Balancing Gradient Flow for Universally Better Vision Transformer Attributions (CVPR 2025 · similarity 0.86)
• DePass: Unified Feature Attributing by Simple Decomposed Forward Pass (NeurIPS 2025 · similarity 0.85)
• A Comprehensive Information-Decomposition Analysis of Large Vision-Language Models (ICLR 2026 · similarity 0.85)