Research map / Architectures and efficiency / CNN design
Revisiting Dynamic Convolution via Matrix Decomposition
Yunsheng Li, Yinpeng Chen, Xiyang Dai, mengchen liu, Dongdong Chen, Ye Yu, Lu Yuan, Zicheng Liu, Mei Chen, Nuno Vasconcelos
AI-extracted. The measured effects below were read by AI from the paper's full text (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 dynamic attention over kernel groups with low-dimensional dynamic channel fusion | MobileNetV2 ×0.5 · ImageNet | Top-1 accuracy: 65.4 -> 70.2 | Improved |
| Add dynamic channel-wise attention to the static kernel (ablation) | DCD on MobileNetV2 ×0.5 · ImageNet | Top-1 accuracy: 69.2 -> 69.8 | Improved |
| Add dynamic channel fusion to the static kernel (ablation) | DCD on MobileNetV2 ×0.5 · ImageNet | Top-1 accuracy: 65.4 -> 69.2 | Improved |
| Add dynamic channel-wise attention to the static kernel (ablation) | DCD on ResNet-18 · ImageNet | Top-1 accuracy: 70.4 -> 71.5 | Improved |
| Add dynamic channel fusion to the static kernel (ablation) | DCD on ResNet-18 · ImageNet | Top-1 accuracy: 70.4 -> 72.8 | Improved |
| Reduce latent dimension L to ×0.25 of default (ablation) | DCD on MobileNetV2 ×0.5 · ImageNet | Top-1 accuracy: 69.8 -> 68.7 | Hurt |
| Reduce latent dimension L to ×0.50 of default (ablation) | DCD on MobileNetV2 ×0.5 · ImageNet | Top-1 accuracy: 69.8 -> 69.0 | Hurt |
| Reduce latent dimension L to ×0.75 of default (ablation) | DCD on MobileNetV2 ×0.5 · ImageNet | Top-1 accuracy: 69.8 -> 69.6 | Hurt |
| Use a block-diagonal dynamic residual with B=8 (ablation) | DCD on MobileNetV2 ×0.5 · ImageNet | Top-1 accuracy: 69.8 -> 68.5 | Hurt |
| Apply DCD to pointwise convolution and classifier rather than all three layer types (ablation) | MobileNetV2 ×0.5 · ImageNet | Top-1 accuracy: 70.0 -> 70.2 | Improved |
| Add DCD to depthwise convolution when pointwise convolution and classifier already use DCD (ablation) | MobileNetV2 ×0.5 · ImageNet | Top-1 accuracy: 70.2 -> 70.0 | Hurt |
| Use channel fusion alone instead of joint channel and kernel-element fusion (ablation) | DCD on ResNet-18 3×3 convolutions · ImageNet | Top-1 accuracy: 71.3 -> 73.1 | Improved |
Closest papers by meaning
- Dynamic Convolution: Attention Over Convolution Kernels (CVPR 2020 · similarity 0.89)
- KernelWarehouse: Rethinking the Design of Dynamic Convolution (ICML 2024 · similarity 0.87)
- Pick-or-Mix: Dynamic Channel Sampling for ConvNets (CVPR 2024 · similarity 0.85)
- BilevelPruning: Unified Dynamic and Static Channel Pruning for Convolutional Neural Networks (CVPR 2024 · similarity 0.84)
- Dynamic Clustering Convolutional Neural Network (AAAI 2025 · similarity 0.84)
- Dynamic Group Normalization: Spatio-Temporal Adaptation to Evolving Data Statistics (CVPR 2025 · 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.
