Justification for variational autoencoders over basic autoencoders

Determine why a probabilistic variational autoencoder should be preferred to a basic autoencoder for unsupervised feature extraction.

Background

The paper discusses autoencoders as a longstanding approach to unsupervised feature extraction and contrasts basic autoencoders with variational autoencoders, which provide a probabilistic framework. It explicitly identifies as unresolved the rationale for preferring the probabilistic variational formulation over a basic autoencoder in settings where both can be used to extract features.

References

However, it is often not clear why such a probabilistic VAE should be preferred to a very basic autoencoder.

— What is the goal of unsupervised machine learning?  (2610.11697 - Hyvärinen, 8 Oct 2026) in Section “Goal 3: Extracting features for downstream tasks,” subsection “Methods”