KindXiaoming/pykan

Health v2 · maintenance only

Adoption not part of the score

16334 stars · 1559 forks observed · 2026-08-28

What it is AI-extracted, prompt v1, taxonomy v1, 2026-08-29, confidence not recorded

Pykan is a Python library implementing Kolmogorov-Arnold Networks (KANs), an alternative to multi-layer perceptrons with learnable activation functions on edges, built on PyTorch. It emphasizes improved accuracy and interpretability, including symbolic regression and visualization of learned functions.

Use cases

• train interpretable neural networks instead of MLPs
• perform symbolic regression to discover formulas from data
• fit scientific models like PDEs with learnable activation networks
• visualize and prune neural network architectures
• experiment with Kolmogorov-Arnold representation theorem based models

When to choose

• you need more interpretable models than standard MLPs
• you want to extract symbolic formulas from learned networks
• your research explores alternatives to conventional neural architectures
• you work on small-to-medium scientific ML problems runnable on CPU

When to avoid

• you need highly optimized GPU-scale training throughput
• you require a mature ecosystem with broad production tooling
• your workload depends on the standard MLP tooling and hardware acceleration
• you need fast inference at scale, since symbolic branches can be slow

Facets

library · maturity active

machine-learning deep-learning data-science math machine-learning deep-learning data-science python cross-platform kolmogorov-arnold-networks kan pytorch interpretable-ml neural-networks symbolic-regression algorithms

Member repositories

For agents

markdown · JSON · MCP: product_card(name="KindXiaoming/pykan")