Autoencoder
Compress inputs through a tiny bottleneck and watch genuine reconstructions improve.
A tiny autoencoder teaches encoder, latent bottleneck, decoder, and reconstruction learning.
Step by step
- Encode the input.
- Pass through a latent bottleneck.
- Decode the latent value.
- Measure reconstruction error.
- Backpropagate decoder and encoder gradients.
Core formulas
Encoder
z = tanh(W_e x + b_e)The input is compressed into a bottleneck.
Reconstruction loss
L = ½||x̂−x||²Training reduces reconstruction error.
When to use Autoencoder
- Learning representation learning and bottlenecks.