You are currently viewing SemiWiki as a guest which gives you limited access to the site. To view blog comments and experience other SemiWiki features you must be a registered member. Registration is fast, simple, and absolutely free so please, join our community today!
GAN, generative adversarial networks looks like it represents a major step in AI in that it gives it imagination. GAN is being used on Nvidia machines with some startling results and great success. It has been used on imagery and particle physics. Having to AIs in an adversarial mode introduces a whole new method of creating something much closer to true AI. I put the links below as a starting point and would appreciate any thoughts or comments on this what I feel is a very important step in AI. If anyone is working on semis for this their comments would be especially appreciated.
I've played around with these a bit in TensorFlow, really interesting possibilities since essentially it's an inversion of a typical NN. Where a NN can look at data and tell you something about it, with a GAN you can tell it what you want the data to look like and it can generate it. So you can say you want a picture of small yellow bird and a GAN will generate a realistic picture of a small yellow bird. A cool site I found that I'm pretty sure has a GAN under the hood is Brandmark - Deep learning for logo design, which uses "deep learning" to design corporate logos.
1. GANs can also address a huge challenge in AI - that of requirement of a large number of data sets for training (another one in that space is "capsule networks")
2. Magic Pony that got acquired by Twitter a year or two back employed a tech with similar outcomes; am not sure if GANs is used there.
3. GANs require a large amount of training time.Hence GPUs are more favourable there than traditional CPUs. That may also account for an additional interest from nVidia
4. There is a lot of work ongoing in the startup space as well as system companies in development of suitable hardware and architectures for optimal leveraging of AI (AI algorithms and underlying hardware). Maybe worthwhile to check in that space for ones working for GAN.
