AI Paper Atlas

Research map / Architectures and efficiency / CNN design

When Random Saliency Looks Trained: Architectural Center Bias in CNN Interpretability

Keying Kuang, Iain Carmichael, Elizabeth Purdom

ICML 2026 · Architectures and efficiency › CNN design · approach: Biologically inspired CNN explanation

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 changedWhereMeasuredEffect
replace zero padding with reflect, replicate or circular padding (ablation)random 3-layer CNN · Random-init CNN, 100 ImageNet imagesCenter-Bias AUCImproved
increase network depth so receptive field grows (ablation)random depth-variable CNN · Random-init CNN, 224x224Center-Bias AUCMixed
use ViT or MLP instead of CNN (ablation)random-init models · Random-init, 100 ImageNet imagesCenter-Bias AUC: center AUC near 0.5Improved
train on corner-shrunk images instead of center-shrunkCNNs (fine-tuned) · Corner-shift benchmarkCenter-Shift Index: 0.064 -> 0.457Improved
move object from center to cornerInception-v4 trained vs random · Trained-random saliency similaritySpearman rho: 0.40 -> 0.15Hurt
use corner-shrunk instead of center-shrunk trainingInception-v4 · ImageNet valTop-1 accuracy (%): 68.18 -> 66.00Hurt
use corner-shrunk instead of center-shrunk trainingViT-B/16 · ImageNet valTop-1 accuracy (%): 74.59 -> 75.53No clear effect
change center crop fraction from 0.5 to 0.3 (ablation)center-bias architecture ranking · Center-bias rankingSpearman correlation: 1.000 -> 0.993No clear effect
change center crop fraction from 0.5 to 0.7 (ablation)center-bias architecture ranking · Center-bias rankingSpearman correlation: 1.000 -> 0.972No clear effect
shrink object scale from 0.5 to 0.4 (top-left)ResNet18 · Corner-shift benchmarkMean CSI: 0.457 -> 0.590No clear effect
move object placement from top-left to bottom-rightResNet18 · Corner-shift benchmarkMean CSI: 0.457 -> 0.524No clear effect
swap Grad-CAM for input-gradient methods (ablation)center-bias analysis · Random-init modelsCenter-bias metricsMixed

Closest papers by meaning

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.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy