Signal or Noise? Auditing Rotation-Induced Saliency Drift in Medical and Aerial Imaging

Post-hoc saliency maps such as Grad-CAM are increasingly used to audit why a deployed vision model made a decision, yet the heatmap drifts when the input is rotated, even when the prediction is unchanged. In domains with no canonical orientation, such as histopathology and aerial imagery, this undermines using saliency as evidence. We ask whether that drift is faithful signal or noise introduced by the CAM operator, and answer it by measuring equivariance at every stage of the operator rather than inferring it from the network's output. The instability is not where one would guess: the channel weights are the most rotation-stable stage, and on ResNet-50 exactly stable, because a GAP+linear head makes the class gradient field spatially constant. What moves is the spatial activation tensor, and the classifier's own pooling discards that movement. A causal test confirms the consequence: occluding the pixels whose saliency drifts costs the model less than occluding random pixels, at either orientation. The drift is carried by degrees of freedom the classifier throws away, which is what makes removing it faithful rather than destructive. EquiGrad-CAM is a training-free wrapper that takes T rotated views, inverse-rotates each view's saliency into a common canonical frame, and averages. On the full ImageNet-1K validation set it raises equivariance over single-view Grad-CAM by +36.0% (ResNet-50), +87.5% (VGG-16) and +247% (ViT-B/16); a scale-matched ablation isolates alignment before averaging, not the locus of aggregation, as the driver. It beats rotation-augmented training without retraining, lifts zero-shot CLIP by +145%, and yields rotation-consistent explanations on PatchCamelyon and RESISC45. Its by-product PEUM ranks explanations by how reproducible they are, at no cost beyond the views already taken. Code: https://github.com/Khawaja-Murad/EquiGrad-CAM

HiRA-CAM: Preserving Fine-Grained Spatial Relevance in Gradient-Based Visual Explanations

Deep Learning models can include billions of parameters or more, making it difficult to explain their internal transformations and outputs. However, explainability is increasing in importance due to the use of AI in crucial applications. This paper focuses on the interpretability of convolutional neural networks (CNNs). Building on the popular gradient based method LayerCAM for extracting internal features in CNNs, we propose an improved method named HiRA-CAM, and show that it outperforms both LayerCAM and Grad-CAM on creating useful saliency maps for object classification. The main feature of HiRA-CAM is its adaptive use of activation maps from all the layers of the CNN to arrive at a more focused saliency map.

Revisiting Shape and Texture Reliance with Category-Separability-Calibrated Suppression

Feature-suppression evaluations infer model reliance on shape or texture from the accuracy loss caused by attenuating each type of information. Such losses, however, conflate feature reliance with the amount of category-relevant information removed by the corresponding transformation. Because shape and texture are suppressed using different operators, their effects are not directly comparable. We introduce the Semantic Degradation Index (SDI), which quantifies the suppression-induced reduction in category separability relative to clean images in a fixed clean-reference discriminative space constructed from handcrafted features. On an ImageNet16-like benchmark, we use SDI to compare Gaussian blur for texture suppression with grid distortion for shape suppression over their overlapping degradation range. At comparable SDI values, all five evaluated ImageNet-trained convolutional neural networks (CNNs) retain less accuracy under Gaussian blur than under grid distortion. This results supports stronger texture than shape reliance under the evaluated operators, contrasting with the shape-dominant conclusion obtained from unmatched suppression conditions. The evaluated Vision Transformers (ViTs) also generally retain more accuracy than CNNs under both operators. To determine whether this advantage extends beyond classification, We evaluate fixed brain-encoding models using clean and suppressed images from the Natural Scenes Dataset. Under both operators, ViT features show smaller suppression-induced decreases in noise-ceiling-normalized explained variance than CNN features. These findings establish category separability as an important reference for interpreting suppression-based feature reliance and show that the CNN-ViT robustness difference extends to model representations predictive of human visual cortical responses.

Beyond Routing: Characterising Expert Tuning and Representation in Vision Mixture-of-Experts

Mixture-of-Experts (MoE) models are often interpreted by analysing which categories are routed to which experts. However, routing alone does not reveal what each expert actually encodes. We train sparsely-gated convolutional MoE models with a contrastive objective on natural images and characterise expert specialisation using tools from visual neuroscience. Extending from gating-level to expert-level analyses, we measure per-expert category separability, and per-expert tuning using the most exciting inputs. Extending from category-level to feature-level explanations, we interpret tuning via semantic dimensions derived from a dataset of human behavioural judgements (THINGS). Finally, we use tuning and representational similarity analysis to assess the stability of expertise-allocation across independent initialisations. We find that an animate-inanimate distinction dominates expert partitioning, apparent from gating through to expert readout, and is stable across independently trained models. Although routing statistics suggest relatively sparse, categorical preferences, expert analyses reveal broader tuning to continuous visual and semantic dimensions that extend beyond category boundaries. Experts exhibit similar category-separability to one another, despite distinct feature tuning, demonstrating the explanatory benefits of moving beyond category-level analyses. Together, these results show that expert specialisation in vision MoEs extends well beyond category routing and is better understood by probing fine-grained expert-level tuning and representational structure.