Autoencoders & VAEs

An autoencoder is two networks end-to-end: an encoder squeezes the input into a tiny latent vector, a decoder rebuilds the input from it. Train on reconstruction loss — 'make the output match the input' — and the bottleneck forces the encoder to keep only what matters.

You can't rebuild a face from 32 numbers unless those numbers capture pose, lighting, and expression. The latent space emerges from the squeeze.