Back to Glossary
Variational Autoencoder (VAE)
Models & Architectures
Probabilistic autoencoder for generative modeling.
VAEs model a probability distribution in latent space, allowing new samples to be drawn.
- Core idea: Evidence Lower Bound (ELBO) combining reconstruction and KL divergence terms.
- Strengths: Smooth latent space, probabilistic interpretation.
- Weaknesses: May produce blurrier samples than GANs.