Research map / Architectures and efficiency / CNN design
Revealing the Dark Secrets of Extremely Large Kernel ConvNets on Robustness
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.
Evaluates robustness of large-kernel ConvNets (RepLKNet, ConvNeXt) against small-kernel CNNs and ViTs across six ImageNet robustness benchmarks, and analyzes occlusion invariance, adversarial attacks, lesion robustness, frequency noise, and kernel attention to explain the robustness.
Problem
Whether large-kernel convolutional networks are robust to distribution shifts and corruptions, how their robustness compares to typical CNNs and ViTs, and what underlies it.
Approach
Shows that pure CNNs with extremely large kernels (RepLKNet-31B) achieve robustness comparable to or better than ViTs across diverse benchmarks, and attributes this to occlusion invariance, stable feature variance, block-removal tolerance, frequency robustness, and mixed local/global kernel attention, with kernel size ablations showing kernel size is the key factor.
Key results
- RepLKNet-31B reaches 36.5% mCE on ImageNet-C, versus 45.5% for ViT-L/16 and 76.7% for ResNet-50 (lower is better).
- On ImageNet-A, RepLKNet-31B's top-1 accuracy is 2.8x that of BiT-m r101x3 and slightly surpasses ViT-L.
- Replacing RepLKNet-31B's large kernels with 3x3 kernels (RepLKNet-3B) drops ImageNet-A accuracy from 29.4 to 14.6 and ImageNet-C from 49.2 to 39.1.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| increase ConvNeXt-Tiny kernel size from 3×3 to 13×13 (ablation) | ConvNeXt-Tiny · ImageNet-A | top-1 accuracy: 5.20 -> 9.89 | Helped |
| increase ConvNeXt-Tiny kernel size from 3×3 to 51×51 (ablation) | ConvNeXt-Tiny · ImageNet-A | top-1 accuracy: 5.20 -> 10.71 | Helped |
| increase ConvNeXt-Tiny kernel size from 3×3 to 51×51 (ablation) | ConvNeXt-Tiny · ImageNet-R | top-1 accuracy: 28.87 -> 31.77 | Helped |
| replace RepLKNet-31B 31×31 kernels with 3×3 kernels (ablation) | RepLKNet-31B · ImageNet-A | top-1 accuracy: 29.4 -> 14.6 | Hurt |
| replace RepLKNet-31B 31×31 kernels with 3×3 kernels (ablation) | RepLKNet-31B · ImageNet-C | top-1 accuracy: 49.2 -> 39.1 | Hurt |
| replace RepLKNet-31B 31×31 kernels with 3×3 kernels (ablation) | RepLKNet-31B · ImageNet-R | top-1 accuracy: 43.9 -> 32.9 | Hurt |
| replace RepLKNet-31B 31×31 kernels with 3×3 kernels (ablation) | RepLKNet-31B · ImageNet-O | AUPR: 78.9 -> 51.6 | Hurt |
| replace RepLKNet-31B 31×31 kernels with 3×3 kernels (ablation) | RepLKNet-31B · Salient-Drop-50% | accuracy drop: -25.1 -> -61.8 | Hurt |
| replace RepLKNet-31B 31×31 kernels with 3×3 kernels (ablation) | RepLKNet-31B · Noise-0.7π | accuracy drop: -5.0 -> -10.6 | Hurt |
| use larger kernels | RepLKNet-31B · ImageNet-C | mCE (%): 58.3 -> 36.5 | Helped |
| use large-kernel RepLKNet instead of ViT-L/16 | ImageNet-P · mFR (%) | mFR: 33.1 -> 29.2 | Helped |
| use large-kernel RepLKNet instead of ViT-L/16 | ImageNet-P · mT5D (%) | mT5D: 50.2 -> 49.8 | Helped |
Surprising findings
- Reducing kernel size from 31 to 3 makes RepLKNet fall below ViT-B on robustness metrics despite similar pre-training.
- Scaling kernel size in ConvNeXt-T up to 51 improves robustness even though ImageNet accuracy gains saturate.
Limitations
- No theoretical proof of the observed robustness mechanisms.
- Kernel size ablation done only on ImageNet-1K, not ImageNet-21K, due to compute.
- Evaluated mainly on a single large-kernel architecture family (RepLKNet) with ConvNeXt as secondary.
Experimental setup
Datasets: ImageNet-21K ImageNet-1K ImageNet-A ImageNet-C ImageNet-O ImageNet-P ImageNet-R ImageNet-9
Benchmarks: ImageNet-A ImageNet-C ImageNet-O ImageNet-P ImageNet-R ImageNet-9 ImageNet
Metrics: top-1 accuracy mCE AUPR mFR mT5D attack success rate accuracy drop under noise
Baselines: ResNet-50 BiT ViT-B/16 ViT-L/16 ConvNeXt AugMix DeepAugment
Builds on: RepLKNet ConvNeXt Vision Transformer
Closest papers by meaning
- More ConvNets in the 2020s: Scaling up Kernels Beyond 51x51 using Sparsity (ICLR 2023 · similarity 0.89)
- Are Large Kernels Better Teachers than Transformers for ConvNets? (ICML 2023 · similarity 0.89)
- Towards Robust Vision Transformer (CVPR 2022 · similarity 0.88)
- Can CNNs Be More Robust Than Transformers? (ICLR 2023 · similarity 0.88)
- Vision Transformers Beat WideResNets on Small Scale Datasets Adversarial Robustness (AAAI 2025 · similarity 0.87)
- UniRepLKNet: A Universal Perception Large-Kernel ConvNet for Audio Video Point Cloud Time-Series and Image Recognition (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.