Research map / Architectures and efficiency / CNN design
Grounding inductive biases in natural images: invariance stems from variations in data
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 no augmentation with RRC | ResNet18 · ImageNet val | Top-1 accuracy: 63.49± 0.1 -> 70.05± 0.1 | Helped |
| Use translation augmentation instead of no augmentation (ablation) | ResNet18 · ImageNet val | Top-1 accuracy: 63.49± 0.1 -> 69.14± 0.0 | Helped |
| Add random-size center crop (scale) to 30% translation (ablation) | ResNet18 · ImageNet val | Top-1 accuracy: 69.14± 0.0 -> 69.30± 0.0 | No clear effect |
| Use translation plus random-size center crop rather than RRC (ablation) | ResNet18 · ImageNet val | Top-1 accuracy: 70.05± 0.1 -> 69.30± 0.0 | Hurt |
| Use scale-only random-size center crop instead of no augmentation (ablation) | ResNet18 · ImageNet val | Top-1 accuracy: 63.49± 0.1 -> 67.84± 0.1 | Helped |
| Use translation-only fixed-size random crop instead of no augmentation (ablation) | ResNet18 · ImageNet val | Top-1 accuracy: 63.49± 0.1 -> 67.93± 0.0 | Helped |
| Set scale-distribution β to 0.1 instead of 1 (ablation) | RRC on ResNet18 · ImageNet val | Top-1 accuracy: 70.11± 0.0 -> 69.32± 0.0 | Hurt |
| Set scale-distribution β to 10 instead of 1 (ablation) | RRC on ResNet18 · ImageNet val | Top-1 accuracy: 70.11± 0.0 -> 63.74± 0.1 | Hurt |
| Increase translation magnitude to 50% with random-size center crop (ablation) | ResNet18 · ImageNet val | Top-1 accuracy: ~≈ 67.5% | Hurt |
| Add standard augmentation during training | ResNet18 and ViT · ImageNet transformations | translation invariance | Helped |
| Add standard augmentation during training | ResNet18 and ViT · ImageNet transformations | scale invariance | No clear effect |
| Train the model rather than use its untrained architecture (ablation) | ResNet18 and ViT · ImageNet transformations | rotation invariance | Hurt |
Closest papers by meaning
- Learning Invariances in Neural Networks from Training Data (NeurIPS 2020 · similarity 0.91)
- The Lie Derivative for Measuring Learned Equivariance (ICLR 2023 · similarity 0.90)
- Automatic Data Augmentation via Invariance-Constrained Learning (ICML 2023 · similarity 0.89)
- Understanding Invariance via Feedforward Inversion of Discriminatively Trained Classifiers (ICML 2021 · similarity 0.88)
- Tilt your Head: Activating the Hidden Spatial-Invariance of Classifiers (ICML 2024 · similarity 0.87)
- From Bricks to Bridges: Product of Invariances to Enhance Latent Space Communication (ICLR 2024 · similarity 0.87)
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.