Tutorial: VAE as an inference paradigm for neuroimaging

Variational Autoencoders (VAEs) have emerged as a powerful tool for unsupervised learning, offering a framework to model complex, high-dimensional data through probabilistic inference [1, 2]. Unlike traditional autoencoders, VAEs integrate principles from Bayesian inference, allowing them to generate and reconstruct data by learning latent representations that are both interpretable and continuous [3]. This paradigm has proven particularly valuable in fields dealing with intricate and multidimensional datasets, such as neuroimaging. Neuroimaging data, which includes structural and functional brain scans, often exhibits high dimensionality, noise, and heterogeneity. These characteristics make traditional machine learning approaches prone to overfitting or limited generalization [4]. Moreover, the integration of neuroimaging data with other modalities--such as cognitive assessments, cerebrospinal fluid markers, or genetic information--requires robust generative models capable of capturing complex relationships while preserving interpretability. This challenge is particularly crucial when integrating multimodal datasets, such as combining structural brain scans with genetic or cognitive markers.