Why are generative opposing networks so creative

Generative Adversarial Networks: The Creative Side of Machine Learning

Before we get into what Generative Adversarial Networks can do for us, let's see what they really are.

It is a machine learning system ( machine learning ) developed in 2014 by a team led by Ian Goodfellow. The purpose of a generative adversary network is to create your own designs based on a range of real-world data. The result is so deceptively real that it is impossible to know that the picture was not made by a human hand. To achieve these results, two competing neural networks are used.

The task of the generating network is to create a fake. The network feeds on data, for example photos of random people, and creates its own photo from the information received. To do this, the network must first learn the common characteristics of all the photos displayed. In this way, the new image is not an imitation of the original data, but a completely new but similar work. In our example it would be a photo of a person (who does not exist).

The basic data and the information generated by it are delivered together to the second network. The task of the discriminatory network is to decide whether the data received is true or false. The image is not only declared as wrong if it deviates too much from the basic data, but also if it is too perfect an imitation. If the generating network limits itself to extracting an average of the data and creating a new work from it, the result will look artificial. In this way, the discriminating network also filters out unnatural looking data.

Both networks compete with each other . When the discriminating network detects the falsified data, it returns it. In this case, the generator grid is not yet good enough and so you need to keep learning. However, the discriminating network learns again. Since both neural networks train each other, they are connected with deep learning systems. The generating network tries to create data sets that look so authentic that they deceive the discriminator. This in turn tries to analyze and understand the real examples with such accuracy that the wrong data has no chance of being identified as real.