Autoencoders

Autoencoders learn a compact or structured latent representation by encoding data and training a decoder to reconstruct it, supporting dimensionality reduction, compression, and feature learning without labels. Common approaches include convolutional, recurrent, and transformer autoencoders; variational and sparse variants; and models with geometric, physical, or domain-specific constraints. They are used on images, audio and video, scientific simulations, medical data, and industrial time series, with evaluation ranging from reconstruction and compression to anomaly detection and downstream prediction. A central tension is that latents optimized for faithful reconstruction are not always useful for generation, interpretation, or stable long-horizon prediction, so current work increasingly evaluates and shapes representations for their intended downstream use.