Research map / Architectures and efficiency / CNN design
PeLK: Parameter-efficient Large Kernel ConvNets with Peripheral Convolution
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 peripheral convolution, a parameter-sharing dense large-kernel operator inspired by human peripheral vision, and builds PeLK, a pure CNN scaling kernels up to 101x101 with far fewer parameters.
Problem
Dense large-kernel convolutions grow quadratically in parameters and suffer optimization difficulties, so prior CNNs fell back to stripe convolutions (up to 51x51) that saturate as kernel size grows.
Approach
Parameterize a large dense kernel with a small kernel via spatial parameter sharing: fine-grained weights in the center and exponentially growing sharing granularity in the periphery (O(K^2) to O(log K) parameters), plus kernel-wise positional embedding to restore local detail and partial channel processing.
Key results
- PeLK-B reaches 84.2% top-1 on ImageNet-1K at 224x224 vs. SLaK-B 84.0% and ConvNeXt-B 83.8%; PeLK-B-101 reaches 85.8% at 384x384.
- PeLK-B achieves 50.4 mIoU on ADE20K vs. SLaK-B 50.2 and RepLKNet-B 49.9 with fewer parameters (126M vs 131M).
- Peripheral convolution reduces parameters of dense convolution by over 90%, and PeLK-T has 1.5x higher throughput than SLaK-T (1138 vs 754 img/s on A100).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Replace stripe convolution with dense grid convolution at 51×51 (ablation) | SLaK framework · ADE20K | mIoU (%): 46.1 -> 46.6 (+0.5) | Helped |
| Use dense grid convolution instead of stripe convolution (ablation) | SLaK framework · ImageNet-1K | accuracy: 81.3 -> 81.6 | Helped |
| Use PeLK-T instead of ConvNeXt-T | ImageNet-1K, 224² · ImageNet-1K | top-1 accuracy: 82.1 -> 82.6 | Helped |
| Use PeLK-S instead of ConvNeXt-S | ImageNet-1K, 224² · ImageNet-1K | top-1 accuracy: 83.1 -> 83.9 | Helped |
| Use PeLK-B instead of SLaK-B | ImageNet-1K, 224² · ImageNet-1K | top-1 accuracy: 84.0 -> 84.2 | Helped |
| Use PeLK-B-101 instead of SLaK-B | ImageNet-1K, 384² · ImageNet-1K | top-1 accuracy: 85.5 -> 85.8 | Helped |
| Use PeLK-T instead of SLaK-T | ADE20K · ADE20K | mIoU (%): 47.6 -> 48.1 | Helped |
| Use PeLK-B-101 instead of ConvNeXt-B | COCO · COCO | APmask: 45.6 -> 46.1 | Helped |
| Keep the central 5×5 region unshared rather than sharing across the central region (ablation) | PeLK-T · ImageNet-1K, 224² | top-1 accuracy: 81.0 -> 81.4 | Helped |
| Use an exponentially increasing sharing grid rather than a non-exponential grid (ablation) | PeLK-T · ImageNet-1K, 224² | top-1 accuracy: 81.3 -> 81.4 | Helped |
| Set the central fine-grained kernel size to 1×1 instead of 5×5 (ablation) | PeLK-T · ImageNet-1K, 224² | top-1 accuracy: 81.6 -> 80.8 | Hurt |
| Increase the central fine-grained kernel size from 5×5 to 7×7 (ablation) | PeLK-T · ImageNet-1K, 224² | top-1 accuracy: 81.6 -> 81.6 | No clear effect |
Surprising findings
- Dense grid convolution outperforms stripe convolution consistently across kernel sizes, with gains growing as kernel size increases.
- Scaling the kernel to 101x101 at 224x224 input gives no extra benefit, and 151x151 degrades performance, which the authors attribute to the stem's 4x downsampling making large kernels cover the global feature map.
Limitations
- Very large kernels saturate or degrade when kernel size exceeds the input-resolution-dependent global coverage (e.g., 151x151).
- Benefits on ImageNet classification are modest; the authors note very large dense kernels are not designed for classification.
- Relies on ImageNet-scale training with dynamic sparsity and high-resolution pretraining for the largest kernels.
Experimental setup
Datasets: ImageNet-1K ADE20K MS COCO
Benchmarks: ImageNet-1K ADE20K MS COCO
Metrics: top-1 accuracy mIoU box AP mask AP params FLOPs throughput
Baselines: Swin ConvNeXt RepLKNet SLaK PerViT T2T-ViT DeiT PVT
Builds on: SLaK RepLKNet ConvNeXt UperNet Cascade Mask R-CNN
Closest papers by meaning
- Scaling Up Your Kernels to 31x31: Revisiting Large Kernel Design in CNNs (CVPR 2022 · similarity 0.90)
- UniRepLKNet: A Universal Perception Large-Kernel ConvNet for Audio Video Point Cloud Time-Series and Image Recognition (CVPR 2024 · similarity 0.86)
- KernelWarehouse: Rethinking the Design of Dynamic Convolution (ICML 2024 · similarity 0.86)
- CAM Back Again: Large Kernel CNNs from a Weakly Supervised Object Localization Perspective (CVPR 2024 · similarity 0.86)
- InceptionNeXt: When Inception Meets ConvNeXt (CVPR 2024 · similarity 0.85)
- Recurrent Partial Kernel Network for Efficient Optical Flow Estimation (AAAI 2024 · similarity 0.85)
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.