What is Variational Autoencoder (VAE) architecture?

VAE is a probabilistic autoencoder that learns latent distributions instead of fixed vectors.

VAEs encode inputs into mean and variance vectors defining a Gaussian distribution. A latent vector is sampled using reparameterization trick (z = μ + σ * ε). The model optimizes reconstruction loss plus KL divergence to enforce smooth latent space. This allows generative sampling and interpolation.