How does GANs Work?
GAN works on a completely newer approach of solving things, rather than a conventional approach and already built up neural networks. The network seems simple as it, itself, learns to generate the data through a training distribution in a way of a 2-player game, and the two players or entities doing all the work are Discriminator and Generator. These two are the entities that in a constant gameplay battle throughout the training process. The generator, as we get cleared from the word, generates new data or instances of data completely different from the real data by adding random noise obeying the similar data distribution, i.e., the new data is not differentiable as the new data is just like the data that could have been the real data. The discriminator comes into the role and its role is to guess or find the data that is from the real dataset or made from the generator. Both these models are Generatively Deep Neural Networks. The generator learns from the feedback provided by the discriminator as the discriminator is unbiased. Thus, to have a more successful model, we have to get the maximum probability of discriminator making a mistake and generator to cover that mistake. The generator wants a chance of low error to become better so that the data produced by it is accurate like the original. Thus, this shows us that there is a trade-off between the two entities to make each other and oneself better. Thus, the optimization goal for both entities is to reach the Nash Equilibrium.
The Noise is the input to the generator giving us the data mixed with the noise as the output, whereas, the discriminator’s output is the result that whether the data is the original or fake. All valid signals are labelled as 1.0, i.e. telling that the corresponding signals are 100% original and the ones from the generator are labelled as 0.0, telling that those signals are fake. Since all the labelling process is automatic, GAN is still the sub-part of unsupervised learning in Deep learning with the help of Deep Neural Networks.
