Research map / Architectures and efficiency / CNN design
CMT: Convolutional Neural Networks Meet Vision Transformers
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 a 4-stage hierarchy and lightweight attention (ablation) | DeiT-S · ImageNet-1K | Top-1 accuracy: 79.8% -> 81.4% | Helped |
| Add a convolution-based stem (ablation) | DeiT-S-4Stage · ImageNet-1K | Top-1 accuracy: 81.4% -> 81.9% | Helped |
| Add the local perception unit (ablation) | DeiT-S-4Stage + Stem · ImageNet-1K | Top-1 accuracy: 81.9% -> 82.7% | Helped |
| Add a shortcut connection to LPU (ablation) | LPU · ImageNet-1K | Top-1 accuracy: 82.0% -> 82.7% | Helped |
| Add the inverted residual feed-forward network (ablation) | DeiT-S-4Stage + Stem + LPU · ImageNet-1K | Top-1 accuracy: 82.7% -> 83.3% | Helped |
| Add a shortcut connection to IRFFN (ablation) | IRFFN · ImageNet-1K | Top-1 accuracy: 82.5% -> 83.3% | Helped |
| Add the projection layer (ablation) | CMT block ablation · ImageNet-1K | Top-1 accuracy: 83.3% -> 83.5% | Helped |
| Replace all layer normalization with batch normalization (ablation) | CMT-S · ImageNet-1K | Training convergence | Hurt |
| Replace all batch normalization with layer normalization (ablation) | CMT-S · ImageNet-1K | Top-1 accuracy: 83.5% -> 83.0% | Hurt |
| Scale depth, dimension, and resolution together instead of scaling depth alone (ablation) | CMT-S variants at 9.3B FLOPs · ImageNet-1K | Top-1 accuracy: 83.4% -> 84.5% | Helped |
| Use CMT-S instead of DeiT-S | ImageNet-1K classification · ImageNet-1K | Top-1 accuracy: 79.8% -> 83.5% | Helped |
| Use CMT-S instead of Twins-PCPVT-S | RetinaNet · COCO val2017 | mAP: 1.3% mAP | Helped |
Closest papers by meaning
- An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale (ICLR 2021 · similarity 0.91)
- Do Vision Transformers See Like Convolutional Neural Networks? (NeurIPS 2021 · similarity 0.90)
- ConViT: Improving Vision Transformers with Soft Convolutional Inductive Biases (ICML 2021 · similarity 0.90)
- AsCAN: Asymmetric Convolution-Attention Networks for Efficient Recognition and Generation (NeurIPS 2024 · similarity 0.88)
- Vision Transformers Need Registers (ICLR 2024 · similarity 0.88)
- An Image is Worth More Than 16x16 Patches: Exploring Transformers on Individual Pixels (ICLR 2025 · similarity 0.88)
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.