KAN Experiments

These experiments were carried out for a course project as part of the Machine Learning for Signal Processing course, taught by Sriram Ganapathy at the Indian Institute of Science, Bengaluru.

Prior Work

The work builds on

Papers

• KAN: Kolmogorov-Arnold Networks
• KAN 2.0: Kolmogorov-Arnold Networks Meet Science
• Convolutional Kolmogorov-Arnold Networks
• LSS-SKAN: Efficient Kolmogorov-Arnold Networks based on Single-Parameterized Function
• LArctan-SKAN: Simple and Efficient Single-Parameterized Kolmogorov-Arnold Networks using Learnable Trigonometric Function

Libraries

• KindXiaoming/pykan
• AntonioTepsich/Convolutional-KANs
• chikkkit/skan_library
• chikkkit/LSS-SKAN
• chikkkit/LArctan-SKAN

Work done

We introduce the multiplicative nodes introduced in the KAN2.0 paper and the pykan library to the single-parametrized non-linear functions in SKAN, forming MultSKAN.

We also introduce a simplified form of convolution to the above, forming ConvMultSKAN.

Experiments

Basic experimentation done for MNIST and CIFAR-10 datasets.

Contact

Purva Parmar, MTech AI 2024-26
purvaparmar@iisc.ac.in