Autoencoder: hourglass architecture and latent space holes
Architecture
Latent space & holes
Autoencoder hourglass architecture
Encoder compresses input to a small latent vector z, decoder reconstructs back to original size
Input
D = 16
Hidden
64
Hidden
32
z
dim = 2
Hidden
32
Hidden
64
Output
x̂ ≈ x
ENCODER E(x)
latent z
DECODER D(z)
Minimize: L = ||x − D(E(x))||²
Loss is reconstruction error (MSE)
Click anywhere in latent space to see what the decoder would produce from that point.