I've been reading about the importance of Generative Adversarial Networks (GANs), and I would like to double check that I understood correctly why they are so relevant.

Before GANs, what people did was to train a neural network and use a set of functions for classification/prediction purposes. The innovation introduced by GANs is that the model is converted from a set of functions into an object, which could be an image or something similar. Then the similarity of the input with this model is evaluated.

I'm not sure why this approach is groundbreaking: we still make a model and we evaluate how close the input is to it. Am I missing something?


I think that the difference is that you are starting to "generate" stuff, the name gives it away.

The key difference with GAN is that you have two different models which compete against each other, but at the very end you'll get a win-win situation from this.

The innovative part is the competition. With GAN you have two models, where the output from one model is the input to the other.

You'll have a zero-sum game, which is different from a "classical" optimization problem.

It's believed that they will be helpful in the process of generating realistic input data (really close to the original).

One main problem of Deep Learning is often the lack of input data, with GAN this may change.

Interesting articles:

Hype around Dueling neural networks

Interesting Use


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.