Authors
Ian J. Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley,
Sherjil Ozair, Aaron Courville, Yoshua Bengio
Published - 2014
number of citation : 59043 last checked at 23/08/10
paper - https://arxiv.org/pdf/1406.2661.pdf
official code - https://www.github.com/goodfeli/adversarial
my code - https://github.com/JaeikBae/gan
Simultaneously train two models
- Generative model : G
- Discriminative model : D
G captures the data distribution.
D estimates the probability that a sample came from the training data rather then G.
G trained for maximize the probability of D making a mistake.
So far, the most striking successes in deep learning have involved discriminative models, usually those that map a high-dimensional, rich sensory input to a class label.
Deep generative models have had less of an impact, due to the difficulty of approximating…
Before GAN, Boltzmann machines(BM) were most reasonable generative model.
BM studies probability distributions and uses them for estimation(or generation). But it needs so many probabilistic computations which hard to use the benefit of piecewise linear units.
BM use probabilistic computations like Markov Chain Monte Carlo (MCMC).
The generative model can be thought of as analogous to a team of counterfeiters, trying to produce fake currency and use it without detection, while the discriminative model is analogous to the police, trying to detect the counterfeit currency.
GAN has a very simple and novel idea. It trains two models at the same time. The first is a discriminator that finds sample sources (training data or generator). The second is a generator that tries to deceive the discriminator. It's like a counterfeiters and a police.
Then auther said, this is framework. It means, another structures can be applied to the generator & discriminator. We used to apply GAN to image processing. But that is sample case in paper. In sample, auther used multilayer perceptron structure.