Galaxy Morphology Classification Using Light-Weight Convolutional Kolmogorov Arnold Network

Ang, Yu Shien (2025) Galaxy Morphology Classification Using Light-Weight Convolutional Kolmogorov Arnold Network. Final Year Project (Bachelor), Tunku Abdul Rahman University of Management and Technology.

Text Ang Yu Shien - FULL TEXT.pdf Restricted to Registered users only Download (2MB)

Abstract

The Galaxy Morphology Classification using deep learning is gaining more attention due to its effectiveness in analysing the galaxy image accurately. However, deep learning models typically are always designed to be complex and large in order to achieve satisfying performance. This work proposes a knowledge distillation based Convolutional Kolmogorov-Arnold Network, a simple and effective method to perform galaxy morphology classification. In our proposed method, pioneering convolutional KAN with attention module is proposed and designed for performing galaxy morphology classification. Then, knowledge distillation is applied to compress the model, at the same time achieving an average testing accuracy exceeding 95%. To further understand how the model predict, popular explainable artificial intelligence model, namely Local Interpretable Model-Agnostic Explanations is adopted for interpretation and visualisation purpose. Notably, our proposed framework outperforms other models, such as CapsuleNet, ConViT, MobileViT, and ResNeXt, in terms of average accuracy. This research demonstrates the effectiveness of knowledge distillation in improving model performance and addressing real-world challenges, such as handling imbalanced class 2 and 3 galaxy morphology datasets and model efficiency in recognizing galaxy features.