We can simply assume the encoder learn the posterior p(z|x) with a network with parameters \phi, hence the encode is a function that approximates the posterior q_{\phi}(z|x).
Decode takes a sample from latent space (when infer without encoder) z \sim N(0, 1), and map them to the space x.
We must train the encoder and decoder jointly. To achieve this, we must design a joint loss.
Leave a Reply