## TL;DR
Autoencoders are unsupervised neural networks that learn compressed representations by reconstructing input: Encoder → compressed latent space → Decoder → reconstruction. They learn the most salient features by forcing data through a bottleneck. Applications: dimensionality reduction, denoising, anomaly detection.
## Core Explanation
Basic autoencoder: input = output, loss = reconstruction error (MSE). Denoising autoencoder: corrupt input (add noise), train to reconstruct clean output — learns robust features. Variational Autoencoder (VAE, Kingma & Welling 2013): probabilistic encoder outputs distribution parameters (μ, σ), enabling generation by sampling latent space. VAE-generated images are smoother but blurrier than GANs.
## Further Reading
- [Deep Learning (Goodfellow, Bengio, Courville)](https://www.deeplearningbook.org/)