Research map / Label-efficient and robust learning / Recognition and long-tail
Does enhanced shape bias improve neural network robustness to common corruptions?
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 |
|---|---|---|---|
| Train on Stylized ImageNet instead of natural images | ResNet18 · ImageNet20 / ImageNet-C | mean corruption accuracy (%): 64.69 -> 77.64 | Helped |
| Train on edge maps instead of natural images (ablation) | ResNet18 · ImageNet20 / ImageNet-C | mean corruption accuracy (%): 64.69 -> 62.01 | Hurt |
| Add stylization to edge-map training (ablation) | ResNet18 · ImageNet20 / ImageNet-C | mean corruption accuracy (%): 62.01 -> 71.81 | Helped |
| Add natural-image content by superposing natural images with stylized edges (ablation) | ResNet18 · ImageNet20 / ImageNet-C | mean corruption accuracy (%): 71.81 -> 78.96 | Helped |
| Pretrain on edges, then train on stylized and natural images (ablation) | ResNet18 · ImageNet20 / ImageNet-C | mean corruption accuracy (%): 77.64 -> 71.55 | Hurt |
| Increase shape bias through edge pretraining and SIN training (ablation) | ResNet18 · ImageNet200 / ImageNet-C | mean corruption accuracy (%): 52.4 -> 47.7 | Hurt |
| Fine-tune normalization affine parameters on the target corruption | IN ResNet18 · ImageNet20 / ImageNet-C speckle noise | accuracy (%): 61.28 -> 82.7 | Helped |
| Fine-tune normalization affine parameters on the target corruption | IN ResNet18 · ImageNet20 / ImageNet-C Gaussian blur | accuracy (%): 42.96 -> 77.3 | Helped |
| Fine-tune normalization affine parameters on SIN | IN ResNet18 · ImageNet20 / SIN validation | accuracy (%): 42.0 -> 68.0 | Helped |
| Use style randomization rather than no feature-space styling (ablation) | IN ResNet18 · ImageNet20, 4 × 4 shuffled patches | accuracy (%): 67.22 -> 41.97 | Hurt |
| Replace style blending with style randomization (ablation) | IN ResNet18 · ImageNet20, shape-based cue conflict #400 | shape-based predictions: 82 -> 86 | Helped |
| Increase shape bias using E-SIN rather than SIN (ablation) | ResNet18 · ImageNet20 / ImageNet-C | mean corruption accuracy (%): 77.64 -> 71.55 | Hurt |
Closest papers by meaning
- Informative Dropout for Robust Representation Learning: A Shape-bias Perspective (ICML 2020 · similarity 0.90)
- The Origins and Prevalence of Texture Bias in Convolutional Neural Networks (NeurIPS 2020 · similarity 0.89)
- Emergence of Shape Bias in Convolutional Neural Networks through Activation Sparsity (NeurIPS 2023 · similarity 0.89)
- ImageNet-trained CNNs are not biased towards texture: Revisiting feature reliance through controlled suppression (NeurIPS 2025 · similarity 0.89)
- Enhancing Intrinsic Features for Debiasing via Investigating Class-Discerning Common Attributes in Bias-Contrastive Pair (CVPR 2024 · similarity 0.86)
- Improving robustness to corruptions with multiplicative weight perturbations (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.