Research map / Image generation / Autoregressive and latent generation
Unlocking Feature Visualization for Deep Network with MAgnitude Constrained Optimization
Thomas FEL, Thibaut Boissin, Victor Boutin, Agustin Martin Picard, Paul Novello, Julien Colin, Drew Linsley, Tom ROUSSEAU, Remi Cadene, Lore Goetschalckx, Laurent Gardes, Thomas Serre
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 |
|---|---|---|---|
| Use magnitude-constrained phase optimization instead of Fourier optimization | Feature visualization · FlexiViT, ViT, and ResNetV2 | plausibility score (1-KNN): 1815 -> 1473; 1817 -> 1097; 1837 -> 1248 | Improved |
| Use magnitude-constrained phase optimization instead of Fourier optimization | Feature visualization · FlexiViT, ViT, and ResNetV2 | FID: 250.25 -> 230.68; 257.81 -> 241.68; 318.15 -> 312.66 | Improved |
| Remove transparency mask (ablation) | MACO · FlexiViT | plausibility: 571.68 -> 617.9 (+46.2) | Hurt |
| Remove transparency mask (ablation) | MACO · FlexiViT | FID: 211.0 -> 208.1 (-2.9) | No clear effect |
| Remove random crop augmentation (ablation) | MACO · FlexiViT | plausibility: 571.68 -> 680.1 (+62.2) | Hurt |
| Remove random crop augmentation (ablation) | MACO · FlexiViT | FID: 211.0 -> 299.2 (-91.1) | Hurt |
| Remove random crop augmentation (ablation) | MACO · FlexiViT | logit magnitude: 5.12 -> 8.18 (+3.1) | Improved |
| Remove additive noise augmentation (ablation) | MACO · FlexiViT | plausibility: 571.68 -> 707.3 (+27.1) | Hurt |
| Remove additive noise augmentation (ablation) | MACO · FlexiViT | FID: 211.0 -> 324.5 (-25.3) | Hurt |
| Remove additive noise augmentation (ablation) | MACO · FlexiViT | logit magnitude: 5.12 -> 11.7 (+3.5) | Improved |
| Show MACO visualization instead of Olah visualization | Human prediction of model-favored query · VGG16, ResNet50, and ViT | logodds of correct query choice: 0.43, SE=0.13, z=3.31, p=.005; 0.66, SE=0.13, z=4.99, p<.0001 | Improved |
| Add MACO visualization to the no-visualization control | Human prediction of model-favored query · VGG16, ResNet50, and ViT | logodds of correct query choice: 0.69, SE=0.13, z=5.38, p<.0001; 0.92, SE=0.13, z=7.07, p<.0001 | Improved |
Closest papers by meaning
- Don't trust your eyes: on the (un)reliability of feature visualizations (ICML 2024 · similarity 0.87)
- VITAL: More Understandable Feature Visualization through Distribution Alignment and Relevant Information Flow (ICCV 2025 · similarity 0.87)
- Visualization of Supervised and Self-Supervised Neural Networks via Attribution Guided Factorization (AAAI 2021 · similarity 0.86)
- Linking in Style: Understanding learned features in deep learning models (ECCV 2024 · similarity 0.86)
- Understanding Visual Feature Reliance through the Lens of Complexity (NeurIPS 2024 · similarity 0.85)
- Manipulating Feature Visualizations with Gradient Slingshots (NeurIPS 2025 · similarity 0.84)
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.
