Research map / Architectures and efficiency / CNN design
All You Need Is a Few Shifts: Designing Efficient Convolutional Neural Networks for Image Classification
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 |
|---|---|---|---|
| Use FE-Net 1.375x with SSL instead of depthwise convolution (ablation) | FE-Net 1.375x · ImageNet | GPU inference time: 30.4ms -> 23.1ms | Helped |
| Replace SSL with depthwise convolution (ablation) | FE-Net 1.375x · ImageNet | Top-1 accuracy: 75.0% -> 75.2% | No clear effect |
| Place SE in the inverted bottleneck expansion part rather than at the block output (ablation) | FE-Net 1.375x · ImageNet | Top-1 accuracy: 75.6% -> 76.5% | Helped |
| Increase the shift penalty from λ=0 to λ=5e-4 (ablation) | ShiftResNet-20 · CIFAR10 | Accuracy: 91.7% -> 89.8% | Hurt |
| Increase the shift penalty from λ=0 to λ=5e-4 (ablation) | ShiftResNet-20 · CIFAR100 | Accuracy: 69.2% -> 67.7% | Hurt |
| Add shift operations rather than use 1x1 convolutions only (ablation) | ShiftResNet-20 · CIFAR100 | Accuracy: 56.7% -> 69.2% | Helped |
| Remove the 4 least important shift layers (ablation) | ShiftResNet-20 SSL · CIFAR100 | Accuracy: 69.2% -> 68.2% | Hurt |
| Remove all shift layers (ablation) | ShiftResNet-20 SSL · CIFAR10 | Accuracy: 91.7% -> 81.5% | Hurt |
| Remove all shift layers (ablation) | ShiftResNet-20 SSL · CIFAR100 | Accuracy: 69.2% -> 56.7% | Hurt |
| Use shift learning rather than heuristic grouped shifts (ablation) | ShiftResNet-20 · CIFAR10 | Accuracy: 90.6% -> 91.7% | Helped |
| Use shift learning rather than heuristic grouped shifts (ablation) | ShiftResNet-20 · CIFAR100 | Accuracy: 68.6% -> 69.2% | Helped |
| Add SE modules (ablation) | FE-Net 1.375x · ImageNet | Top-1 accuracy: 75.0% -> 75.6%; 75.0% -> 76.5% | Helped |
Closest papers by meaning
- Constructing Fast Network through Deconstruction of Convolution (NeurIPS 2018 · similarity 0.89)
- Shift: A Zero FLOP, Zero Parameter Alternative to Spatial Convolutions (CVPR 2018 · similarity 0.89)
- SBNet: Sparse Blocks Network for Fast Inference (CVPR 2018 · similarity 0.85)
- S$^3$: Sign-Sparse-Shift Reparametrization for Effective Training of Low-bit Shift Networks (NeurIPS 2021 · similarity 0.85)
- Improving the Learning Capability of Small-size Image Restoration Network by Deep Fourier Shifting (NeurIPS 2024 · similarity 0.85)
- Super Sparse Convolutional Neural Networks (AAAI 2019 · 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.