Research map / Architectures and efficiency / CNN design
Learning to Convolve: A Generalized Weight-Tying Approach
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 |
|---|---|---|---|
| learn filter rotations with a learned basis rather than handcrafted interpolation | CIFAR-10 group CNN · CIFAR-10 | test accuracy: 90.72 -> 92.75 | Helped |
| learn filter rotations with a learned basis rather than handcrafted interpolation | CIFAR-10 group CNN · CIFAR-10 with full rotation augmentation | test accuracy: 89.45 -> 89.84 | Helped |
| limit basis learning to 0°–90° and manually rotate filters for other angles (ablation) | Full learned-basis model · CIFAR-10 | test accuracy: 92.75 -> 90.93 | Hurt |
| limit basis learning to 0°–90° and manually rotate filters for other angles (ablation) | Full learned-basis model · CIFAR-10 with full rotation augmentation | test accuracy: 89.84 -> 89.60 | Hurt |
| increase the number of basis elements threefold (ablation) | Full learned-basis model · CIFAR-10 | test accuracy: 92.75 -> 91.16 | Hurt |
| increase the number of basis elements threefold (ablation) | Full learned-basis model · CIFAR-10 with full rotation augmentation | test accuracy: 89.84 -> 89.93 | Helped |
| use the Partial basis rather than the Full basis (ablation) | learned-basis group CNN · MNIST and CIFAR-10 rotated test inputs | test error fluctuation: At its worst it has a ~2% error increase over its best error on MNIST and 20% on CIFAR-10; other interpolation methods s | Helped |
| add only 90° rotation augmentation during training (ablation) | Partial model · MNIST and CIFAR-10 | validation accuracy | No clear effect |
| increase rotational data augmentation (ablation) | Partial model · MNIST and CIFAR-10 at intermediate input angles | validation performance | Helped |
Closest papers by meaning
- Learning Steerable Filters for Rotation Equivariant CNNs (CVPR 2018 · similarity 0.90)
- RotDCF: Decomposition of Convolutional Filters for Rotation-Equivariant Deep Networks (ICLR 2019 · similarity 0.89)
- FILTRA: Rethinking Steerable CNN by Filter Transform (ICML 2021 · similarity 0.88)
- Group Downsampling with Equivariant Anti-aliasing (ICLR 2025 · similarity 0.86)
- Variational Partial Group Convolutions for Input-Aware Partial Equivariance of Rotations and Color-Shifts (ICML 2024 · similarity 0.86)
- The Master Key Filters Hypothesis: Deep Filters Are General (AAAI 2025 · 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.