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

Kolmogorov-Arnold Networks (KANs) extend classical neural architectures by replacing fixed activation functions with learnable univariate transformations on network edges, yielding a function-theoretic alternative to Multi-Layer Perceptrons. To evaluate their suitability for unsupervised biomedical signal modelling, we compare fully connected and convolutional autoencoders with parameter-matched Kolmogorov-Arnold counterparts across reconstruction, denoising and inpainting tasks using stethoscope-derived cardiologic signals from the AbnormalHeartbeat dataset. Convolutional variants substantially outperform dense architectures, reflecting the importance of local receptive fields in capturing temporal structure. Within this class, KAN-based convolutional autoencoders (KCAE, KCAE-PS) consistently achieve the lowest test MSE and exhibit superior robustness to noise and missing segments while maintaining reduced parameter counts. PixelShuffle-enhanced KAN models provide the most favourable accuracy-efficiency trade-off, although KAN layers introduce significant computational overhead due to the cost of spline-based functional evaluations. These results demonstrate that Kolmogorov-Arnold parametrisations can enhance the expressive capacity and compactness of convolutional autoencoders for biomedical time-series analysis, while also delineating current performance bottlenecks for large-scale or real-time deployment.