Research map / Architectures and efficiency / CNN design
Truly Shift-Invariant Convolutional Neural Networks
Check 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 |
|---|---|---|---|
| Replace conventional subsampling with APS | ResNet-18 · CIFAR-10 test with circular padding | consistency: 90.88% -> 100% | Helped |
| Replace conventional subsampling with APS | ResNet-18 · CIFAR-10 test with circular padding | accuracy: 91.96% -> 93.97% | Helped |
| Add APS instead of conventional subsampling | ResNet-18 · ImageNet validation with circular padding | consistency: 80.39% -> 100% | Helped |
| Add APS instead of conventional subsampling | ResNet-18 · ImageNet validation with circular padding | accuracy: 64.88% -> 67.05% | Helped |
| Add APS to BlurPool with filter size 5 (ablation) | ResNet-18 · CIFAR-10 test with standard padding | consistency: 95.13% -> 96.25% | Helped |
| Add APS to BlurPool with filter size 5 (ablation) | ResNet-18 · CIFAR-10 test with standard padding | accuracy: 94.22% -> 94.81% | Helped |
| Add blur filter size 5 to APS (ablation) | ResNet-18 · CIFAR-10 test with circular padding | accuracy: 93.97% -> 94.48% | Helped |
| Add blur filter size 3 to APS (ablation) | ResNet-18 · ImageNet validation with circular padding | accuracy: 67.05% -> 67.43% | Helped |
| Use maximum l-infinity norm rather than maximum l2 norm to select a polyphase component (ablation) | APS on ResNet-18 · CIFAR-10 test | accuracy: $94.03\pm$ $0.26\%$ -> $\boldsymbol{94.14\pm}$ $\boldsymbol{0.25\%}$ | Helped |
| Use minimum l2 norm rather than maximum l2 norm to select a polyphase component (ablation) | APS on ResNet-18 · CIFAR-10 test | accuracy: $94.03\pm$ $0.26\%$ -> $93.90\pm$ $0.16\%$ | Hurt |
| Replace conventional sampling with APS | ResNet-18 · Vertically flipped CIFAR-10 test images | consistency: 84.94% -> 100% | Helped |
| Train conventional sampling with random-shift data augmentation (ablation) | ResNet-18 · CIFAR-10 test | accuracy: Baseline 91.96%; Baseline + DA 94.33%; APS-3 94.53%; APS-3 + DA 94.61% | No clear effect |
Closest papers by meaning
- Making Vision Transformers Truly Shift-Equivariant (CVPR 2024 · similarity 0.84)
- Shifting the Paradigm: A Diffeomorphism Between Time Series Data Manifolds for Achieving Shift-Invariancy in Deep Learning (ICLR 2025 · similarity 0.83)
- Alias-Free ViT: Fractional Shift Invariance via Linear Attention (NeurIPS 2025 · similarity 0.83)
- On Making $SE(2)$-Invariant Networks Optimal (NeurIPS 2026 · similarity 0.82)
- Don’t Think It Twice: Exploit Shift Invariance for Efficient Online Streaming Inference of CNNs (AAAI 2025 · similarity 0.82)
- Equivariant Machine Learning on Graphs with Nonlinear Spectral Filters (NeurIPS 2024 · similarity 0.82)
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.