Research map / Architectures and efficiency / CNN design
Alias-Free Convnets: Fractional Shift Invariance via Polynomial Activations
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 |
|---|---|---|---|
| Use the complete alias-free ConvNeXt modifications | ConvNeXt · ImageNet translation adversarial, half-pixel grid | adversarial accuracy: 73.65 -> 81.04 | Helped |
| Use the complete alias-free ConvNeXt modifications | ConvNeXt · ImageNet validation | test accuracy: 82.12 -> 81.04 | Hurt |
| Replace GeLU with polynomial activation (ablation) | ConvNeXt-Baseline · ImageNet validation | test accuracy: 82.12 -> 81.77 | Hurt |
| Add BlurPool (ablation) | ConvNeXt with polynomial activation · ImageNet validation | test accuracy: 81.77 -> 78.99 | Hurt |
| Add BlurPool (ablation) | ConvNeXt with polynomial activation · ImageNet validation | fractional shift consistency: 92.708 -> 96.572 | Helped |
| Add activation before the first downsampling layer (ablation) | ConvNeXt with BlurPool · ImageNet validation | test accuracy: 78.99 -> 81.51 | Helped |
| Replace per-pixel LayerNorm scaling with alias-free LayerNorm (ablation) | ConvNeXt with first-layer activation · ImageNet validation | test accuracy: 81.51 -> 80.66 | Hurt |
| Replace per-pixel LayerNorm scaling with alias-free LayerNorm (ablation) | ConvNeXt with first-layer activation · ImageNet validation | fractional shift consistency: 97.347 -> 96.990 | Hurt |
| Use upsampled alias-free polynomial activation (ablation) | ConvNeXt with alias-free LayerNorm · ImageNet validation | fractional shift consistency: 96.990 -> 100.000 | Helped |
| Replace GeLU with degree-2 polynomial activation | ConvNeXt-tiny · ImageNet | test accuracy: 82.1 -> 81.98 | Hurt |
| Replace GeLU with degree-2 polynomial activation | ViT · CIFAR10 | test accuracy: 97.04 -> 97.08 | Helped |
| Replace GeLU with degree-4 polynomial activation | Deit-tiny · ImageNet | test accuracy: 72.29 -> 71.96 | Hurt |
Closest papers by meaning
- Alias-Free ViT: Fractional Shift Invariance via Linear Attention (NeurIPS 2025 · similarity 0.92)
- Making Convolutional Networks Shift-Invariant Again (ICML 2019 · similarity 0.90)
- Truly Shift-Invariant Convolutional Neural Networks (CVPR 2021 · similarity 0.86)
- On Making $SE(2)$-Invariant Networks Optimal (NeurIPS 2026 · similarity 0.84)
- Alias-Free Latent Diffusion Models: Improving Fractional Shift Equivariance of Diffusion Latent Space (CVPR 2025 · similarity 0.84)
- Group Downsampling with Equivariant Anti-aliasing (ICLR 2025 · similarity 0.83)
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.