Open Questions about Generative Adversarial Networks (GANS)

Question

New distill.pub article about future direction of GAN research

Open Questions about Generative Adversarial Networks
What we’d like to find out about GANs that we don’t know yet.
What are the trade-offs between GANs and other generative models?
What sorts of distributions can GANs model?
How can we Scale GANs beyond image synthesis?
What can we say about the global convergence of the training dynamics?
How should we evaluate GANs and when should we use them?
How does GAN training scale with batch size?
What is the relationship between GANs and adversarial examples?

https://distill.pub/2019/gan-open-problems/

Chris Gregori · Accepted Answer

Question: How does copyright work for GAN output? If I input 300,000 copyright protected photos of celebrities and generate images of new celebrities that do not exist, are the generated images public domain or would there be copyright issues?

Gufran Mirza · Answer

As I understand it, there is no "general infringement", it has to infringe on a particular work (there needs to be a plaintiff). So if it's not really similar to any of the inputs, there shouldn't be a problem. BUT there is another problem: are we really allowed to use datasets for any purpose? For example, I've worked with an app that can automatically mix between tracks from spotify. For that to work, we needed to extract metadata such as tempo, musical key, downbeat positions, etc. But in the contract with spotify, they have a clause where they state that any data derived from their data also belongs to them. So that would mean that they own the metadata files. That may not hold up in court, but it's at least a problem to consider. . For faces, there is a third issue, which is that real people have the right to their own likeness, as argued by the family of frank sinatra: https://www.npr.org/2015/12/11/459313019/hypothetical-coffee-mug-jolts-sinatra-into-action-to-protect-his-image

Chris Gregori · Answer

How about, "What's the relationship between GANs and Metropolis samplers, and can we train GANs well enough that we can compute accurate probabilities by integrating over the sampling distribution?"

Artificial Intelligence

Similar Discussions

An implementation of "Ego-splitting Framework: from Non-Overlapping to Overlapping Clusters" (KDD 2017).

I trained a face predictor that detects fulls bounds of face (81 facial landmarks vs dlib's 68)

On a mission to build Next-Gen Community Platform for Developers

Facebook

Twitter

Linkedin

Instagram