# Reparameterization Trick

> Source: https://sukruyusufkaya.com/en/glossary/reparameterization-trick
> Updated: 2026-05-13T20:00:57.813Z
> Type: glossary
> Category: derin-ogrenme
**TLDR:** A core VAE technique that makes latent-variable models with stochastic sampling differentiable.

<p>The reparameterization trick is one of the key ideas that makes training variational autoencoders possible. Because direct stochastic sampling in latent space is not differentiable, the sampling step is rewritten as a deterministic transformation plus independent noise. This allows backpropagation to flow through stochastic latent layers. It is a landmark idea in the history of deep generative modeling.</p>