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

Qin Su; Tie Luo

cs.CVMar 2, 2026

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

Authors: Qin Su, Tie Luo

Organizations: Department of Electrical and Computer Engineering University of Kentucky, Lexington, KY, USA

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.

Figures & tables

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.

Panel of Figure 2 — Figure 2: Single-class attribution on ImageNet (grasshopper class; ViT-B/16). BiCAM highlights the target in red and background in blue.

ImageNet-1k
Method	Pix. Acc.	IoU	F1	Prec.	Rec.
Attn Rollout	0.6209	0.3597	0.4893	0.7326	0.4657
LRP-based	0.5863	0.2029	0.3055	0.9110	0.2176
AGCAM	0.7341	0.5212	0.6515	0.8299	0.6276
BiCAM	0.6253	0.5419	0.6624	0.5901	0.9288
VOC 2012

Table 1: Localization results on 3 datasets. ViT-Shapley is excluded because its amortized explainer could not scale to 1,000-class evaluation.

ImageNet
	Attn Rollout	LRP	AGCAM	BiCAM
LIF $\uparrow$	0.4739	0.5140	0.5298	0.5478
MIF $\downarrow$	0.2053	0.1736	0.1607	0.1654
Faith $\uparrow$	0.2685	0.3404	0.3691	0.3824
VOC 2-class
LIF $\uparrow$	0.8623	0.9231	0.9280	0.9313

Table 2: Faithfulness scores. ViT-Shapley is excluded because it achieves only 0.294 on the 10-class ImageNette.

Panel of Figure 4 — 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.

Attack	$\Delta$ PNR	std	AUROC	AUPR	Threshold
Clean	0.00	N/A	-	-	-
PGD	+0.57	0.31	0.781	0.749	0.248
C&W	+0.79	0.37	0.842	0.808	0.332
MI-FGSM	+0.46	0.27	0.764	0.731	0.219
Avg.	+0.61	0.32	0.796	0.763	0.266

Table 3: PNR-based adversarial detection on VOC 2012 multi-object scenes.

Method	ms/img	MB/img	Train time
Attn Rollout	23.3	0.15	None
LRP	134.6	97.2	None
ViT-Shapley	10.1 ∗	–	19 hrs ∗
AGCAM	20.0	0.22	None
BiCAM	16.0	0.24	None

Table 4: Runtime comparison (ViT-B/16, RTX 4090, ImageNet-1k).

	Layer window $\ell$ ( $T{=}2$ )				Temperature $T$ ( $\ell{=}8$ )
	$\ell{=}12$	$\ell{=}8$	$\ell{=}4$	$\ell{=}1$	$T{=}1$	$\mathbf{T{=}2}$	$T{=}3$
Pix.Acc	0.8505	0.8559	0.8498	0.8413	0.8402	0.8559	0.8492
IoU	0.2840	0.3700	0.2093	0.1491	0.1051	0.3700	0.2060
F1	0.4043	0.5104	0.3173	0.2277	0.1711	0.5104	0.3140
Prec.	0.5980	0.6095	0.6765	0.5343	0.6401	0.6095	0.6789
Rec.	0.3949	0.5863	0.2649	0.1824	0.1144	0.5863	0.2589

Table 5: Ablation on layer window

\ell

and temperature

T

(ViT-B/16, VOC 2012).

Figure 5: BiCAM generalizes to ViT variants, tested on ImageNet. All models adopt the base version. Largely coherent bidirectional attributions are observed.

Explore similar work

Grad-CAM for Vision Transformers: A Systematic Taxonomy and Audit of Methodological Ambiguity in Explainable AI
Aug 5, 2026Casey Wall, Longwei Wang, Rodrigue Rizk +1Gradient-Based Attribution Transformer Interpretability
USD Artificial Intelligence Research, Department of Computer Science, University of South Dakota
Learn to Rank: Visual Attribution by Learning Importance Ranking
Apr 7, 2026David Schinagl, Christian Fruhwirth-Reisinger, Alexander Prutsch +2Feature Attribution Explainable Image Classification
Institute of Visual Computing (IVC), Graz University of Technology, Graz, Austria · Nomadic AI, USA · Amazon, USA
Decision-Aware Attention Propagation for Vision Transformer Explainability
Apr 20, 2026Sehyeong Jo, Gangjae Jang, Haesol ParkTransformer Interpretability Vision Transformer
Aarhus University · University of Colorado Boulder · KIST

Aug 5, 2026J/K move · Enter open · S save

Grad-CAM for Vision Transformers: A Systematic Taxonomy and Audit of Methodological Ambiguity in Explainable AI

Casey Wall, Longwei Wang, Rodrigue Rizk, KC Santosh

USD Artificial Intelligence Research, Department of Computer Science, University of South Dakota

Gradient-Based Attribution Transformer Interpretability Vision Transformer

Open paper AskPDF arXiv

Gradient-weighted Class Activation Mapping (Grad-CAM) is widely used to visualize model decisions, but it was originally formulated for convolutional neural networks, where spatial feature maps and channel dimensions have clear architectural meanings. Vision Transformers (ViTs) do not provide the same structure, instead representing images through tokens, attention, residual streams, and multimodal interactions. This paper presents a systematic taxonomy and literature audit of how Grad-CAM and related methods are adapted, justified, and reported for ViT-based architectures. From an initial search of more than 550 papers, we identify 175 papers that apply Grad-CAM or Grad-CAM-adjacent methods to ViTs. We find that most papers do not provide a full mathematical or implementation-level account of how Grad-CAM is adapted to transformer representations. To characterize this gap, we introduce a descriptive taxonomy of ViT Grad-CAM adaptations that makes explicit the feature locations, gradient targets, spatial reconstruction steps, and aggregation choices that are often left implicit. This taxonomy is not intended to prescribe a single correct adaptation, but to clarify the range of methodological choices being made. The study shows that Grad-CAM on ViTs is often treated as a trivial extension of CNN-based Grad-CAM, despite requiring nontrivial choices that affect rigor, reproducibility, and interpretation.

Abstract

Explore similar work

Grad-CAM for Vision Transformers: A Systematic Taxonomy and Audit of Methodological Ambiguity in Explainable AI

Learn to Rank: Visual Attribution by Learning Importance Ranking

Decision-Aware Attention Propagation for Vision Transformer Explainability