GAN

1. encoder decoder
2. latent vector/feature
3. input/ output feature
4. denoise autoencoder
5. sparse autoencoder
6. contractive autoencoder
7. vae(变分自动编码器)

x should be identity of x~

f(g(x)) ~~ x

1. add stachostic noise to improve robustness / overcome overfitting (when parameter is larger than data point)
2. sensitive is measure by F-norm of Jacobian matrix of encoder activations inspect to input