Continual learning in KANs

This repository contains the codes for the paper A preliminary study on continual learning in computer vision using Kolmogorov-Arnold Networks. We investigate the ability of Kolmogorov-Arnold Networks (KANs) to deal with computer vision tasks in a class-incremental learning scenario.

KANs were presented by Liu and colleagues in their work KAN: Kolmogorov-Arnold Networks.

---

📙 The slideshow presented on the day of the exam:

KANs Continual Learning [Slideshow PPTX] - Morelli Valerio Paganica Federica.pptx

KANs Continual Learning [Slideshow PDF] - Morelli Valerio Paganica Federica.pdf

KANs Continual Learning [Thesis PDF] - Morelli Valerio Paganica Federica.pdf

---

📘 Table of Contents

• ⬇️ Class-IL Scenario
  • MLPs vs KANs
  • KAN-based and non-KAN-based convolutional nets
• ❗ The Gaussian Peaks Problem
• 👨🏻‍💻 Authors

⬇️ Class-IL Scenario

MLP vs PyKAN vs EffKAN

🎬 The following videos highlight the difference between MLP, PyKAN (PyKAN), and EffKAN (EfficientKAN) in a Class-IL scenario on the MNIST dataset. Each video shows the per-epoch predicitons of the corresponding model in the optimal hyper-parameter configuration.

The following test accuracy plots show the same trainin runs as the confusion matrices.

KAN-based and non-KAN-based convolutional nets

Based on Convolutional-KANs by AntonioTepsich.

❗ The Gaussian Peaks Problem

Here we show how the can be solved by EfficientKAN with the same performance as PyKAN.

Read more on Something different from the official results for KAN_

After introducing the sb_trainable and sp_trainable on the EfficientKAN class, and setting them to False just like PyKAN does, the same results can be achieved:

👨🏻‍💻 Authors

Name	Email	GitHub
Valerio Morelli	s1118781@studenti.univpm.it	MrPio
Federica Paganica	s1116749@studenti.univpm.it	Federica
Alessandro Cacciatore	a.cacciatore1@unimc.it	geronimaw