I drew the encoder on its own first, because the autoencoder is much easier to read once we have seen this half. An encoder is an MLP that gets narrower. The same layer as before, stacked three times, except each one is thinner than the one above it. 64 features become 32, then 16, then 8.
Every layer has to throw something away, because there is no room to keep it all. What comes out at the bottom is a code: the whole input rewritten in 8 numbers. Whether those 8 numbers are any good depends on what we ask the network to do with them next.
The tighter we squeeze the code, the more the encoder has to decide what matters.
Next:
6. Decoder