OVERVIEW

Variational AutoEncoders are 2 neural networks

1. Decoder (simulates p(Z|X)) - Outputs a vector of $\sigma,\mu$ - parameters for normal distributions
2. Endoder (simulates p(X|Z))

Random variables

• Z is Random Variable in latent space — Assumption is that it has normal distribution.
• X is Random Variable modeling the actual data (Usually an image)

The learning process tries to make NN1 and NN2 to most accurately regenerate process

$X$→ NN1 → $\mu,\sigma$→ draw Z→ NN2 → $\hat{X}$

The cost function used is