Architectural patterns Introduction
Autoencoders
Learn a representation by reconstructing an input.
Before this
The intuition
An encoder maps an input to a representation. A decoder tries to reconstruct the original input from that representation. Training reduces reconstruction loss.
A concrete example
A network might encode ten input measurements into three numbers, then reconstruct the ten measurements from those three. The bottleneck encourages it to retain information useful for reconstruction.
What to remember
A bottleneck does not guarantee a useful representation. Reconstruction quality and usefulness for another task are different objectives.
Check your understanding
Where do the reconstruction targets come from?
Reveal explanation
From the inputs themselves; no separate human-provided target labels are needed.