Introduction
The autoencoder in topic 1.1 had just one hidden layer on each side. Like any neural network, an autoencoder becomes more powerful with depth: more layers can learn more complex, hierarchical features. A deep autoencoder is called a stacked autoencoder.
For images, we can do better still by using convolutional layers (Computer Vision tab) instead of dense ones, so the autoencoder respects the 2-D structure of pictures.
Think of describing a photo over the phone. A good describer works in layers: first the big picture ("a beach at sunset"), then the main objects ("two people, a dog"), then details. A deep encoder learns these levels of description; a deep decoder rebuilds the picture level by level in reverse.