Variational autoencoder

A variational autoencoder (VAE) is a type of generative neural network model capable of learning compressed latent representations for data. Key characteristics:

An autoencoder architecture with encoder and decoder networks.
The encoder compresses data into a latent vector.
The decoder reconstructs data from the latent space.
Latent space captures high-level features and variations.
Sampling the space allows generating new data points.
Training involves a reconstruction loss plus a regularization term.

VAEs are useful for:

Dimensionality reduction and feature extraction.
Generating new data similar to training examples.
Applications in image and text generation.

Unlike regular autoencoders, VAEs impose structure on the latent space allowing meaningful interpolation and exploration. VAEs are a foundational technique for representation learning.