Kolmogorov-Arnold Networks (KAN) Guide

Overview

Kolmogorov-Arnold Networks (KANs) are a novel neural network architecture that places learnable activation functions on edges (weights) instead of fixed activations on nodes. Based on the Kolmogorov-Arnold representation theorem, KANs use B-spline functions as learnable edge activations, achieving better accuracy and interpretability than MLPs with fewer parameters in certain domains. This collection tracks the rapidly growing KAN literature.

Core Concept

Traditional MLP:
  x → [fixed activation(linear transform)] → y
  Activations on nodes, weights on edges

KAN:
  x → [learnable spline functions on edges] → sum → y
  Each edge learns its own activation function (B-spline)

Kolmogorov-Arnold Theorem:
  f(x₁,...,xₙ) = Σ Φᵢ(Σ φᵢⱼ(xⱼ))
  Any multivariate continuous function = composition of
  univariate functions and addition

Key Papers

@article{liu2024kan,
  title={KAN: Kolmogorov-Arnold Networks},
  author={Liu, Ziming and Wang, Yixuan and Vaidya, Sachin and
          Ruehle, Fabian and Halverson, James and
          Solja{\v{c}}i{\'c}, Marin and Hou, Thomas Y. and
          Tegmark, Max},
  journal={arXiv:2404.19756},
  year={2024}
}

Implementation

# Using pykan (official implementation)
# pip install pykan

from kan import KAN
import torch

# Create a KAN model
model = KAN(
    width=[2, 5, 1],    # Input: 2, Hidden: 5, Output: 1
    grid=5,               # Spline grid resolution
    k=3,                  # Spline order (cubic)
)

# Training data
x = torch.randn(1000, 2)
y = torch.sin(x[:, 0]) + torch.cos(x[:, 1])
y = y.unsqueeze(1)

# Train
dataset = {"train_input": x[:800], "train_label": y[:800],
           "test_input": x[800:], "test_label": y[800:]}
model.train(dataset, steps=100, lr=0.01)

# Visualize learned functions
model.plot()

# Prune and simplify
model = model.prune()
model.plot()

KAN vs MLP Comparison

# Comparison on function approximation
from kan import KAN
import torch.nn as nn

# KAN: learnable activations on edges
kan_model = KAN(width=[2, 5, 1], grid=5, k=3)
# Parameters: ~150 (spline coefficients)

# MLP: fixed activations on nodes
class MLP(nn.Module):
    def __init__(self):
        super().__init__()
        self.net = nn.Sequential(
            nn.Linear(2, 50),
            nn.ReLU(),
            nn.Linear(50, 50),
            nn.ReLU(),
            nn.Linear(50, 1),
        )
    def forward(self, x):
        return self.net(x)

mlp_model = MLP()
# Parameters: ~2,700

# KAN advantages:
# - Fewer parameters for same accuracy
# - Interpretable (visualize learned functions)
# - Better for scientific discovery (symbolic regression)
# - Grid refinement for progressive accuracy

# MLP advantages:
# - Faster training
# - Better scaling to high dimensions
# - More mature tooling and optimization