Autoencoder

An unsupervised neural network that compresses input data into a low-dimensional bottleneck to learn efficient, non-linear representations.

Autoencoders utilize a symmetric architecture (an encoder and a decoder) to map data into a latent-space representation. This bottleneck forces the model to capture only the most salient features while discarding noise. Practitioners deploy specific variants like Denoising Autoencoders (DAE) to restore corrupted signals or Variational Autoencoders (VAE) for complex generative modeling. In 2006, Geoffrey Hinton proved these networks could outperform traditional Principal Component Analysis (PCA) for dimensionality reduction. Most implementations optimize the reconstruction loss (often Mean Squared Error) to ensure the output matches the input as closely as possible.

https://www.deeplearningbook.org/contents/autoencoders.html