# Posterior Collapse

> Source: https://sukruyusufkaya.com/en/glossary/posterior-collapse
> Updated: 2026-05-13T20:58:21.856Z
> Type: glossary
> Category: derin-ogrenme
**TLDR:** A VAE training issue in which the decoder ignores the latent variable, weakening representation learning.

<p>Posterior collapse can make latent space effectively meaningless, especially in variational autoencoders with strong decoders. If the model learns to generate outputs without relying on latent information, the latent representation may collapse under KL regularization pressure. This is a serious issue for generative quality, representation learning, and controllable sampling. A significant portion of VAE research has focused on mitigating this problem.</p>