Reproduce “KAConvNet: Kolmogorov-Arnold Convolutional Networks for Vision Recognition” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Zhaoxiang Liu, Zhicheng Ma, Kaikai Zhao, Kai Wang, Shiguo Lian
KAConvNet proposes a novel architecture that deeply integrates the Kolmogorov-Arnold representation theorem into convolutional layers for vision recognition. Utilizing the Kolmogorov–Arnold Convolutional Layer (KAConvLayer) with a new Grid Linear (GLinear) activation function, it achieves competitive performance on ImageNet-1K with KAConvNet-L reaching 80.1% Top-1 accuracy and demonstrates improved inference speed over previous KAN-based convolutions.
Key insights
- Rigorous integration of the Kolmogorov–Arnold representation theorem into convolutional layers, by distinctly mapping inner and outer functions, leads to more interpretable and effective KAN-CNN architectures.
- The piecewise linear Grid Linear (GLinear) activation function provides superior computational efficiency and better generalization performance than B-splines in KAN-based computer vision models.
- KAConvLayers are effective across all network depths when integrated with residual structures, maintaining stable performance in deep layers and contributing to overall network robustness.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.