Generative Adversarial Networks (Gans) - 01: Main Notions About Gans
Generative Adversarial Networks (Gans) - 01: Main Notions About Gans
definition
Generative Adversarial Networks (GANs) are deep generative models that can produce new
pieces of content. This kind of neural network can generate, for example, images, texts, or
music. It was introduced in 2014 by Ian J. Goodfellow and his co-authors in the article “Generative
Adversarial Nets”.
For example, the problem of generating a new human face is equivalent to the problem of generat-
ing a new data point following the "human faces probability distribution."
To generate data from our target distribution, we use the inverse transform method. Once trained,
the generative network in a GAN takes points following a simple distribution as input and
turns them into points following the target distribution.
Generated
distribution
Inputs follow a
simple probability
distribution
generative True target
network distribution
When training the generative network, the target and the generated distributions are not directly
compared. Instead, a discriminative network is trained to take true data and generated data
and to classify them.
During training, the two networks (generator and discriminator) have opposing goals. The
discriminator always wants to classify the data as accurately as it can. The generator always tries to
produce fake data that looks like the true data to fool the discriminator.
Inputs of the The generative The generated The discriminative The classification
generative network network is trained to distribution and the network is trained to error is the reference
follow a simple maximise the final true distribution are minimise the final metric for the training
distribution classification error not compared directly classification error of both networks
generative discriminative
network network
generative discriminative
network network
r
It becomes harder fo
the discriminator...
?
Final iterations
generative discriminative
network network