Research map / Architectures and efficiency / CNN design
Shift: A Zero FLOP, Zero Parameter Alternative to Spatial Convolutions
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 |
|---|---|---|---|
| Replace ResNet spatial convolutions with shift-based modules | ResNet110 · CIFAR100 | Top-1 accuracy: 72.11% -> 74.10% | Helped |
| Use shift modules instead of convolutions at matched parameter count | ResNet110 · CIFAR-100 | CIFAR-100 Acc: 60.44% -> 67.84% | Helped |
| Use ShiftResNet56-3 instead of ResNet56 | ResNet56 · CIFAR-100 | Top1 Acc: 69.27% -> 69.77%; 151M -> 54M FLOPs; 0.87M -> 0.29M parameters | Helped |
| Increase expansion from 1 to 9 (ablation) | ShiftResNet20 · CIFAR100 | Top-1 accuracy: 55.62% -> 69.82% | Helped |
| Increase expansion from 1 to 9 (ablation) | ShiftResNet56 · CIFAR100 | Top-1 accuracy: 65.21% -> 73.64% | Helped |
| Increase expansion from 1 to 9 (ablation) | ShiftResNet110 · CIFAR100 | Top-1 accuracy: 67.84% -> 74.10% | Helped |
| Use ShiftNet-A instead of MobileNet-224 (1.0) | ImageNet classification models · ImageNet validation | Top-1 accuracy: 70.6 vs 70.1; 4.2 vs 4.1 parameters (Millions) | Mixed |
| Shrink channels across ShiftNet-A to form ShiftNet-B (ablation) | ShiftNet-A · ImageNet validation | Top-1 accuracy: 70.1 -> 61.2 | Hurt |
| Use the smaller, shallower ShiftNet-C design (ablation) | ShiftNet · ImageNet validation | Top-1 accuracy: 58.8 vs 57.5; 0.78 vs 1.2 parameters (Millions) | Helped |
| Replace FaceNet with the smaller ShiftFaceNet | FaceNet · LFW | Verification accuracy: 97.1±1.3 -> 96.0±1.4 | Hurt |
| Replace FaceNet with the smaller ShiftFaceNet | FaceNet · YTF | Verification accuracy: 92.0±1.1 -> 90.1±0.9 | Hurt |
| Replace FaceNet with the smaller ShiftFaceNet | FaceNet · MSC | Verification accuracy: 79.2±1.7 -> 77.6±1.7 | Hurt |
Closest papers by meaning
- All You Need Is a Few Shifts: Designing Efficient Convolutional Neural Networks for Image Classification (CVPR 2019 · similarity 0.89)
- Constructing Fast Network through Deconstruction of Convolution (NeurIPS 2018 · similarity 0.88)
- Warped Convolutions: Efficient Invariance to Spatial Transformations (ICML 2017 · similarity 0.86)
- Truly Shift-Invariant Convolutional Neural Networks (CVPR 2021 · similarity 0.85)
- Improving the Learning Capability of Small-size Image Restoration Network by Deep Fourier Shifting (NeurIPS 2024 · similarity 0.85)
- KernelWarehouse: Rethinking the Design of Dynamic Convolution (ICML 2024 · similarity 0.84)
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.