Research map / Architectures and efficiency / CNN design
Are Large Kernels Better Teachers than Transformers for ConvNets?
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 |
|---|---|---|---|
| Use SLaK-T rather than CSWin-T as the NKD teacher | ResNet-50 · ImageNet | top-1 accuracy: 77.68 -> 78.57 | Helped |
| Use SLaK-T rather than CSWin-T as the NKD teacher | ConvNeXt-T · ImageNet | top-1 accuracy: 81.65 -> 82.17 | Helped |
| Use SLaK-T rather than ConvNeXt-T as the NKD teacher | ResNet-50 · ImageNet | top-1 accuracy: 78.04 -> 78.57 | Helped |
| Use SLaK-T rather than Swin-T as the KD teacher | ConvNeXt-T · ImageNet | top-1 accuracy: 80.93 -> 81.86 | Helped |
| Combine feature distillation with NKD rather than feature distillation alone | ResNet-50 distilled from SLaK-T · ImageNet | top-1 accuracy: 76.85 -> 77.99 | Helped |
| Use SLaK-T rather than Swin-T for feature distillation | ResNet-50 · ImageNet | top-1 accuracy: 76.77 -> 76.85 | Helped |
| Use SLaK-T rather than Swin-T as the NKD teacher for 300 epochs | ResNet-50 · ImageNet | top-1 accuracy: 79.44 -> 80.24 | Helped |
| Increase feature-distillation depth from L=1 to L=3 (ablation) | ResNet-50 distilled from SLaK-T · ImageNet | top-1 accuracy: 76.85 -> 76.94 | Helped |
| Increase feature-distillation depth from L=1 to L=2 (ablation) | ResNet-50 distilled from SLaK-T · ImageNet | top-1 accuracy: 76.85 -> 76.79 | Hurt |
| Increase feature-distillation depth from L=1 to L=4 (ablation) | ResNet-50 distilled from SLaK-T · ImageNet | top-1 accuracy: 76.85 -> 76.92 | Helped |
| Extend distillation training from 120 to 300 epochs (ablation) | ResNet-50 with NKD · ImageNet | top-1 accuracy: 78.57 -> 80.24 | Helped |
| Distill from SLaK-T rather than Swin-T | ResNet-50 · ImageNet-Sketch | top-1 accuracy: 26.07 -> 28.54 | Helped |
Closest papers by meaning
- ScaleKD: Strong Vision Transformers Could Be Excellent Teachers (NeurIPS 2024 · similarity 0.90)
- UniRepLKNet: A Universal Perception Large-Kernel ConvNet for Audio Video Point Cloud Time-Series and Image Recognition (CVPR 2024 · similarity 0.89)
- Revealing the Dark Secrets of Extremely Large Kernel ConvNets on Robustness (ICML 2024 · similarity 0.89)
- VkD: Improving Knowledge Distillation using Orthogonal Projections (CVPR 2024 · similarity 0.87)
- KernelWarehouse: Rethinking the Design of Dynamic Convolution (ICML 2024 · similarity 0.87)
- PeLK: Parameter-efficient Large Kernel ConvNets with Peripheral Convolution (CVPR 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.