a ([[function|deterministic]]) [[1991KramerNonlinearPrincipalComponent|autoencoder]] is a mapping $\text{dec}[\theta_{\text{dec}}] : x \mapsto z$ and $\text{enc}[\theta_{\text{enc}}] : z \mapsto x$ where $\dim \mathcal{Z} \ll \dim \mathcal{X}$.
the task is
$\max_{ \theta_{\text{enc}}, \theta_{\text{dec}} } d(x, \text{dec}(\text{enc}(x)))$
where $d$ is some [[loss function]]
ie measurement of distance (eg [[metric]]).
does ([[unsupervised]]) "nonlinear [[principal component]] analysis".
can also be adjusted without [[backpropagation]] by using the [[recirculation algorithm]] (see [[biologically plausible]])
in [[neural gradient representation by activity differences]] framework, backpropagate activations from higher layers
see [Lilian Weng's blog post](https://lilianweng.github.io/posts/2018-08-12-vae/) for more details