Hands-On Machine Learning on Google Cloud Platform

Adversarial autoencoder

Adversarial autoencoder (AAE) is a generative model produced by the union of VAE and GAN. To explain the model, we start by defining the following terms:

• x: Autoencoder input
• z: Code produced from x,
• p(z): The distribution we want to impose
• q(z|x): Distribution learned from the encoder
• p(x|z): Distribution learned from the decoder
• pdata: Distribution of the data
• p(x): Distribution of the model

We consider the function q(z|x) as a posterior distribution of q(z), which is defined as follows:

We try to impose the equality q(z)=p(z) on the model. The difference with a VAE is due to the fact that what drives q (z) towards ...