Variational Autoencoder

Diederik Kingma & Max Welling, 2014

Introduced by Kingma and Welling in 2014, the Variational Autoencoder (VAE) combines deep learning with Bayesian inference. The encoder compresses input data into a probability distribution in latent space, parameterized by mean (μ) and standard deviation (σ). Using the reparameterization trick (z = μ + σ·ε), it samples from this distribution and the decoder reconstructs the data. The visualization shows the encoder shrinking layers, the latent space as a 2D scatter cloud, and the decoder expanding layers back to the original dimensionality.