Signal or Noise? Auditing Rotation-Induced Saliency Drift in Medical and Aerial Imaging
Khawaja Murad ul Hassan, Mehran Ebrahimi
cs.CV · cs.AI
Quick verdict
Claim: Grad-CAM rotation drift is noise the classifier discards; aligned multi-view averaging restores equivariance training-free. Evidence strong; faithfulness thin.
Summary
The paper asks whether Grad-CAM's drift under input rotation is faithful signal (the network genuinely attending elsewhere) or operator noise. It decomposes the operator stage-wise on ImageNet-1K and finds channel weights are the most rotation-stable stage (exactly so on ResNet-50, where a GAP+linear head makes the class gradient field spatially constant), while spatial activations carry the drift that the classifier's pooling discards. A causal occlusion test supports this: blurring the pixels whose saliency drifts costs less probability than blurring random pixels at either orientation. The proposed fix, EquiGrad-CAM, is a training-free wrapper computing Grad-CAM over 18 rotated views, inverse-rotating each map into a canonical frame, then averaging. It raises equivariance 0.703 to 0.956 (ResNet-50), 0.511 to 0.958 (VGG-16), 0.250 to 0.867 (ViT-B/16) on the full ImageNet validation set, 0.317 to 0.777 on zero-shot CLIP, and works without retraining on PatchCamelyon and RESISC45. Its by-product PEUM ranks unstable explanations for auditor triage. Limitations: faithfulness results are descriptive with small insertion costs on CNNs, main tables are single-seed, and no CAM method localizes PatchCamelyon's label-defining region well.
Review complete
Assessment at a glance
5
risk flags identified
● high 0
● medium 3
● low 2
5 strengths · 5 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- Grad-CAM's rotation-induced saliency drift is mostly operator noise: it is carried by degrees of freedom the classifier discards (exact on ResNet-50, where the GAP+linear head makes channel weights orientation-invariant).
- A causal occlusion test shows the drifting pixels matter less to the model than random pixels at either orientation (0.100/0.038 vs 0.152/0.121 on ResNet-50, n=1,000).
- EquiGrad-CAM, a training-free wrapper of aligned multi-view averaging, raises equivariance on the full ImageNet validation set from 0.703 to 0.956 (ResNet-50), 0.511 to 0.958 (VGG-16) and 0.250 to 0.867 (ViT-B/16), and 0.317 to 0.777 on zero-shot CLIP.
- Canonical-frame alignment — not where the averaging happens — is the mechanism: unaligned averaging falls to or below single-view Grad-CAM on both CNN backbones (Table 6).
- PEUM, a by-product uncertainty score, surfaces 4.4x (aerial) and 2.8x (histopathology) more rotation-unstable explanations than random inspection at a 10% budget, while not predicting model error.
Strengths
Mechanistic operator decomposition rather than output-level inference
Table 1 measures equivariance at every Grad-CAM stage (activations 0.321, gradient field, channel weights 1.000 on ResNet-50, final map 0.700) and validates against published Eq (0.703) to within 0.003; the ResNet-50 exactness of alpha is derived from the GAP+linear structure (gradient spatial std exactly 0 in all 2,048 channels) and directly verified, not merely observed.
Causal, not just correlational, evidence for the signal-vs-noise question
Section 3(5): occluding the drift region costs 0.100/0.038 probability (canonical/rotated) versus 0.152/0.121 for random and 0.264/0.225 for the agreement region (ResNet-50, n=1,000); the orientation-asymmetry null is reported even though the test construction is tilted toward finding faithful signal.
Scale-matched ablation that cleanly isolates the mechanism
Unusually honest reporting practices
Breadth of evaluation with a practical by-product
Weaknesses
The central faithfulness framing is under-supported by the faithfulness numbers themselves
Large relative gains ride on low absolute baselines
Statistical support for the means is uneven
Non-trivial compute overhead for an auditing tool
Novelty claim rests on the authors' own literature positioning
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
This paper (arXiv:2609.02224) asks whether the drift of Grad-CAM saliency maps under input rotation is *faithful signal* (the network genuinely attending elsewhere on a partly out-of-distribution input) or *operator noise*. It answers in three moves: (1) a stage-wise decomposition of the Grad-CAM operator (Table 1), (2) a causal occlusion test (§3, item 5), and (3) a training-free remedy, EquiGrad-CAM, evaluated at ImageNet scale (Table 5) and in two rotation-natural domains, PatchCamelyon and RESISC45 (Table 2), plus zero-shot CLIP (Table 4).
The headline claims are well separated from their support. The signal-vs-noise verdict is stated at calibrated strength (§3 conclusion, Limitations): the drift is *mostly* noise, exact on ResNet-50 where the GAP+linear head makes the class gradient field spatially constant in all 2,048 channels (spatial std exactly 0), empirical on VGG-16 and ViT. The mechanism claim — channel weights are the most stable stage (alpha Pearson 1.000/0.889/0.685) while spatial activations move (0.321/0.246/0.453) — is verified structurally, not just observed.
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 5 strengths & 5 weaknesses, analyzed in detail
- 5 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What is EquiGrad-CAM and how does it differ from ordinary Grad-CAM?
EquiGrad-CAM is a training-free wrapper around Grad-CAM. It computes Grad-CAM on T rotated copies of the input (18 by default), inverse-rotates each resulting saliency map back into the original orientation, and averages them. The alignment step is the key difference: averaging maps without aligning them first can perform worse than single-view Grad-CAM.
Does EquiGrad-CAM require retraining the model?
No. It wraps any pretrained network that Grad-CAM already supports (CNNs and ViTs, including zero-shot CLIP) with no weight changes. The cost is extra compute: about 19-20x a single Grad-CAM pass at 18 views, or about 6-7x at the recommended 6-view setting.
How large are the reported equivariance improvements?
On the full ImageNet-1K validation set, Pearson equivariance rises from 0.703 to 0.956 on ResNet-50 (+36%), from 0.511 to 0.958 on VGG-16 (+87.5%), and from 0.250 to 0.867 on ViT-B/16 (+247%); on zero-shot CLIP on aerial images it rises from 0.317 to 0.777 (+145%). These percentages come from low baselines on the weaker backbones, so the absolute numbers matter most.
Does a more rotation-consistent saliency map mean a more correct one?
Not necessarily. The paper's own label-region check shows that no CAM method it tested — EquiGrad-CAM included — concentrates strongly on the region that defines the PatchCamelyon label. The method makes explanations reproducible under rotation and flags unstable ones for human review; it does not certify that the explanation is correct.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Figures and the supplement (e.g., the angle-resolved plot, cost table, PEUM scatter) are not part of the extracted text; all conclusions above rest on in-text tables and numbers, which appear complete for the main claims.
- The released code repository was not inspected; reproducibility assessment is based on the protocol description in the paper only.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.