To understand GANs, we must know how discriminative and generative algorithms work. Discriminative algorithms try to predict a label and classify the input data, or categorize them to where the data belongs. On the other hand, generative algorithms make an attempt to predict features to give a certain label.

For example, a discriminative algorithm can predict whether an email is spam or not. Here, spam is one of the labels, and the text that's captured from the message is considered the input data. If you consider the label as y and the input as x, we can formulate this as follows:

On the other hand, generative algorithms try to guess how likely these input features (x, in the previous equation) are. Generative models care about how you get x, while discriminative models care about the relation between x and y.

Using the MNIST database as an example, the generator will generate images and pass them on to the discriminator. The discriminator will authenticate the image if it is truly from the MNIST dataset. The generator generates images with the hope that it will pass through the discriminator and be authenticated, even though it is fake (as shown in the preceding diagram).