AI
]
Generative Adversarial Networks
The course slides:
Generative models
Explicit
What kind of distribution could have generated this data?
-
We can learn an explicit model of the data, this
-
Answer quesitons using the model
Impilicit
simulator that cna generate same properties
In ML, we train the models to maximize the probability distribution of the data under our model.
train latent variable models with maximum lax gets tricky use approxiate maximum likelihood
2014 GAN Ppaer show we can go from simple image of digits to images of faces, small resolution
black and white to colored
break the imagenet barrier in 2018, trained
generate faces at very high resolution quite photo realistic
not only on high diversity but also high resolution
how are gans able to learn the probabilic so accuaraly to generate such data
player:
- discriminator
- generator
have a distribution on the input of our model
variational auto-encoder
distribution - multi-variate gaussian noise
noise is much more dimensional than the data
pass it to deterministic nn
discriminator given some set of samples from our data and given form models,
are these real or are t hese generated?
discriminator as a teacher guide and also improve itself
ge