# Feature Extraction and Disentanglement

Feature Extraction and Disentanglement

## Feature Disentangle

• The GAN receives one vector as input and output a desirable result.
• We hope to control the characteristic of the output by mopdifying each specific value in the input vector.
• For general GAN, modifying a specific dimension of the vector commonly change the feature of the result unconsciously
• Because the actual distribution of each feature are intricate and entangled in the latent space.

### InfoGAN

• Split input z to two parts, c encodes the different feature in each dimension and z' as input noise.
• Classifier recover the predict c from the output x from the Generator, which supervises the generate to output x with the feature of c.
• Discriminator still output a scalar to represent the result good or not, but shares the parameter with Classifier except for the last output layer.
• Without Discriminator, the output from Generator will only focus on c which benefit for the Classifier to predict, but generate bad results.

*reference-2

### VAE-GAN

• Based on VAE (variational auto-encoder), VAE-GAN combines VAE and GAN
• VAE only generates obscure results.
• Discriminator

TODO