Kolmogorov-Arnold Network (KAN) - Analysis Project

Description

This project implements and analyzes the Kolmogorov-Arnold Network (KAN), an innovative type of neural network based on the Kolmogorov-Arnold theorem. The project includes practical implementations and comparisons with other machine learning approaches.

Requirements

• Python 3.9.6
• Main dependencies:

matplotlib==3.6.2
numpy==1.24.4
scikit_learn==1.1.3
torch==2.2.2
pandas==2.0.1

Installation

1. Clone the repository:

git clone https://github.com/xtreamsrl/KAN
cd KAN

2. Install dependencies:

pip install -r requirements.txt

Project Structure

Main Notebooks

1. kan_presentation.ipynb: A comprehensive introduction to the Kolmogorov-Arnold Network (KAN) and its applications
2. kaggle_vs_kan.ipynb: Demonstration of using KAN to solve a Kaggle competition

Key Features

• KAN implementation using PyTorch
• Network simplification techniques:
  • Sparsification
  • Refinement
  • Network pruning
• Learning visualization through animations
• Comparison with traditional neural networks

Core Functionality

Model Creation

model = KAN(width=[2,5,1], grid=3, k=3, device=device)

Training

_ = model.fit(dataset, opt="LBFGS", steps=150)

Network Optimization

model = model.prune()  # Pruning
model = model.refine(10)  # Refinement

Visualizations

The project includes functionality to generate dynamic visualizations of network learning, saved as MP4 videos.

References

For in-depth tutorials and video content, visit:

https://www.youtube.com/@umarjamilai/videos

Notes

• Project is optimized for both CPU and CUDA GPU execution
• Includes implementations for both regression and classification problems
• Provides tools for comparative analysis with other machine learning models

License

This project is distributed under an open source license. See LICENSE file for details.