Research map / Architectures and efficiency / CNN design
Emergence of Shape Bias in Convolutional Neural Networks through Activation Sparsity
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 |
|---|---|---|---|
| add 20% spatial Top-K layer during training | ResNet-18 · IN-S1 | top-1 accuracy: 87.8 -> 89.4 | Helped |
| add 20% spatial Top-K layer during training | ResNet-18 · IN-S2 | top-1 accuracy: 81.3 -> 83.4 | Helped |
| add 20% spatial Top-K layer during training | ResNet-18 · Stylized IN-S1 | top-1 accuracy: 49.3 -> 55.4 | Helped |
| add 20% spatial Top-K layer during training | ResNet-18 · Stylized IN-S2 | top-1 accuracy: 52.4 -> 59.7 | Helped |
| replace Top-K values with channel mean vs plain Top-K (ablation) | ResNet-18 w. Top-K · IN-S1 | top-1 accuracy: 89.4 -> 84.9 | Hurt |
| replace Top-K values with channel mean vs plain Top-K (ablation) | ResNet-18 w. Top-K · IN-S2 | top-1 accuracy: 83.4 -> 75.5 | Hurt |
| replace Top-K values with channel mean, vs baseline ResNet-18 (ablation) | ResNet-18 · Stylized IN-S1 | top-1 accuracy: 49.3 -> 56.8 | Helped |
| replace Top-K values with channel mean, vs baseline ResNet-18 (ablation) | ResNet-18 · Stylized IN-S2 | top-1 accuracy: 52.4 -> 53.1 | No clear effect |
| add Top-K at 32x32 layer (K=5%) | FastGAN · Jeep-100 | FID: 49.0 -> 38.7 | Helped |
| add Top-K at 32x32 layer (K=5%) | FastGAN · Jeep-100 | KID: 12.0 -> 5.9 | Helped |
| add Top-K at 32x32 layer (K=5%) | FastGAN · Fish-100 | FID: 46.2 -> 38.2 | Helped |
| add Top-K at 32x32 layer (K=5%) | FastGAN · Train-100 | FID: 46.1 -> 40.2 | Helped |
Closest papers by meaning
- Shape or Texture: Understanding Discriminative Features in CNNs (ICLR 2021 · similarity 0.91)
- Does enhanced shape bias improve neural network robustness to common corruptions? (ICLR 2021 · similarity 0.89)
- ImageNet-trained CNNs are not biased towards texture: Revisiting feature reliance through controlled suppression (NeurIPS 2025 · similarity 0.88)
- Visual Anagrams Reveal Hidden Differences in Holistic Shape Processing Across Vision Models (NeurIPS 2025 · similarity 0.86)
- From Flat to Hierarchical: Extracting Sparse Representations with Matching Pursuit (NeurIPS 2025 · similarity 0.85)
- Vision Transformers Learn Gestalt-Like Figure-Ground Cues from Natural Images (NeurIPS 2026 · 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.