Research map / Architectures and efficiency / CNN design
Structured Convolutions for Efficient Neural Network Design
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 |
|---|---|---|---|
| compress ResNet with structured convolutions (A versions, 2x smaller) | ResNet18/34/50/56 · CIFAR-10 and ImageNet | accuracy (%): ResNet50 76.15 -> 75.65; ResNet56 93.03 -> 92.65 | Hurt |
| Struct-V2-A instead of other compression methods for MobileNetV2 | MobileNetV2 · ImageNet val | top-1 accuracy (%): 71.29 vs 70.80 (AMC-MV2), 68.99 (ChPrune) | Helped |
| Struct-18-A vs WeightSVD and channel pruning at 2x compression | ResNet18 · ImageNet val | top-1 accuracy (%): 69.13 vs 67.71 / 67.69 | Helped |
| apply structured convolutions to HRNet | HRNetV2-W18-Small-v2 · Cityscapes | mean IoU (%): 76.1 -> 74.6 | Hurt |
| train with structural regularization loss then decompose, instead of directly training decomposed architecture (ablation) | Struct-18-A / Struct-MV2-A · ImageNet val | top-1 accuracy: +1.5% (Struct-18-A), +0.9% (Struct-MV2-A) | Helped |
| lambda too small (0.001) in structural regularization (ablation) | Struct-18-B · ImageNet val | post-decomposition top-1 accuracy (%): 66.19 -> 58.19 | Hurt |
| vary lambda from 0.01 to 1.0 (ablation) | Struct-18-A · ImageNet val | post-decomposition top-1 accuracy (%): post-decomposition accuracy changed by 0.08% | No clear effect |
| replace 3x3 depthwise kernels with 2x2 kernels (no sum-pooling) (ablation) | MobileNetV2 · ImageNet val | top-1 accuracy: 4.5% degradation vs Struct-V2-A | Hurt |
| add structural regularization loss during training (ablation) | Struct-18-A / Struct-MV2-A · training cost, batch size 256 | seconds per iteration: 0.44s -> 0.46s (Struct-18-A) | No clear effect |
Closest papers by meaning
- Structure-Aware Convolutional Neural Networks (NeurIPS 2018 · similarity 0.89)
- Linear CNNs Discover the Statistical Structure of the Dataset Using Only the Most Dominant Frequencies (ICML 2023 · similarity 0.88)
- Refining Architectures of Deep Convolutional Neural Networks (CVPR 2016 · similarity 0.88)
- Structured Receptive Fields in CNNs (CVPR 2016 · similarity 0.88)
- Compute Better Spent: Replacing Dense Layers with Structured Matrices (ICML 2024 · similarity 0.87)
- Which Frequencies do CNNs Need? Emergent Bottleneck Structure in Feature Learning (ICML 2024 · similarity 0.87)
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.