Research map / Architectures and efficiency / CNN design
When Random Saliency Looks Trained: Architectural Center Bias in CNN Interpretability
Keying Kuang, Iain Carmichael, Elizabeth Purdom
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| replace zero padding with reflect, replicate or circular padding (ablation) | random 3-layer CNN · Random-init CNN, 100 ImageNet images | Center-Bias AUC | Improved |
| increase network depth so receptive field grows (ablation) | random depth-variable CNN · Random-init CNN, 224x224 | Center-Bias AUC | Mixed |
| use ViT or MLP instead of CNN (ablation) | random-init models · Random-init, 100 ImageNet images | Center-Bias AUC: center AUC near 0.5 | Improved |
| train on corner-shrunk images instead of center-shrunk | CNNs (fine-tuned) · Corner-shift benchmark | Center-Shift Index: 0.064 -> 0.457 | Improved |
| move object from center to corner | Inception-v4 trained vs random · Trained-random saliency similarity | Spearman rho: 0.40 -> 0.15 | Hurt |
| use corner-shrunk instead of center-shrunk training | Inception-v4 · ImageNet val | Top-1 accuracy (%): 68.18 -> 66.00 | Hurt |
| use corner-shrunk instead of center-shrunk training | ViT-B/16 · ImageNet val | Top-1 accuracy (%): 74.59 -> 75.53 | No clear effect |
| change center crop fraction from 0.5 to 0.3 (ablation) | center-bias architecture ranking · Center-bias ranking | Spearman correlation: 1.000 -> 0.993 | No clear effect |
| change center crop fraction from 0.5 to 0.7 (ablation) | center-bias architecture ranking · Center-bias ranking | Spearman correlation: 1.000 -> 0.972 | No clear effect |
| shrink object scale from 0.5 to 0.4 (top-left) | ResNet18 · Corner-shift benchmark | Mean CSI: 0.457 -> 0.590 | No clear effect |
| move object placement from top-left to bottom-right | ResNet18 · Corner-shift benchmark | Mean CSI: 0.457 -> 0.524 | No clear effect |
| swap Grad-CAM for input-gradient methods (ablation) | center-bias analysis · Random-init models | Center-bias metrics | Mixed |
Closest papers by meaning
- Understanding and Visualizing Deep Visual Saliency Models (CVPR 2019 · similarity 0.89)
- CAMERAS: Enhanced Resolution and Sanity Preserving Class Activation Mapping for Image Saliency (CVPR 2021 · similarity 0.88)
- Modeling Saliency Dataset Bias (ICCV 2025 · similarity 0.88)
- There and Back Again: Revisiting Backpropagation Saliency Methods (CVPR 2020 · similarity 0.88)
- Great Minds Think Alike: The Universal Convergence Trend of Input Salience (NeurIPS 2024 · similarity 0.86)
- Explainable Saliency: Articulating Reasoning with Contextual Prioritization (CVPR 2025 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
