Variational Autoencoders and the Variable Collapse Phenomenon

In Variational Autoencoders, when working in high-dimensional latent spaces, there is a natural collapse of latent variables with minor significance, that get altogether neglected by the generator. We discuss this known but controversial phenomenon, sometimes referred to as overpruning, to emphasize the under-use of the model capacity. In fact, it is an important form of self-regularization, with all the typical benefits associated with sparsity: it forces the model to focus on the really important features, enhancing their disentanglement and reducing the risk of overfitting. In this article, we discuss the issue, surveying past works, and particularly focusing on the exploitation of the variable collapse phenomenon as a methodological guideline for the correct tuning of the model capacity, and of the loss function parameters.