Benchmarking Variational Autoencoders for Dimensionality Reduction and Application to Medical Heartbeats Data

This report explores the implementation of Variational Autoencoders (VAEs) for dimensionality reduction in multivariate statistics. We benchmark VAEs against traditional methods such as PCA, ICA, and KPCA, using datasets including Iris Flowers, Wine, and Breast Cancer. The study further extends the application of VAEs to non-columnar data, specifically medical heartbeat audio recordings. The results demonstrate the strengths and limitations of VAEs in clustering, manifold learning, and anomaly detection, particularly in the context of medical data. Our findings suggest that while VAEs offer flexible non-linear representations, their performance is highly sensitive to the type of activation function used and the nature of the data.