Research map / Architectures and efficiency / Attention and transformers
Designing Concise ConvNets with Columnar Stages
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes CoSNet, a concise ConvNet built from columnar stages with parallel narrow convolutions fed by input replication, minimal 1x1 layers, and a single uni-branch structure, evaluated on ImageNet against ResNets, RepVGG, VanillaNet, ViTs and others.
Problem
Designing ConvNets that are simultaneously shallow, low in parameters and FLOPs, low-branching and fast on parallel hardware, for resource-constrained deployment, without relying on attention.
Approach
A CoSNet-unit uses two 1x1 convolutions around M parallel columns of 3x3 convolutions with N kernels each, fed by input replication (not channel splitting), with columns batched into one uni-branch computation; fusion happens once at the end, plus shallow and deep projections.
Key results
- CoSNet-B2 reaches 81.3% top-1 on ImageNet with 30.0M params, 5.10B FLOPs and depth 26, versus 81.8% for ConvNeXt-T at 29.0M params and depth 59.
- CoSNet-C2 reaches 82.1% top-1 with depth 26, matching ConvNeXt-V2-T-level accuracy with higher FPS (90 vs 62).
- CoSNet-B1 trains 300 epochs in about 25 hours versus 40 hours for VanillaNet-6, with 79.50% top-1.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Increase the number of parallel convolutions M from 1 to 4 (ablation) | CoSNet ablation with N=16/32/64/128 · ImageNet validation | Top-1 accuracy: 74.45 -> 75.65 | Helped |
| Increase the number of parallel convolutions M from 4 to 5 (ablation) | CoSNet ablation with N=16/32/64/128 · ImageNet validation | Top-1 accuracy: 75.65 -> 76.01 | Helped |
| Increase the number of kernels N from 16/32/64/128 to 32/64/128/256 at M=1 (ablation) | CoSNet ablation · ImageNet validation | Top-1 accuracy: 74.45 -> 75.37 | Helped |
| Use 5 parallel columnar convolutions instead of 4 (ablation) | CoSNet R1/R2 · ImageNet validation | Top-1 accuracy: 75.65 -> 76.01 | Helped |
| Use multiple PCCs rather than the R3 configuration (ablation) | CoSNet ablation · ImageNet validation | Top-1 accuracy: 75.37 -> 75.65 | Helped |
| Use the shallower R2 configuration instead of deeper R9 (ablation) | CoSNet ablation · ImageNet validation | Top-1 accuracy: 75.88 -> 76.01 | Helped |
| Add residual connections within PCC (ablation) | CoSNet shallow and deep ablations · ImageNet validation | Top-1 accuracy: +0.09%; +0.70% | Helped |
| Add deep projections (ablation) | CoSNet R10 versus R11 · ImageNet validation | Top-1 accuracy: 73.61 -> 74.15 | Helped |
| Add pooling to deep projections (ablation) | CoSNet R11/R12 · ImageNet validation | Top-1 accuracy: 1.22% | Helped |
| Restrict N to a very small value in deeper layers (ablation) | CoSNet R14 · ImageNet validation | Top-1 accuracy: 76.76 -> 71.20 | Hurt |
| Use pairwise frequent fusion | CoSNet-B2 · ImageNet validation | Top-1 accuracy: 81.3 -> 82.7 | Helped |
| Replace the ResNet-50 backbone with CoSNet-C2 | DN-DETR · MS-COCO | AP: 38.3 -> 39.2 | Helped |
Surprising findings
- Adding a 1x1 fusion after each column (group-conv style) reduced accuracy by about 1%, suggesting frequent fusion is unnecessary in this design.
- Parameters and FLOPs alone do not guarantee lower latency; EfficientNet-B0 has fewer params and FLOPs but runs about 37% slower than CoSNet-like models.
Limitations
- Results are reported mainly against models trained with differing batch sizes and recipes, requiring extra retraining for fair comparison.
- Latency measurements are at batch size 1 on specific hardware and the paper does not focus on the mobile regime.
Experimental setup
Datasets: ImageNet-1K MS-COCO
Benchmarks: ImageNet-1K MS-COCO ADE20K
Metrics: Top-1 accuracy FLOPs parameters latency FPS AP mIoU
Baselines: ResNet-18 ResNet-50 ResNeXt-50 RepVGG VanillaNet ParNet DeiT-S Swin-T ConvNeXt-T EfficientViT-M4
Builds on: VanillaNet ResNet ResNeXt RepVGG ParNet
Closest papers by meaning
- Are Large Kernels Better Teachers than Transformers for ConvNets? (ICML 2023 · similarity 0.85)
- LSNet: See Large, Focus Small (CVPR 2025 · similarity 0.84)
- InceptionNeXt: When Inception Meets ConvNeXt (CVPR 2024 · similarity 0.84)
- UniRepLKNet: A Universal Perception Large-Kernel ConvNet for Audio Video Point Cloud Time-Series and Image Recognition (CVPR 2024 · similarity 0.84)
- MogaNet: Multi-order Gated Aggregation Network (ICLR 2024 · similarity 0.83)
- ConvNet vs Transformer, Supervised vs CLIP: Beyond ImageNet Accuracy (ICML 2024 · similarity 0.83)
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.