Comparing Kolmogorov-Arnold Network Autoencoders versus MLP Autoencoders for the analysis of biomedical data

Abstract Current research extensively utilises deep learning architectures like Multi-Layer Perceptrons and Convolutional Neural Networks. Topologi- cally, these can be viewed as graphs where node functions are learned, and fixed edges facilitate information flow. A novel architecture, Kolmogorov- Arnold Networks, has been proposed, demonstrating improved performance across various applications by incorporating learnable activation functions on network edges. Ongoing research aims to enhance KANs through fea- tures such as dropout regularisation, Autoencoders, model benchmarking, and the development of KAN Convolutional Networks for matrix convo- lution. This study compares the performance of standard Autoencoders with their Kolmogorov-Arnold counterparts, which possess an equivalent or smaller parameter count, using cardiologic performance signals as input. Specifically, some classic AE tasks such as reconstruction, , denoising, and inpainting, were evaluated on the AbnormalHeartbeat dataset, which comprises audio signals recorded via stethoscope.