Autoencoder

Autoencoder is a type of artificial neural network trained to copy its input to its output, typically through a compressed intermediate representation called a bottleneck or latent space. The architecture is symmetric: an encoder maps input to a lower-dimensional representation, and a decoder reconstructs the input from that representation. By penalizing reconstruction error, the network learns to capture the most salient features of the input distribution.

Autoencoders are foundational to modern representation learning. Variants such as denoising autoencoders and contractive autoencoders impose different constraints on the latent space, enabling applications from dimensionality reduction to generative modeling. The variational autoencoder in particular introduced a probabilistic framing that bridges neural networks and Bayesian inference, though it remains an open question whether the latent spaces learned by autoencoders correspond to meaningful generative factors or merely to compressions that happen to be useful for downstream tasks.