More ConvNets in the 2020s: Scaling up Kernels Beyond 51x51 using Sparsity

TL;DR

This paper demonstrates that by applying sparsity techniques, convolutional neural networks can effectively scale to extremely large kernels beyond 51x51, achieving performance comparable to or better than state-of-the-art transformers across multiple vision tasks.

Contribution

The paper introduces a sparsity-based recipe for training extremely large kernels in CNNs, enabling kernels up to 61x61 and proposing the Sparse Large Kernel Network (SLaK) architecture.

Findings

SLaK with sparse 51x51 kernels matches or exceeds transformer performance.

Scaling kernels beyond 51x51 improves accuracy across vision tasks.

Sparse large kernels enable CNNs to compete with or outperform modern transformers.

Abstract

Transformers have quickly shined in the computer vision world since the emergence of Vision Transformers (ViTs). The dominant role of convolutional neural networks (CNNs) seems to be challenged by increasingly effective transformer-based models. Very recently, a couple of advanced convolutional models strike back with large kernels motivated by the local-window attention mechanism, showing appealing performance and efficiency. While one of them, i.e. RepLKNet, impressively manages to scale the kernel size to 31x31 with improved performance, the performance starts to saturate as the kernel size continues growing, compared to the scaling trend of advanced ViTs such as Swin Transformer. In this paper, we explore the possibility of training extreme convolutions larger than 31x31 and test whether the performance gap can be eliminated by strategically enlarging convolutions. This study ends up with a recipe for applying extremely large kernels from the perspective of sparsity, which can smoothly scale up kernels to 61x61 with better performance. Built on this recipe, we propose Sparse Large Kernel Network (SLaK), a pure CNN architecture equipped with sparse factorized 51x51 kernels that can perform on par with or better than state-of-the-art hierarchical Transformers and modern ConvNet architectures like ConvNeXt and RepLKNet, on ImageNet classification as well as a wide range of downstream tasks including semantic segmentation on ADE20K, object detection on PASCAL VOC 2007, and object detection/segmentation on MS COCO.