Deep Learning

Autoencoder & Latent Space

Force a picture through a gap of two numbers and make the network rebuild it from those two alone. Nothing says what the numbers should mean, and yet the only way to pass is to lay every picture out on a map.

the point you are decodingone training image

Drag anywhere on the map, including the empty parts

Reconstruction error—

—

Numbers in the gap 2

onefour

Pictures

Numbers

reconstruction error—

training steps—

compression—

best with 1 number—

best with 2—

best with 3—

What to observe

Train on digits. Nobody says what a digit is, yet similar pictures land near each other: rebuilding from two numbers forces a map.
Drag into the empty space between clusters. Nothing in the loss covers those points, so what comes out there is a blend or junk: a plain autoencoder's map has holes. A variational autoencoder adds noise to the codes and pulls them towards one smooth cloud, which is what lets you sample from it.
With 1 number everything lies on a line and the error jumps; each extra number helps less. The panel keeps the best error per width.
On blobs only a spot's position varies, so two numbers suffice: the data really is two-dimensional, seeThe Data Manifold.

Shortcuts:space run/pause · r reset · f fullscreen