An autoencoder compresses 8×8 pixel patterns through a 2-D bottleneck and reconstructs them. Watch the reconstructions sharpen and the latent space cluster by shape class.

Originals vs. reconstructions

Top row: inputs. Bottom row: network reconstructions. Each column is a different shape class.

Latent space & loss

Top: 2-D scatter of encoded positions colored by class. Bottom: reconstruction loss curve.