Autoencoder — Encoder-Decoder & Latent Space

Compress high-dim data to a bottleneck, then reconstruct. Interpolate in latent space.

Controls

Encoder: Input → latent z (bottleneck)
Decoder: z → reconstruction x̂

Loss: ||x - x̂||²

Interpolation:
z_α = (1-α)·z_A + α·z_B
x̂_α = Decoder(z_α)

Latent space is smooth — linear interpolation gives meaningful outputs.