Research map / Architectures and efficiency / CNN design
Exemplary Natural Images Explain CNN Activations Better than State-of-the-Art Feature Visualization
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 synthetic feature visualizations with strongly activating natural reference images | Human prediction of Inception V1 feature-map activations · ImageNet feature-map prediction, Experiment I | accuracy: 82±4% -> 92±2% | Helped |
| Provide synthetic reference images instead of no reference images | Human prediction of Inception V1 feature-map activations · ImageNet feature-map prediction, Experiment I | accuracy | Helped |
| Replace synthetic with natural reference images when showing one maximally activating image | Human prediction of Inception V1 feature-map activations · ImageNet feature-map prediction, Experiment II, Max 1 | accuracy: 64±5% -> 73±4% | Helped |
| Replace synthetic with natural reference images when showing Min+Max 9 | Human prediction of Inception V1 feature-map activations · ImageNet feature-map prediction, Experiment II, Min+Max 9 | accuracy: 72±4% -> 91±3 | Helped |
| Show nine rather than one reference image (ablation) | Human prediction with natural or synthetic references · ImageNet feature-map prediction, Experiment II | accuracy | Helped |
| Add minimally activating reference images alongside maximally activating images (ablation) | Human prediction with natural or synthetic references · ImageNet feature-map prediction, Experiment II | accuracy | Helped |
| Use natural rather than synthetic references for a single-image presentation (ablation) | Human prediction of Inception V1 feature-map activations · ImageNet feature-map prediction, Experiment II, Max 1 | accuracy: 64±5% -> 73±4% | Helped |
| Use hand-picked rather than randomly selected feature maps (ablation) | Human prediction with synthetic or natural references · Inception V1 feature-map prediction | accuracy | No clear effect |
| Have expert rather than lay participants perform the task (ablation) | Human prediction of Inception V1 feature-map activations · ImageNet feature-map prediction, Experiment II | accuracy | No clear effect |
Closest papers by meaning
- How Well do Feature Visualizations Support Causal Understanding of CNN Activations? (NeurIPS 2021 · similarity 0.88)
- Synthesizing the preferred inputs for neurons in neural networks via deep generator networks (NeurIPS 2016 · similarity 0.87)
- Don't trust your eyes: on the (un)reliability of feature visualizations (ICML 2024 · similarity 0.85)
- Finer-CAM: Spotting the Difference Reveals Finer Details for Visual Explanation (CVPR 2025 · similarity 0.85)
- The Unmet Promise of Synthetic Training Images: Using Retrieved Real Images Performs Better (NeurIPS 2024 · similarity 0.85)
- Parallel Backpropagation for Shared-Feature Visualization (NeurIPS 2024 · 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.