autoencoders

Definition

Models trained to compress an input into a bottleneck representation and reconstruct it from there, using the input itself as the target so no labels are needed. The bottleneck is the point: whatever survives it is what the model treats as structure. Variants dominate practice — variational autoencoders that learn a distribution, the latent encoders that make high-resolution diffusion tractable, and sparse autoencoders used to decompose activations into interpretable features.