Technical GlossaryDeep Learning

Variational Autoencoder

A generative autoencoder architecture that models the latent space probabilistically for both representation and generation.

A variational autoencoder extends the classical autoencoder idea with a probabilistic generative perspective. Instead of learning fixed latent points, it learns latent distributions, which makes it possible to generate new samples. Because it combines representation learning with generative modeling, it occupies a special place in deep learning history. It is especially important for understanding latent-space geometry and performing controlled sampling.