Kan.py ★ Updated

: The library includes specific tools for "symbolic regression," where the model attempts to simplify learned splines into exact mathematical formulas (e.g., turning a learned curve into x2x squared

Supports CPU and GPU, though GPU support may require specific configurations in early versions.

). In a KAN, each connection is a small, learnable spline function ( kan.py

from kan import KAN import torch # Create a KAN with 2 inputs, 5 hidden neurons, and 1 output model = KAN(width=[2, 5, 1], grid=5, k=3) # Training follows a standard loop structure # model.train(dataset, opt="LBFGS", steps=20) Use code with caution. Copied to clipboard

The fundamental shift in KANs is the replacement of fixed linear weights with univariate functions. : The library includes specific tools for "symbolic

: Because the functions are univariate splines, they are easier for humans to visualize and understand, making KANs particularly useful for AI for Science . The pykan Library

: In a standard MLP, a connection is just a single number ( Copied to clipboard The fundamental shift in KANs

: Nodes in a KAN simply sum the incoming signals; they do not have their own activation functions like ReLU or Sigmoid.