Generative adversarial network

src: i.ytimg.com

The generative hostility network ( GANs ) is a class of artificial intelligence algorithms used in unattended machine learning, implemented by a system of two neural networks that compete with each other in a zero-sum game frame. They were introduced by Ian Goodfellow et al. in 2014. This technique can produce the most invisible photographs that are authentic to human observers, have many realistic characteristics (though in tests people can say real from those generated in many cases).

Video Generative adversarial network

Metode

One network produces the candidate (generative) and the other evaluates them (discriminatory). Typically, generative networks learn to map from latent space to particular interesting data distributions, while discriminatory networks distinguish between instances of the correct data distribution and generator generated candidates. The purpose of generative network training is to increase the error rate of discriminatory networks (ie, "cheat" discriminator networks by producing new synthesized examples that seem to come from the correct data distribution).

In practice, known datasets serve as initial training data for discriminators. Discriminator training involves presenting it with samples from a dataset, to a certain degree of accuracy. Generally the generator is provided with random inputs sampled from a predetermined latent space (eg, multivariate normal distribution). After that, the sample synthesized by the generator is evaluated by the discriminator. Backpropagation is applied in both networks so that the generator produces better images, while the discriminator becomes more skilled in marking synthetic images. The generator is usually a deconvolutional nerve tissue, and the discriminator is a convolutional neural network.

The idea to conclude the model in a competitive setting (model versus discriminator) was proposed by Li, Gauci and Gross in 2013. Their method was used for behavioral inference. This is called Turing Learning, because the setting is similar to the Turing test. Turing Learning is a generalization of the GAN. Models other than neural networks can be considered. In addition, discriminators are allowed to influence the process from which the dataset is obtained, making them active interrogators as in the Turing test. The idea of ​​hostility training can also be found in previous works, such as Schmidhuber in 1992.

Maps Generative adversarial network

Apps

GAN has been used to produce photorealistic image samples for the purpose of visualizing new/industrial interior designs, shoes, bags and clothing or items for computer game scenes. This network is reportedly used by Facebook. Recently, GAN modeled motion patterns in the video. They have also been used to reconstruct 3D object models from images and to enhance astronomical images. By 2017, fully convolutional feedback is used for image enhancement using automatic texture synthesis in combination with loss of perception. The system focuses on realistic textures rather than pixel accuracy. The result is higher image quality at high magnification.

src: 3dgan.csail.mit.edu

References

src: i.ytimg.com

External links

• Knight, Will. "What to expect from artificial intelligence in 2017". MIT Technology Review . Retrieved 2017-01-05 .

Source of the article : Wikipedia