Research map / Architectures and efficiency / CNN design
Truly Scale-Equivariant Deep Nets with Fourier Layers
Md Ashiqur Rahman, Raymond A. Yeh
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 |
|---|---|---|---|
| Add scale-consistency loss (ablation) | Ours · MNIST-scale 5000 samples | accuracy: 0.9831 -> 0.9835 | Improved |
| Add scale-consistency loss (ablation) | Ours · MNIST-scale 5000 samples | scale-consistency: 0.9150 -> 0.9296 | Improved |
| Add scale-consistency loss (ablation) | Ours · MNIST-scale 1000 samples | scale-consistency: 0.8144 -> 0.8183 | Improved |
| Add spatially local Fourier layers (ablation) | Ours · MNIST-scale ideal | accuracy: 0.9835 -> 0.9889 | Improved |
| Add scale-equivariant ReLU (ablation) | Ours · MNIST-scale ideal | equivariance error: 7.32 -> 0.00 | Improved |
| Add scale-equivariant ReLU (ablation) | Ours · MNIST-scale ideal | accuracy: 0.9897 -> 0.9889 | No clear effect |
| Ideal upsampling instead of zero-padding for baselines (ablation) | SI-CovNet · STL10-scale ideal | accuracy: 0.5588 -> 0.6722 | Improved |
| Ideal upsampling instead of zero-padding for baselines (ablation) | SS-CNN · STL10-scale ideal | accuracy: 0.4788 -> 0.3246 | Hurt |
| Use Fourier scale-equivariant layers instead of Fourier CNN | Ours · MNIST-scale ideal | scale-consistency: 0.2421 -> 0.9716 | Improved |
| Reduce training samples from 5000 to 1000 (ablation) | Per Res. CNN · MNIST-scale | accuracy: 0.9118 -> 0.5815 | Hurt |
| Reduce training samples from 5000 to 1000 | Ours · MNIST-scale | accuracy: 0.9835 -> 0.9606 | Hurt |
| Train only on full-resolution MNIST (unseen scales) | Ours · MNIST-scale missing scales | accuracy: 0.9888 vs 0.9914 | No clear effect |
Closest papers by meaning
- Empowering Networks With Scale and Rotation Equivariance Using A Similarity Convolution (ICLR 2023 · similarity 0.91)
- Scaling-Translation-Equivariant Networks with Decomposed Convolutional Filters (JMLR 2022 · similarity 0.90)
- Scale-Equivariant Steerable Networks (ICLR 2020 · similarity 0.89)
- A Probabilistic Approach to Learning the Degree of Equivariance in Steerable CNNs (ICML 2024 · similarity 0.87)
- Scale Equivariant Graph Metanetworks (NeurIPS 2024 · similarity 0.87)
- On the Fourier analysis in the SO(3) space : the EquiLoPO Network (ICLR 2025 · similarity 0.86)
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.
