Reproduce “PolyKAN: Efficient Fused GPU Operators for Polynomial Kolmogorov-Arnold Network Variants” 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

Mingkun Yu, Heming Zhong, Dan Huang, Yutong Lu, Jiazhi Jiang

PolyKAN introduces efficient fused GPU operators for Polynomial Kolmogorov–Arnold Network (KAN) variants, achieving up to 12.6x faster training and 14.8x faster inference on an A100 GPU compared to baseline implementations. This optimization enables practical deployment of KANs, maintaining or improving accuracy and convergence speed, particularly beneficial for scientific applications.

Key insights

  • A systematic combination of LUT-based evaluation, 2D tiling, two-stage reduction, and coefficient layout reordering provides a generalized, variant-agnostic optimization framework for KANs.
  • Approximating gradients through LUT-based interpolation can unexpectedly lead to faster convergence rates by producing a smoother gradient landscape, acting as an implicit regularizer.
  • Bridging the gap between novel, theoretically advantageous neural network architectures and high-performance deployment requires specialized, hardware-aware system design that targets computational and memory bottlenecks.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.