Research map / Detection and segmentation / CNN features
A ConvNet for the 2020s
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 the enhanced Transformer-style training recipe | ResNet-50 · ImageNet-1K | top-1 accuracy: 76.1% to 78.8% (+2.7%) | Helped |
| Change stage blocks to (3, 3, 9, 3) (ablation) | ResNet-50 with enhanced recipe · ImageNet-1K | top-1 accuracy: 78.8% to 79.4% | Helped |
| Replace the ResNet stem with a 4×4 stride-4 patchify stem (ablation) | ResNet-50 modernization · ImageNet-1K | top-1 accuracy: 79.4% to 79.5% | No clear effect |
| Adopt depthwise convolution without increasing width (ablation) | ResNet-50 modernization · ImageNet-1K | top-1 accuracy: 79.51 ± 0.18 to 78.28 ± 0.08 | Hurt |
| Increase width after adopting depthwise convolution (ablation) | ResNet-50 modernization · ImageNet-1K | top-1 accuracy: 78.28 ± 0.08 to 80.50 ± 0.02 | Helped |
| Use an inverted bottleneck (ablation) | ResNet-50 modernization · ImageNet-1K | top-1 accuracy: 80.50 ± 0.02 to 80.64 ± 0.03 | Helped |
| Move the depthwise convolution before the pointwise layers (ablation) | ResNet-50 modernization · ImageNet-1K | top-1 accuracy: 80.64 ± 0.03 to 79.92 ± 0.08 | Hurt |
| Increase depthwise kernel size from 3×3 to 7×7 (ablation) | ResNet-50 modernization · ImageNet-1K | top-1 accuracy: 79.9% to 80.6% | Helped |
| Increase depthwise kernel size beyond 7×7 (ablation) | ResNet-50 modernization · ImageNet-1K | top-1 accuracy: 80.57 ± 0.14 to 80.57 ± 0.06 (9×9); 80.47 ± 0.11 (11×11) | No clear effect |
| Replace ReLU with GELU (ablation) | ResNet-50 modernization · ImageNet-1K | top-1 accuracy: 80.57 ± 0.14 to 80.62 ± 0.14 | No clear effect |
| Keep only one GELU activation per block (ablation) | ResNet-50 modernization · ImageNet-1K | top-1 accuracy: +0.7% to 81.3% | Helped |
| Remove two BatchNorm layers per block (ablation) | ResNet-50 modernization · ImageNet-1K | top-1 accuracy: 81.27 ± 0.06 to 81.41 ± 0.09 | Helped |
Closest papers by meaning
- ConvNet vs Transformer, Supervised vs CLIP: Beyond ImageNet Accuracy (ICML 2024 · similarity 0.90)
- ConvNeXt V2: Co-Designing and Scaling ConvNets With Masked Autoencoders (CVPR 2023 · similarity 0.90)
- More ConvNets in the 2020s: Scaling up Kernels Beyond 51x51 using Sparsity (ICLR 2023 · similarity 0.90)
- DECO: Unleashing the Potential of ConvNets for Query-based Detection and Segmentation (ICLR 2025 · similarity 0.89)
- Do Vision Transformers See Like Convolutional Neural Networks? (NeurIPS 2021 · similarity 0.89)
- MogaNet: Multi-order Gated Aggregation Network (ICLR 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.