Autoencoders are a type of neural network.
For autoencoders the goal for the output layer is the reconstructed input layer, rather than a classification.
By including sparsity in the neural network we can reduce the dimensions. This splits the network into an encoder and a decoder.
Middle vector is called latent variables.
Like AE, but force middle vector to have unit gaussian by adding new loss function
Now we can generate new images by sampling for latent normal of unit 1.