Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

What Helps -- and What Hurts: Bidirectional Explanations for Vision Transformers

Qin Su, Tie Luo

TL;DR

BiCAM is proposed, a bidirectional class activation mapping method that captures both supportive and suppressive contributions to model predictions and introduces a Positive-to-Negative Ratio (PNR) that summarizes attribution balance and enables lightweight detection of adversarial examples without retraining.

Abstract

Vision Transformers (ViTs) achieve strong performance in visual recognition, yet their decision-making remains difficult to interpret. We propose BiCAM, a bidirectional class activation mapping method that captures both supportive (positive) and suppressive (negative) contributions to model predictions. Unlike prior CAM-based approaches that discard negative signals, BiCAM preserves signed attributions to produce more complete and contrastive explanations. BiCAM further introduces a Positive-to-Negative Ratio (PNR) that summarizes attribution balance and enables lightweight detection of adversarial examples without retraining. Across ImageNet, VOC, and COCO, BiCAM improves localization and faithfulness while remaining computationally efficient. It generalizes to multiple ViT variants, including DeiT and Swin. These results suggest the importance of modeling both supportive and suppressive evidence for interpreting transformer-based vision models.

What Helps -- and What Hurts: Bidirectional Explanations for Vision Transformers

TL;DR

BiCAM is proposed, a bidirectional class activation mapping method that captures both supportive and suppressive contributions to model predictions and introduces a Positive-to-Negative Ratio (PNR) that summarizes attribution balance and enables lightweight detection of adversarial examples without retraining.

Abstract

Vision Transformers (ViTs) achieve strong performance in visual recognition, yet their decision-making remains difficult to interpret. We propose BiCAM, a bidirectional class activation mapping method that captures both supportive (positive) and suppressive (negative) contributions to model predictions. Unlike prior CAM-based approaches that discard negative signals, BiCAM preserves signed attributions to produce more complete and contrastive explanations. BiCAM further introduces a Positive-to-Negative Ratio (PNR) that summarizes attribution balance and enables lightweight detection of adversarial examples without retraining. Across ImageNet, VOC, and COCO, BiCAM improves localization and faithfulness while remaining computationally efficient. It generalizes to multiple ViT variants, including DeiT and Swin. These results suggest the importance of modeling both supportive and suppressive evidence for interpreting transformer-based vision models.
Paper Structure (16 sections, 4 equations, 4 figures, 5 tables, 1 algorithm)

This paper contains 16 sections, 4 equations, 4 figures, 5 tables, 1 algorithm.

Table of Contents

  1. Introduction
  2. Related Work
  3. Methodology
  4. Strategic Layer Aggregation
  5. BiCAM Attribution Mechanism
  6. Positive-to-Negative Ratio (PNR)
  7. Evaluation
  8. Experimental Setup
  9. Qualitative Results
  10. Localization Performance
  11. Faithfulness via Feature Removal
  12. PNR Analysis
  13. Computational Efficiency
  14. Ablation Studies
  15. Generalization to ViT Variants
  16. ...and 1 more sections

Figures (4)

  • Figure 1: BiCAM applied to ViT-B/16 on COCO. A query of "tusker" yields supportive (red) and suppressive (blue) attributions on the elephant and the zebra, respectively; yet querying "zebra" reverses the pattern. Baseline methods do not consistently produce this contrastive explanation.
  • Figure 2: Single-class attribution on ImageNet (grasshopper class; ViT-B/16). BiCAM highlights the target in red and background in blue.
  • Figure 4: Faithfulness evaluation via perturbation on VOC and COCO (ViT-B/16). Curves show the mean across five random seeds. Higher faithfulness is indicated by steeper accuracy degradation in MIF removal and flatter curves in LIF removal.
  • Figure 5: BiCAM generalizes to ViT variants, tested on ImageNet. All models adopt the base version. Largely coherent bidirectional attributions are observed.