KANs 入门

A Beginner-friendly Introduction to Kolmogorov Arnold Networks (KAN)

简介-Understanding Kolmogorov–Arnold Networks (KAN)

• Kolmogorov-Arnold Networks (KANs) are an innovative deep learning architecture proposed in 2024 by Ziming Liu and researchers at MIT.

• They serve as a powerful alternative to traditional Multi-Layer Perceptrons (MLPs) and are inspired by the mathematical Kolmogorov-Arnold Representation Theorem.

How KANs Work vs. MLPs

The fundamental difference lies in where the learning happens.

• Traditional MLPs (Standard Neural Networks): Use fixed activation functions on the nodes (neurons) and train fixed linear weights on the edges (connections).
• KANs: Do not have linear weights on the edges at all. Instead, every connection is assigned a learnable, univariate function (typically a spline). The nodes simply perform summation.

Why the Shift?

By placing the learnable functions on the edges, KANs solve some of the intrinsic limitations of traditional deep learning:

**Higher Interpretability: **

• Since KANs map inputs directly through learnable functions, you can easily visualize and understand what the network is doing at any point, turning the "black box" into a readable mathematical diagram.

**Superior Accuracy: **

• In mathematical and physical data fitting, much smaller KANs frequently outperform massive MLPs.

They scale faster and are more efficient in learning continuous functions.

• If you'd like to dive deeper, we can explore:The math behind the Kolmogorov-Arnold Representation TheoremHow to build or train a KAN using PyTorchSpecific scientific use cases, like solving PDEs or discovering physical laws