The compressed representation layer in an autoencoder that forces the model to learn efficient, meaningful encodings of input data.