autoencoder

Definition

A model trained to reproduce its own input through a narrow bottleneck: an encoder compresses to a latent code, a decoder reconstructs from it, and reconstruction error is the loss. Because no labels are needed, autoencoders serve for dimensionality reduction, for anomaly detection — inputs that reconstruct badly are flagged — and for learning latent spaces that generative systems operate in. A plain autoencoder is not a VAE, which adds a probabilistic prior aimed at sampling.