In an [[autoencoder]] neural network, the latent space, also known as the hidden state vector or bottleneck, is the module that contains the compressed knowledge representations and is, therefore, the most important part of the network. This hidden layer contains a latent space representation of the input data.